def __init__(self, **kwargs):
super(WordPieceVectorizer1D, self).__init__(kwargs.get('transform_fn'))
global BERT_TOKENIZER
self.max_seen = 128
handle = kwargs.get('embed_file')
if BERT_TOKENIZER is None:
BERT_TOKENIZER = BertTokenizer.from_pretrained(handle)
self.tokenizer = BERT_TOKENIZER
self.mxlen = kwargs.get('mxlen', -1)
def __init__(self,
pretrained_model: str,
use_starting_offsets: bool = False,
do_lowercase: bool = True,
max_pieces: int = 512) -> None:
bert_tokenizer = BertTokenizer.from_pretrained(pretrained_model, do_lowercase)
super().__init__(vocab=bert_tokenizer.vocab,
wordpiece_tokenizer=bert_tokenizer.wordpiece_tokenizer.tokenize,
namespace="bert",
use_starting_offsets=use_starting_offsets,
max_pieces=max_pieces)
def __init__(self, name, **kwargs):
super(BERTBaseEmbeddings, self).__init__(name=name, **kwargs)
global BERT_TOKENIZER
self.dsz = kwargs.get('dsz')
if BERT_TOKENIZER is None:
BERT_TOKENIZER = BertTokenizer.from_pretrained(kwargs.get('embed_file'))
self.model = BertModel.from_pretrained(kwargs.get('embed_file'))
self.vocab = BERT_TOKENIZER.vocab
self.vsz = len(BERT_TOKENIZER.vocab) # 30522 self.model.embeddings.word_embeddings.num_embeddings
self.layer_indices = kwargs.get('layers', [-1, -2, -3, -4])
self.operator = kwargs.get('operator', 'concat')
def __init__(self,
pretrained_model: str,
use_starting_offsets: bool = False,
do_lowercase: bool = True,
never_lowercase: List[str] = None,
max_pieces: int = 512) -> None:
if pretrained_model.endswith("-cased") and do_lowercase:
logger.warning("Your BERT model appears to be cased, "
"but your indexer is lowercasing tokens.")
elif pretrained_model.endswith("-uncased") and not do_lowercase:
logger.warning("Your BERT model appears to be uncased, "
"but your indexer is not lowercasing tokens.")
bert_tokenizer = BertTokenizer.from_pretrained(pretrained_model, do_lower_case=do_lowercase)
super().__init__(vocab=bert_tokenizer.vocab,
wordpiece_tokenizer=bert_tokenizer.wordpiece_tokenizer.tokenize,
namespace="bert",
use_starting_offsets=use_starting_offsets,
max_pieces=max_pieces,
do_lowercase=do_lowercase,
never_lowercase=never_lowercase,
start_tokens=["[CLS]"],
end_tokens=["[SEP]"])
def main():
args = process_args()
os.makedirs(args.output_dir, exist_ok=True)
if args.enable_butd:
if args.visdial_v == '1.0':
assert (args.len_vis_input == 36)
elif args.visdial_v == '0.9':
assert (args.len_vis_input == 100)
args.region_bbox_file = os.path.join(args.image_root, args.region_bbox_file)
args.region_det_file_prefix = os.path.join(args.image_root,
args.region_det_file_prefix) if args.dataset in (
'cc', 'coco') and args.region_det_file_prefix != '' else ''
device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
# fix random seed
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
tokenizer = BertTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.do_lower_case)
# for generative, there is no [SEP] at the end
args.max_seq_length = args.len_vis_input + 2 + args.max_len_hist_ques + 2 + args.max_len_ans
tokenizer.max_len = args.max_seq_length
bi_uni_pipeline = [Preprocess4TestVisdialGen(list(tokenizer.vocab.keys()),
tokenizer.convert_tokens_to_ids, args.max_seq_length,
new_segment_ids=args.new_segment_ids,
truncate_config={'len_vis_input': args.len_vis_input,
'max_len_hist_ques': args.max_len_hist_ques,
'max_len_ans': args.max_len_ans},
mode="s2s",
region_bbox_file=args.region_bbox_file,
region_det_file_prefix=args.region_det_file_prefix,
image_features_hdfpath=args.image_features_hdfpath,
visdial_v=args.visdial_v, pad_hist=args.pad_hist,
inc_full_hist=args.inc_full_hist)]
amp_handle = None
if args.fp16 and args.amp:
from apex import amp
amp_handle = amp.init(enable_caching=True)
logger.info("enable fp16 with amp")
# Prepare model
cls_num_labels = 2
type_vocab_size = 6 if args.new_segment_ids else 2
mask_word_id, eos_word_ids = tokenizer.convert_tokens_to_ids(
["[MASK]", "[SEP]"])
logger.info('Attempting to recover models from: {}'.format(args.model_recover_path))
if 0 == len(glob.glob(args.model_recover_path.strip())):
logger.error('There are no models to recover. The program will exit.')
sys.exit(1)
for model_recover_path in glob.glob(args.model_recover_path.strip()):
logger.info("***** Recover model: %s *****", model_recover_path)
model_recover = torch.load(model_recover_path)
model = BertForVisDialGen.from_pretrained(args.bert_model,
max_position_embeddings=512,
config_path=args.config_path,
state_dict=model_recover, num_labels=cls_num_labels,
type_vocab_size=type_vocab_size, task_idx=3,
mask_word_id=mask_word_id,
search_beam_size=args.beam_size,
length_penalty=args.length_penalty,
eos_id=eos_word_ids,
ngram_size=args.ngram_size,
min_len=args.min_len,
enable_butd=args.enable_butd, len_vis_input=args.len_vis_input,
tokenizer=tokenizer, decode_verbose=args.decode_verbose,
visdial_v=args.visdial_v)
del model_recover
if args.fp16:
model.half()
# cnn.half()
model.to(device)
# cnn.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
# cnn = torch.nn.DataParallel(cnn)
torch.cuda.empty_cache()
model.eval()
def read_data(src_file):
eval_lst = []
with open(src_file, "r", encoding='utf-8') as f_src:
data = json.load(f_src)['data']
dialogs = data['dialogs']
questions = data['questions']
answers = data['answers']
img_idx = 0
for dialog in tqdm(dialogs):
if img_idx < args.use_num_imgs or args.use_num_imgs == -1:
img_id = dialog['image_id']
cap_tokens = tokenizer.tokenize(dialog['caption'])
ques_id = [item['question'] for item in dialog['dialog']]
ques_tokens = [tokenizer.tokenize(questions[id] + '?') for id in ques_id]
ans_id = [item['answer'] for item in dialog['dialog']]
ans_tokens = [tokenizer.tokenize(answers[id]) for id in ans_id]
gt_id = [item['gt_index'] for item in dialog['dialog']]
ans_opts = [item['answer_options'] for item in dialog['dialog']]
ans_opts_tokens = [[tokenizer.tokenize(answers[id]) for id in ans] for ans in ans_opts]
assert len(ques_tokens) == len(ans_tokens) == len(ans_opts_tokens) == 10, \
"ques num: %d, ans num: %d, ans opt num: %d" % (
len(ques_tokens), len(ans_tokens), len(ans_opts_tokens))
assert all(
[len(ans_opt) == 100 for ans_opt in ans_opts_tokens]), "all the answer have 100 options"
eval_lst.append(
(img_id, cap_tokens, ques_tokens, ans_tokens, ans_opts_tokens, gt_id))
img_idx += 1
return eval_lst
def get_gt_rel_dict(fname):
gt_rel_dict = {}
gt_rel_data = json.load(open(fname))
for item in gt_rel_data:
image_id = item['image_id']
round_id = item['round_id']
gt_relevance = item['gt_relevance']
# each image only at most has one turn having dense annotation
if image_id not in gt_rel_dict:
gt_rel_dict[image_id] = (round_id, gt_relevance)
return gt_rel_dict
if args.gt_rel_file != '':
gt_rel_dict = get_gt_rel_dict(args.gt_rel_file)
input_lines = read_data(args.src_file)
next_i = 0
total_batch = math.ceil(len(input_lines) / args.batch_size)
print('start the visdial decode evaluation...')
t0 = time.time()
ranks_json = []
sparse_metrics = SparseGTMetrics()
ndcg = NDCG()
with tqdm(total=total_batch) as pbar:
while next_i < len(input_lines):
_chunk = input_lines[next_i:next_i + args.batch_size]
buf_id = [x[0] for x in _chunk]
buf = [x[:-1] for x in _chunk]
buf_gt_id = [x[-1] for x in _chunk]
next_i += args.batch_size
instances = []
for instance in buf:
instances.append(bi_uni_pipeline[0](instance))
with torch.no_grad():
buf_gt_id = torch.tensor(buf_gt_id).long().to(device)
batch_data = list(zip(*instances))
img, vis_pe = (torch.stack(x).to(device) for x in batch_data[-2:])
task_idx = torch.tensor(batch_data[-3], dtype=torch.long).to(device)
conv_feats = img.data # Bx100x2048
vis_pe = vis_pe.data
output_scores_turn = []
# (input_ids_turn, segment_ids_turn, position_ids_turn, ans_ids_turn, ans_opts_ids_turn,
# input_mask_turn, self.task_idx, img, vis_pe)
input_ids_turns = [[x[turn_i] for x in batch_data[0]] for turn_i in range(10)]
segment_ids_turns = [[x[turn_i] for x in batch_data[1]] for turn_i in range(10)]
position_ids_turns = [[x[turn_i] for x in batch_data[2]] for turn_i in range(10)]
ans_ids_turns = [[x[turn_i] for x in batch_data[3]] for turn_i in range(10)]
ans_opts_ids_turns = [[x[turn_i] for x in batch_data[4]] for turn_i in range(10)]
input_mask_turns = [[x[turn_i] for x in batch_data[5]] for turn_i in range(10)]
for turn_i in range(10):
input_ids = torch.tensor(input_ids_turns[turn_i], dtype=torch.long).to(device)
segment_ids = torch.tensor(segment_ids_turns[turn_i], dtype=torch.long).to(device)
position_ids = torch.tensor(position_ids_turns[turn_i], dtype=torch.long).to(device)
ans_ids = torch.tensor(ans_ids_turns[turn_i], dtype=torch.long).to(device)
ans_opts_ids = torch.tensor(ans_opts_ids_turns[turn_i], dtype=torch.long).to(device)
input_mask = torch.stack(input_mask_turns[turn_i]).to(device)
output_scores = model(conv_feats, vis_pe, input_ids, segment_ids,
position_ids, input_mask, ans_ids, ans_opts_ids, task_idx=task_idx)
output_scores = output_scores[-1] # [batch_size, num_options]
output_scores_turn.append(output_scores)
output_scores_turn = torch.stack(output_scores_turn, 1) # [batch_size, num_rounds, num_options]
ranks = scores_to_ranks(output_scores_turn)
# output_scores_turn_cheat = output_scores_turn.scatter_(2, buf_gt_id.unsqueeze(2), 100.0)
sparse_metrics.observe(output_scores_turn, buf_gt_id)
for i in range(len(buf_id)):
# Cast into types explicitly to ensure no errors in schema.
# Round ids are 1-10, not 0-9
if args.split == "val":
for j in range(10):
ranks_json.append(
{
"image_id": buf_id[i],
"round_id": int(j + 1),
"ranks": [rank.item() for rank in ranks[i][j]],
}
)
if args.gt_rel_file:
scores = []
gt_rels = []
for i in range(len(buf_id)):
if buf_id[i] in gt_rel_dict:
turn_idx, gt_rel = gt_rel_dict[buf_id[i]]
scores.append(output_scores_turn[i, turn_idx - 1, :])
gt_rels.append(torch.tensor(gt_rel, dtype=torch.float32).to(device))
scores = torch.stack(scores)
gt_rels = torch.stack(gt_rels)
ndcg.observe(scores, gt_rels)
pbar.update(1)
json.dump(ranks_json, open(args.save_ranks_path, "w"))
logger.info("Finish writing rankings into %s" % (args.save_ranks_path))
if args.split == "val":
fw = open(args.save_ranks_path.replace('.json', '_results.txt'), "w")
all_metrics = {}
all_metrics.update(sparse_metrics.retrieve(reset=True))
if args.gt_rel_file:
all_metrics.update(ndcg.retrieve(reset=True))
for metric_name, metric_value in all_metrics.items():
print(f"{metric_name}: {metric_value}")
fw.write("%s: %.6f\n" % (metric_name, metric_value))
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument("--bert_model", default=None, type=str, required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.")
parser.add_argument("--model_recover_path", default=None, type=str,
help="The file of fine-tuned pretraining model.")
parser.add_argument("--max_seq_length", default=512, type=int,
help="The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument('--ffn_type', default=0, type=int,
help="0: default mlp; 1: W((Wx+b) elem_prod x);")
parser.add_argument('--num_qkv', default=0, type=int,
help="Number of different <Q,K,V>.")
parser.add_argument('--seg_emb', action='store_true',
help="Using segment embedding for self-attention.")
# decoding parameters
parser.add_argument('--fp16', action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--amp', action='store_true',
help="Whether to use amp for fp16")
parser.add_argument("--input_file", type=str, help="Input file")
parser.add_argument('--subset', type=int, default=0,
help="Decode a subset of the input dataset.")
parser.add_argument("--output_file", type=str, help="output file")
parser.add_argument("--split", type=str, default="",
help="Data split (train/val/test).")
parser.add_argument('--tokenized_input', action='store_true',
help="Whether the input is tokenized.")
parser.add_argument('--seed', type=int, default=123,
help="random seed for initialization")
parser.add_argument("--do_lower_case", action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument('--new_segment_ids', action='store_true',
help="Use new segment ids for bi-uni-directional LM.")
parser.add_argument('--new_pos_ids', action='store_true',
help="Use new position ids for LMs.")
parser.add_argument('--batch_size', type=int, default=4,
help="Batch size for decoding.")
parser.add_argument('--beam_size', type=int, default=1,
help="Beam size for searching")
parser.add_argument('--length_penalty', type=float, default=0,
help="Length penalty for beam search")
parser.add_argument('--forbid_duplicate_ngrams', action='store_true')
parser.add_argument('--forbid_ignore_word', type=str, default=None,
help="Forbid the word during forbid_duplicate_ngrams")
parser.add_argument("--min_len", default=None, type=int)
parser.add_argument('--need_score_traces', action='store_true')
parser.add_argument('--ngram_size', type=int, default=3)
parser.add_argument('--mode', default="s2s",
choices=["s2s", "l2r", "both","2in1","a2q","q2a","c2a","c2q"])
parser.add_argument('--max_tgt_length', type=int, default=128,
help="maximum length of target sequence")
parser.add_argument('--s2s_special_token', action='store_true',
help="New special tokens ([S2S_SEP]/[S2S_CLS]) of S2S.")
parser.add_argument('--s2s_add_segment', action='store_true',
help="Additional segmental for the encoder of S2S.")
parser.add_argument('--s2s_share_segment', action='store_true',
help="Sharing segment embeddings for the encoder of S2S (used with --s2s_add_segment).")
parser.add_argument('--pos_shift', action='store_true',
help="Using position shift for fine-tuning.")
args = parser.parse_args()
if args.need_score_traces and args.beam_size <= 1:
raise ValueError(
"Score trace is only available for beam search with beam size > 1.")
if args.max_tgt_length >= args.max_seq_length - 2:
raise ValueError("Maximum tgt length exceeds max seq length - 2.")
device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
tokenizer = BertTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.do_lower_case)
tokenizer.max_len = args.max_seq_length
pair_num_relation = 0
bi_uni_pipeline = []
bi_uni_pipeline.append(seq2seq_loader.Preprocess4Seq2seqDecoder(list(tokenizer.vocab.keys()), tokenizer.convert_tokens_to_ids, args.max_seq_length, max_tgt_length=args.max_tgt_length, new_segment_ids=args.new_segment_ids,
mode=args.mode, num_qkv=args.num_qkv, s2s_special_token=args.s2s_special_token, s2s_add_segment=args.s2s_add_segment, s2s_share_segment=args.s2s_share_segment, pos_shift=args.pos_shift))
amp_handle = None
if args.fp16 and args.amp:
from apex import amp
amp_handle = amp.init(enable_caching=True)
logger.info("enable fp16 with amp")
# Prepare model
cls_num_labels = 2
type_vocab_size = 6 + \
(1 if args.s2s_add_segment else 0) if args.new_segment_ids else 2
if args.mode == "a2q":
mask_word_id, eos_word_ids, sos_word_id = tokenizer.convert_tokens_to_ids(
["[MASK]", "[SEP]", "[SEP_1]"])
elif args.mode == "q2a":
mask_word_id, eos_word_ids, sos_word_id = tokenizer.convert_tokens_to_ids(
["[MASK]", "[SEP]", "[SEP_2]"])
elif args.mode == "c2q":
mask_word_id, eos_word_ids, sos_word_id = tokenizer.convert_tokens_to_ids(
["[MASK]", "[SEP]", "[SEP_6]"])
elif args.mode == "c2a":
mask_word_id, eos_word_ids, sos_word_id = tokenizer.convert_tokens_to_ids(
["[MASK]", "[SEP]", "[SEP_7]"])
elif args.mode == "2in1":
mask_word_id, eos_word_ids, sos_word_id = tokenizer.convert_tokens_to_ids(
["[MASK]", "[SEP]", "[SEP_0]"])
else:
mask_word_id, eos_word_ids, sos_word_id = tokenizer.convert_tokens_to_ids(
["[MASK]", "[SEP]", "[S2S_SOS]"])
forbid_ignore_set = None
if args.forbid_ignore_word:
w_list = []
for w in args.forbid_ignore_word.split('|'):
if w.startswith('[') and w.endswith(']'):
w_list.append(w.upper())
else:
w_list.append(w)
forbid_ignore_set = set(tokenizer.convert_tokens_to_ids(w_list))
print(args.model_recover_path)
for model_recover_path in glob.glob(args.model_recover_path.strip()):
logger.info("***** Recover model: %s *****", model_recover_path)
model_recover = torch.load(model_recover_path)
model = BertForSeq2SeqDecoder.from_pretrained(args.bert_model, state_dict=model_recover, num_labels=cls_num_labels, num_rel=pair_num_relation, type_vocab_size=type_vocab_size, task_idx=3, mask_word_id=mask_word_id, search_beam_size=args.beam_size,
length_penalty=args.length_penalty, eos_id=eos_word_ids, sos_id=sos_word_id, forbid_duplicate_ngrams=args.forbid_duplicate_ngrams, forbid_ignore_set=forbid_ignore_set, ngram_size=args.ngram_size, min_len=args.min_len, mode=args.mode, max_position_embeddings=args.max_seq_length, ffn_type=args.ffn_type, num_qkv=args.num_qkv, seg_emb=args.seg_emb, pos_shift=args.pos_shift)
del model_recover
if args.fp16:
model.half()
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
torch.cuda.empty_cache()
model.eval()
next_i = 0
max_src_length = args.max_seq_length - 2 - args.max_tgt_length
with open(args.input_file, encoding="utf-8") as fin:
input_lines = [x.strip() for x in fin.readlines()]
if args.subset > 0:
logger.info("Decoding subset: %d", args.subset)
input_lines = input_lines[:args.subset]
data_tokenizer = WhitespaceTokenizer() if args.tokenized_input else tokenizer
input_lines = [data_tokenizer.tokenize(
x)[-max_src_length:] for x in input_lines]
input_lines = sorted(list(enumerate(input_lines)),
key=lambda x: -len(x[1]))
output_lines = [""] * len(input_lines)
score_trace_list = [None] * len(input_lines)
total_batch = math.ceil(len(input_lines) / args.batch_size)
with tqdm(total=total_batch) as pbar:
while next_i < len(input_lines):
_chunk = input_lines[next_i:next_i + args.batch_size]
buf_id = [x[0] for x in _chunk]
buf = [x[1] for x in _chunk]
next_i += args.batch_size
max_a_len = max([len(x) for x in buf])
instances = []
for instance in [(x, max_a_len) for x in buf]:
for proc in bi_uni_pipeline:
instances.append(proc(instance))
with torch.no_grad():
batch = seq2seq_loader.batch_list_to_batch_tensors(
instances)
batch = [
t.to(device) if t is not None else None for t in batch]
input_ids, token_type_ids, position_ids, input_mask, mask_qkv, task_idx = batch
traces = model(input_ids, token_type_ids,
position_ids, input_mask, task_idx=task_idx, mask_qkv=mask_qkv)
if args.beam_size > 1:
traces = {k: v.tolist() for k, v in traces.items()}
output_ids = traces['pred_seq']
else:
output_ids = traces.tolist()
for i in range(len(buf)):
w_ids = output_ids[i]
output_buf = tokenizer.convert_ids_to_tokens(w_ids)
output_tokens = []
for t in output_buf:
if t in ("[SEP]", "[PAD]","[S2S_SEP]"):
pass
break
output_tokens.append(t)
output_sequence = ' '.join(detokenize(output_tokens))
print(output_sequence)
output_lines[buf_id[i]] = output_sequence
if args.need_score_traces:
score_trace_list[buf_id[i]] = {
'scores': traces['scores'][i], 'wids': traces['wids'][i], 'ptrs': traces['ptrs'][i]}
pbar.update(1)
if args.output_file:
fn_out = args.output_file
else:
fn_out = model_recover_path+'.'+args.split
with open(fn_out, "w", encoding="utf-8") as fout:
for l in output_lines:
fout.write(l)
fout.write("\n")
if args.need_score_traces:
with open(fn_out + ".trace.pickle", "wb") as fout_trace:
pickle.dump(
{"version": 0.0, "num_samples": len(input_lines)}, fout_trace)
for x in score_trace_list:
pickle.dump(x, fout_trace)
def __init__(self,
output_dir,
model="best_model.bin",
topk=50,
bert_model="bert-large-cased",
do_lower_case=False,
train_batch_size=32,
seed=42,
eval_batch_size=64,
max_seq_length=128,
num_labels=2,
entail_label=1,
grad_acc_steps=1,
num_of_epochs=5,
learning_rate=1e-5,
warmup_proportion=0.1,
action="train"):
print("Loading BERT NLI Model")
self.name = "BertNLI"
self.topk = topk
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
tokenizer = BertTokenizer.from_pretrained(bert_model,
do_lower_case=do_lower_case)
self.device = device
self.tokenizer = tokenizer
self.max_seq_length = max_seq_length
self.train_batch_size = int(train_batch_size / grad_acc_steps)
self.eval_batch_size = eval_batch_size
self.num_labels = num_labels
self.entail_label = entail_label
self.grad_acc_steps = grad_acc_steps
self.gradient_accumulation_steps = self.grad_acc_steps
self.num_of_epochs = num_of_epochs
self.output_dir = output_dir
self.learning_rate = learning_rate
self.warmup_proportion = warmup_proportion
self.n_gpu = n_gpu
print("Action:", action)
if action == "train":
print("Params:")
print("Device:", device)
print("Max-Seq-Length:", max_seq_length)
print("Train-Batch-Size:", train_batch_size)
print("Num Labels:", num_labels)
print("Gradient Accumulation Steps:", grad_acc_steps)
print("Num of Train Epochs:", num_of_epochs)
print("Output Dir:", output_dir)
print("Learning Rate:", learning_rate)
print("Warmup Proportion:", warmup_proportion)
print("Number of Gpus:", n_gpu)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(seed)
os.makedirs(self.output_dir, exist_ok=True)
model = BertForSequenceClassification.from_pretrained(
bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(-1),
num_labels=num_labels)
model.to(device)
model = torch.nn.DataParallel(model)
self.model = model
self.best_metric = None
if action == "predict":
output_model_file = os.path.join(output_dir, model)
# Load a trained model that you have fine-tuned
print("Loading Model", output_model_file)
model_state_dict = torch.load(output_model_file)
model = BertForSequenceClassification.from_pretrained(
bert_model, state_dict=model_state_dict, num_labels=num_labels)
model.to(device)
model = torch.nn.DataParallel(model)
self.model = model
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--bert_model", default=None, type=str, required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--output_dir", default=None, type=str, required=True,
help="The output directory where the model checkpoints and predictions will be written.")
## Other parameters
parser.add_argument("--train_file", default=None, type=str, help="SQuAD json for training. E.g., train-v1.1.json")
parser.add_argument("--predict_file", default=None, type=str,
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
parser.add_argument("--max_seq_length", default=384, type=int,
help="The maximum total input sequence length after WordPiece tokenization. Sequences "
"longer than this will be truncated, and sequences shorter than this will be padded.")
parser.add_argument("--doc_stride", default=128, type=int,
help="When splitting up a long document into chunks, how much stride to take between chunks.")
parser.add_argument("--max_query_length", default=64, type=int,
help="The maximum number of tokens for the question. Questions longer than this will "
"be truncated to this length.")
parser.add_argument("--do_train", action='store_true', help="Whether to run training.")
parser.add_argument("--do_predict", action='store_true', help="Whether to run eval on the dev set.")
parser.add_argument("--train_batch_size", default=32, type=int, help="Total batch size for training.")
parser.add_argument("--predict_batch_size", default=8, type=int, help="Total batch size for predictions.")
parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs", default=3.0, type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--warmup_proportion", default=0.1, type=float,
help="Proportion of training to perform linear learning rate warmup for. E.g., 0.1 = 10%% "
"of training.")
parser.add_argument("--n_best_size", default=20, type=int,
help="The total number of n-best predictions to generate in the nbest_predictions.json "
"output file.")
parser.add_argument("--max_answer_length", default=30, type=int,
help="The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.")
parser.add_argument("--verbose_logging", action='store_true',
help="If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument('--gradient_accumulation_steps',
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.")
parser.add_argument("--do_lower_case",
action='store_true',
help="Whether to lower case the input text. True for uncased models, False for cased models.")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--loss_scale',
type=float, default=0,
help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--version_2_with_negative',
action='store_true',
help='If true, the SQuAD examples contain some that do not have an answer.')
parser.add_argument('--null_score_diff_threshold',
type=float, default=0.0,
help="If null_score - best_non_null is greater than the threshold predict null.")
args = parser.parse_args()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError("At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train:
raise ValueError("Output directory () already exists and is not empty.")
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = read_squad_examples(
input_file=args.train_file, is_training=True, version_2_with_negative=args.version_2_with_negative)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
# Prepare model
model = BertForQuestionAnswering.from_pretrained(args.bert_model,
cache_dir=os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank)))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
if args.do_train:
cached_train_features_file = args.train_file+'_{0}_{1}_{2}_{3}'.format(
list(filter(None, args.bert_model.split('/'))).pop(), str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length))
train_features = None
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info(" Saving train features into cached file %s", cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_start_positions, all_end_positions)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
if n_gpu == 1:
batch = tuple(t.to(device) for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, start_positions, end_positions = batch
loss = model(input_ids, segment_ids, input_mask, start_positions, end_positions)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used and handles this automatically
lr_this_step = args.learning_rate * warmup_linear(global_step/num_train_optimization_steps, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if args.do_train:
# Save a trained model and the associated configuration
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
# Load a trained model and config that you have fine-tuned
config = BertConfig(output_config_file)
model = BertForQuestionAnswering(config)
model.load_state_dict(torch.load(output_model_file))
else:
model = BertForQuestionAnswering.from_pretrained(args.bert_model)
model.to(device)
if args.do_predict and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
eval_examples = read_squad_examples(
input_file=args.predict_file, is_training=False, version_2_with_negative=args.version_2_with_negative)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
logger.info("***** Running predictions *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", args.predict_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.predict_batch_size)
model.eval()
all_results = []
logger.info("Start evaluating")
for input_ids, input_mask, segment_ids, example_indices in tqdm(eval_dataloader, desc="Evaluating"):
if len(all_results) % 1000 == 0:
logger.info("Processing example: %d" % (len(all_results)))
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
batch_start_logits, batch_end_logits = model(input_ids, segment_ids, input_mask)
for i, example_index in enumerate(example_indices):
start_logits = batch_start_logits[i].detach().cpu().tolist()
end_logits = batch_end_logits[i].detach().cpu().tolist()
eval_feature = eval_features[example_index.item()]
unique_id = int(eval_feature.unique_id)
all_results.append(RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits))
output_prediction_file = os.path.join(args.output_dir, "predictions.json")
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json")
output_null_log_odds_file = os.path.join(args.output_dir, "null_odds.json")
write_predictions(eval_examples, eval_features, all_results,
args.n_best_size, args.max_answer_length,
args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file, args.verbose_logging,
args.version_2_with_negative, args.null_score_diff_threshold)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
type=str,
required=True,
help=
"The input data dir. Should contain train.tsv and dev.tsv files for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument("--task_name",
type=str,
required=True,
help="The name of the task to train (codah or swag).")
parser.add_argument(
"--output_dir",
default="../bert-output/",
type=str,
required=False,
help="The output directory where the model checkpoints will be written."
)
## Other parameters
parser.add_argument(
'--model_labels_save_filename',
type=str,
default='model_labels.json',
help=
"JSON file to save model outputs/labels to (relative to output_dir)")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--answer_only",
default=False,
action='store_true',
help="Whether to run with answers only (blank out question).")
parser.add_argument(
"--do_lower_case",
default=False,
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=8,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=2e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--optimize_on_cpu',
default=False,
action='store_true',
help=
"Whether to perform optimization and keep the optimizer averages on CPU"
)
parser.add_argument(
'--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=128,
help=
'Loss scaling, positive power of 2 values can improve fp16 convergence.'
)
parser.add_argument(
'--save_final_only',
default=False,
action='store_true',
help=
"Whether to only save the final model from training (vs saving at every epoch)"
)
args = parser.parse_args()
processors = {
"codah": CodahProcessor,
"swag": SwagProcessor,
}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
if args.fp16:
logger.info(
"16-bits training currently not supported in distributed training"
)
args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info("device %s n_gpu %d distributed training %r", device, n_gpu,
bool(args.local_rank != -1))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
if args.answer_only:
for exset in train_examples:
for ex in exset:
ex.text_a = ''
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = BertForSequenceClassification.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank),
num_labels=1)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare loss criterion
criterion = CrossEntropyLoss()
# Prepare optimizer
if args.fp16:
param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
for n, param in model.named_parameters()]
elif args.optimize_on_cpu:
param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.0
}]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
global_step = 0
# Enable logging for eval-only use case
tr_loss = 0
nb_tr_steps = 1
if args.do_train:
# train_features will now be list of lists(of length 4)
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
# all ids are lists of lists(of 4)
all_input_ids = torch.tensor([[ff.input_ids for ff in f]
for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([[ff.input_mask for ff in f]
for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([[ff.segment_ids for ff in f]
for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([[ff.label_id for ff in f]
for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
# each minibatch that dataloader generates is a set of 4 answers for 1 question, each processed individually
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for epoch_num in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
# reshape and reduce the second dimension
# pull label from training data, set aside for softmax
n_batches = batch[0].shape[0]
batch = [ids.view(ids.shape[0] * 4, -1) for ids in batch]
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
logits = model(
input_ids, segment_ids, input_mask
) #, label_ids) label removed to skip softmax in model
logits = logits.view(-1, 4) # reshape to (:, 4)
# run 4-label softmax on each minibatch
loss = criterion(logits, label_ids.view(n_batches, 4)[:, 0])
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16 or args.optimize_on_cpu:
if args.fp16 and args.loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
if param.grad is not None:
param.grad.data = param.grad.data / args.loss_scale
is_nan = set_optimizer_params_grad(
param_optimizer,
model.named_parameters(),
test_nan=True)
if is_nan:
logger.info(
"FP16 TRAINING: Nan in gradients, reducing loss scaling"
)
args.loss_scale = args.loss_scale / 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(
model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad()
global_step += 1
if (epoch_num + 1 == int(
args.num_train_epochs)) or not args.save_final_only:
logger.info("***** Saving model to disk *****")
torch.save(
model.state_dict(), args.output_dir +
"pytorch_model_epoch{}.bin".format(epoch_num))
logger.info("***** Finished saving model to disk *****")
# Probably redundant, since BertAdam configurations remain unchanged throughout training, and params are
# read directly from model
# torch.save(optimizer.state_dict(), args.output_dir + "pytorch_optimizer_epoch{}.bin".format(_))
# logger.info("***** Finished saving optimizer to disk *****")
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
output_eval_result_file = open(os.path.join(args.output_dir,
'full-result.txt'),
'w+',
encoding='utf8')
eval_examples = processor.get_dev_examples(args.data_dir)
if args.answer_only:
for exset in eval_examples:
for ex in exset:
ex.text_a = ''
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([[ff.input_ids for ff in f]
for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([[ff.input_mask for ff in f]
for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([[ff.segment_ids for ff in f]
for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([[ff.label_id for ff in f]
for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
all_model_outputs = []
data_index = 0
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for input_ids, input_mask, segment_ids, label_ids in eval_dataloader:
input_ids = input_ids.view(input_ids.shape[0] * 4, -1).to(device)
input_mask = input_mask.view(input_mask.shape[0] * 4,
-1).to(device)
segment_ids = segment_ids.view(segment_ids.shape[0] * 4,
-1).to(device)
label_ids = label_ids[:, 0].to(device)
with torch.no_grad():
# output label dimension == num_labels == 1
# reshape into (bsz, 4)
logits = model(input_ids, segment_ids, input_mask).view(-1, 4)
tmp_eval_loss = criterion(logits, label_ids)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = accuracy_logging(logits, label_ids,
output_eval_result_file)
for i in range(logits.shape[0]):
output_obj = dict()
output_obj['logits'] = [float(x) for x in logits[i]]
output_obj['true_label'] = int(label_ids[i])
output_obj['model_label'] = int(np.argmax(logits[i]))
output_obj['swag_data'] = {
'context_sentence':
eval_examples[data_index][0].text_a,
'answer_candidates':
[ex.text_b for ex in eval_examples[data_index]]
}
all_model_outputs.append(output_obj)
data_index += 1
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += label_ids.shape[0]
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': tr_loss / nb_tr_steps
}
output_eval_result_file.close()
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
with open(
os.path.join(args.output_dir, args.model_labels_save_filename),
'w') as f:
json.dump(all_model_outputs, f)
import tqdm
import pickle
import re
import collections
import argparse
from sys import path
from data_utils.vocab import Vocabulary
from pytorch_pretrained_bert.tokenization import BertTokenizer
from data_utils.log_wrapper import create_logger
from data_utils.label_map import GLOBAL_MAP
from data_utils.glue_utils import *
DEBUG_MODE = False
MAX_SEQ_LEN = 512
bert_tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
logger = create_logger(__name__,
to_disk=True,
log_file='bert_data_proc_512.log')
def _truncate_seq_pair(tokens_a, tokens_b, max_length):
"""Truncates a sequence pair in place to the maximum length.
Copyed from https://github.com/huggingface/pytorch-pretrained-BERT
"""
# This is a simple heuristic which will always truncate the longer sequence
# one token at a time. This makes more sense than truncating an equal percent
# of tokens from each, since if one sequence is very short then each token
# that's truncated likely contains more information than a longer sequence.
while True:
total_length = len(tokens_a) + len(tokens_b)
norm_gcn_vocab_adj_list=[]
for i in range(len(gcn_vocab_adj_list)):
adj=gcn_vocab_adj_list[i] #.tocsr() #(lr是用非norm时的1/10)
print(' Zero ratio(?>66%%) for vocab adj %dth: %.8f'%(i, 100*(1-adj.count_nonzero()/(adj.shape[0]*adj.shape[1]))))
adj=normalize_adj(adj)
norm_gcn_vocab_adj_list.append(sparse_scipy2torch(adj.tocoo()).to(device))
gcn_adj_list=norm_gcn_vocab_adj_list
del gcn_vocab_adj_tf,gcn_vocab_adj,gcn_vocab_adj_list
gc.collect()
train_classes_num, train_classes_weight = get_class_count_and_weight(train_y,len(label2idx))
loss_weight=torch.tensor(train_classes_weight).to(device)
tokenizer = BertTokenizer.from_pretrained(bert_model_scale, do_lower_case=do_lower_case)
#%%
def get_pytorch_dataloader(examples, tokenizer, batch_size, shuffle_choice, classes_weight=None, total_resample_size=-1):
ds = CorpusDataset(examples, tokenizer, gcn_vocab_map, MAX_SEQ_LENGTH, gcn_embedding_dim)
if shuffle_choice==0: # shuffle==False
return DataLoader(dataset=ds,
batch_size=batch_size,
shuffle=False,
num_workers=4,
collate_fn=ds.pad)
elif shuffle_choice==1: # shuffle==True
return DataLoader(dataset=ds,
batch_size=batch_size,
shuffle=True,
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .csv files (or other data files) for the task.")
parser.add_argument("--bert_model", default=None, type=str, required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written.")
## Other parameters
parser.add_argument("--max_seq_length",
default=128,
type=int,
help="The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--warmup_proportion",
default=0.1,
type=float,
help="Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument('--gradient_accumulation_steps',
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.")
parser.add_argument('--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--loss_scale',
type=float, default=0,
help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
args = parser.parse_args()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError("At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = read_swag_examples(os.path.join(args.data_dir, 'train.csv'), is_training = True)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
# Prepare model
model = BertForMultipleChoice.from_pretrained(args.bert_model,
cache_dir=os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank)),
num_choices=4)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
if args.do_train:
train_features = convert_examples_to_features(
train_examples, tokenizer, args.max_seq_length, True)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = torch.tensor(select_field(train_features, 'input_ids'), dtype=torch.long)
all_input_mask = torch.tensor(select_field(train_features, 'input_mask'), dtype=torch.long)
all_segment_ids = torch.tensor(select_field(train_features, 'segment_ids'), dtype=torch.long)
all_label = torch.tensor([f.label for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step/num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if args.do_train:
# Save a trained model, configuration and tokenizer
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
# If we save using the predefined names, we can load using `from_pretrained`
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
# Load a trained model and vocabulary that you have fine-tuned
model = BertForMultipleChoice.from_pretrained(args.output_dir, num_choices=4)
tokenizer = BertTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
else:
model = BertForMultipleChoice.from_pretrained(args.bert_model, num_choices=4)
model.to(device)
if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
eval_examples = read_swag_examples(os.path.join(args.data_dir, 'val.csv'), is_training = True)
eval_features = convert_examples_to_features(
eval_examples, tokenizer, args.max_seq_length, True)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor(select_field(eval_features, 'input_ids'), dtype=torch.long)
all_input_mask = torch.tensor(select_field(eval_features, 'input_mask'), dtype=torch.long)
all_segment_ids = torch.tensor(select_field(eval_features, 'segment_ids'), dtype=torch.long)
all_label = torch.tensor([f.label for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for input_ids, input_mask, segment_ids, label_ids in tqdm(eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids, input_mask, label_ids)
logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = accuracy(logits, label_ids)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
result = {'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': tr_loss/nb_tr_steps}
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
os.makedirs(args.output_dir)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
output_mode = output_modes[task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
if args.local_rank not in [-1, 0]:
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
model = BertForSequenceClassification.from_pretrained(args.bert_model, num_labels=num_labels)
if args.local_rank == 0:
torch.distributed.barrier()
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_path",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
args = parser.parse_args()
tokenizer = BertTokenizer.from_pretrained(args.bert_model)
model = BertModel.from_pretrained(args.bert_model)
model.to(DEVICE)
# read dataframe
recipe_df = pd.read_csv(args.data_path)
recipe_df['pred_steps'] = recipe_df['pred_steps'].agg(eval)
bert_features_for_all_recipe = []
for index, recipe in tqdm(recipe_df.iterrows(), total=len(recipe_df)):
if len(recipe['pred_steps']) > 15:
steps = recipe['pred_steps'][:15]
else:
steps = recipe['pred_steps']
recipe_bert_like_input = bert_input_prepare(steps)
recipe_bert_features = get_bert_hidden(recipe_bert_like_input,
max_sent_length=256,
tokenizer=tokenizer,
model=model,
device=DEVICE,
max_recipe_steps=15)
bert_features_for_all_recipe.append(recipe_bert_features)
recipe_df['bert_features_steps'] = bert_features_for_all_recipe
recipe_df[['i', 'bert_features_steps']].to_msgpack(
os.path.join(args.output_dir,
'absolute_order_model_baseline_e9_pred.msgpack'))
def main(args):
# hyper param
do_lower_case = args.do_lower_case
root = args.root_dir
assert os.path.exists(root)
is_uncased = False
if 'uncased' in args.model:
is_uncased = True
mt_dnn_suffix = 'bert'
encoder_model = EncoderModelType.BERT
if 'xlnet' in args.model:
encoder_model = EncoderModelType.XLNET
mt_dnn_suffix = 'xlnet'
if 'roberta' in args.model:
encoder_model = EncoderModelType.ROBERTA
mt_dnn_suffix = 'roberta'
if encoder_model == EncoderModelType.ROBERTA:
if args.roberta_path is None or (not os.path.exists(
args.roberta_path)):
print('Please specify roberta model path')
encoder = get_encoder('{}/encoder.json'.format(args.roberta_path),
'{}/vocab.bpe'.format(args.roberta_path))
vocab = load_dict('{}/ict.txt'.format(args.roberta_path))
tokenizer = RoBERTaTokenizer(vocab, encoder)
elif encoder_model == EncoderModelType.XLNET:
tokenizer = spm.SentencePieceProcessor()
if 'large' in args.model:
tokenizer.load('mt_dnn_models/xlnet_large_cased_spiece.model')
else:
tokenizer.load('mt_dnn_models/xlnet_base_cased_spiece.model')
else:
tokenizer = BertTokenizer.from_pretrained(args.model,
do_lower_case=do_lower_case)
if is_uncased:
mt_dnn_suffix = '{}_uncased'.format(mt_dnn_suffix)
else:
mt_dnn_suffix = '{}_cased'.format(mt_dnn_suffix)
if do_lower_case:
mt_dnn_suffix = '{}_lower'.format(mt_dnn_suffix)
mt_dnn_root = os.path.join(root, mt_dnn_suffix)
if not os.path.isdir(mt_dnn_root):
os.mkdir(mt_dnn_root)
task_defs = TaskDefs(args.task_def)
task_def_dic = yaml.safe_load(open(args.task_def))
for task, task_def in task_def_dic.items():
logger.info("Task %s" % task)
data_format = DataFormat[task_def["data_format"]]
task_type = TaskType[task_def["task_type"]]
label_mapper = task_defs.global_map.get(task, None)
split_names = task_def.get("split_names", ["train", "dev", "test"])
for split_name in split_names:
rows = load_data(
os.path.join(root, "%s_%s.tsv" % (task, split_name)),
data_format, task_type, label_mapper)
dump_path = os.path.join(mt_dnn_root,
"%s_%s.json" % (task, split_name))
logger.info(dump_path)
build_data(rows,
dump_path,
tokenizer,
data_format,
encoderModelType=encoder_model,
task_type=task_type,
lab_dict=label_mapper)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--train_file",
default=None,
type=str,
required=True,
help="The input train corpus.")
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument(
"--on_memory",
action='store_true',
help="Whether to load train samples into memory or use disk")
parser.add_argument(
"--do_lower_case",
action='store_true',
help=
"Whether to lower case the input text. True for uncased models, False for cased models."
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
args = parser.parse_args()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train:
raise ValueError(
"Training is currently the only implemented execution option. Please set `do_train`."
)
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
#train_examples = None
num_train_optimization_steps = None
if args.do_train:
print("Loading Train Dataset", args.train_file)
train_dataset = BERTDataset(args.train_file,
tokenizer,
seq_len=args.max_seq_length,
corpus_lines=None,
on_memory=args.on_memory)
num_train_optimization_steps = int(
len(train_dataset) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
model = BertForPreTraining.from_pretrained(args.bert_model)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
if args.local_rank == -1:
train_sampler = RandomSampler(train_dataset)
else:
#TODO: check if this works with current data generator from disk that relies on file.__next__
# (it doesn't return item back by index)
train_sampler = DistributedSampler(train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
loss = model(input_ids, segment_ids, input_mask, lm_label_ids,
is_next)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Save a trained model
logger.info("** ** * Saving fine - tuned model ** ** * ")
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
def main(args):
device = torch.device(
"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("device: {}, n_gpu: {}, 16-bits training: {}".format(
device, n_gpu, args.fp16))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if args.do_train:
assert (args.train_file is not None) and (args.dev_file is not None)
if args.eval_test:
assert args.test_file is not None
else:
assert args.dev_file is not None
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
if args.do_train:
logger.addHandler(
logging.FileHandler(os.path.join(args.output_dir, "train.log"),
'w'))
else:
logger.addHandler(
logging.FileHandler(os.path.join(args.output_dir, "eval.log"),
'w'))
logger.info(args)
tokenizer = BertTokenizer.from_pretrained(args.model,
do_lower_case=args.do_lower_case)
# read query templates
query_templates = read_query_templates(normal_file=args.normal_file,
des_file=args.des_file)
if args.do_train or (not args.eval_test):
eval_examples = read_ace_examples(input_file=args.dev_file,
is_training=False)
gold_examples = read_ace_examples(input_file=args.gold_file,
is_training=False)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
query_templates=query_templates,
nth_query=args.nth_query,
is_training=False)
logger.info("***** Dev *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_if_trigger_ids = torch.tensor(
[f.if_trigger_ids for f in eval_features], dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0),
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_if_trigger_ids,
all_example_index)
eval_dataloader = DataLoader(eval_data,
batch_size=args.eval_batch_size)
if args.do_train:
train_examples = read_ace_examples(input_file=args.train_file,
is_training=True)
train_features = convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
query_templates=query_templates,
nth_query=args.nth_query,
is_training=True)
if args.train_mode == 'sorted' or args.train_mode == 'random_sorted':
train_features = sorted(train_features,
key=lambda f: np.sum(f.input_mask))
else:
random.shuffle(train_features)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_if_trigger_ids = torch.tensor(
[f.if_trigger_ids for f in train_features], dtype=torch.long)
all_start_positions = torch.tensor(
[f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor(
[f.end_position for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_if_trigger_ids,
all_start_positions, all_end_positions)
train_dataloader = DataLoader(train_data,
batch_size=args.train_batch_size)
train_batches = [batch for batch in train_dataloader]
num_train_optimization_steps = \
len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
logger.info("***** Train *****")
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
eval_step = max(1, len(train_batches) // args.eval_per_epoch)
best_result = None
lrs = [args.learning_rate] if args.learning_rate else \
[1e-6, 2e-6, 3e-6, 5e-6, 1e-5, 2e-5, 3e-5, 5e-5]
for lr in lrs:
if not args.add_if_trigger_embedding:
model = BertForQuestionAnswering.from_pretrained(
args.model, cache_dir=PYTORCH_PRETRAINED_BERT_CACHE)
else:
model = BertForQuestionAnswering_withIfTriggerEmbedding.from_pretrained(
args.model, cache_dir=PYTORCH_PRETRAINED_BERT_CACHE)
if args.fp16:
model.half()
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
param_optimizer = list(model.named_parameters())
param_optimizer = [
n for n in param_optimizer if 'pooler' not in n[0]
]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params': [
p for n, p in param_optimizer
if any(nd in n for nd in no_decay)
],
'weight_decay':
0.0
}]
optimizer = BertAdam(optimizer_grouped_parameters,
lr=lr,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
tr_loss = 0
nb_tr_examples = 0
nb_tr_steps = 0
global_step = 0
start_time = time.time()
for epoch in range(int(args.num_train_epochs)):
model.train()
logger.info("Start epoch #{} (lr = {})...".format(epoch, lr))
if args.train_mode == 'random' or args.train_mode == 'random_sorted':
random.shuffle(train_batches)
for step, batch in enumerate(train_batches):
if n_gpu == 1:
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, if_trigger_ids, start_positions, end_positions = batch
if not args.add_if_trigger_embedding:
loss = model(input_ids, segment_ids, input_mask,
start_positions, end_positions)
else:
loss = model(input_ids, segment_ids, if_trigger_ids,
input_mask, start_positions,
end_positions)
if n_gpu > 1:
loss = loss.mean()
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
global_step += 1
if (step + 1) % eval_step == 0:
save_model = False
if args.do_eval:
# result, _, _ = evaluate(args, model, device, eval_dataset, eval_dataloader, eval_examples, eval_features)
result, preds = evaluate(args, model, device,
eval_dataloader,
eval_examples,
gold_examples,
eval_features)
# import ipdb; ipdb.set_trace()
model.train()
result['global_step'] = global_step
result['epoch'] = epoch
result['learning_rate'] = lr
result['batch_size'] = args.train_batch_size
if (best_result is
None) or (result[args.eval_metric] >
best_result[args.eval_metric]):
best_result = result
save_model = True
logger.info(
'Epoch: {}, Step: {} / {}, used_time = {:.2f}s, loss = {:.6f}'
.format(epoch, step + 1,
len(train_batches),
time.time() - start_time,
tr_loss / nb_tr_steps))
# logger.info("!!! Best dev %s (lr=%s, epoch=%d): p_c: %.2f, r_c: %.2f, f1_c: %.2f" %
# (args.eval_metric, str(lr), epoch, result["prec_c"], result["recall_c"], result["f1_c"]))
logger.info(
"!!! Best dev %s (lr=%s, epoch=%d): p_c: %.2f, r_c: %.2f, f1_c: %.2f, p_i: %.2f, r_i: %.2f, f1_i: %.2f"
% (args.eval_metric, str(lr), epoch,
result["prec_c"], result["recall_c"],
result["f1_c"], result["prec_i"],
result["recall_i"], result["f1_i"]))
else:
save_model = True
if save_model:
model_to_save = model.module if hasattr(
model, 'module') else model
output_model_file = os.path.join(
args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(
args.output_dir, CONFIG_NAME)
torch.save(model_to_save.state_dict(),
output_model_file)
model_to_save.config.to_json_file(
output_config_file)
tokenizer.save_vocabulary(args.output_dir)
if best_result:
with open(
os.path.join(args.output_dir,
"eval_results.txt"),
"w") as writer:
for key in sorted(best_result.keys()):
writer.write(
"%s = %s\n" %
(key, str(best_result[key])))
if args.do_eval:
if args.eval_test:
eval_examples = read_ace_examples(input_file=args.test_file,
is_training=False)
gold_examples = read_ace_examples(input_file=args.gold_file,
is_training=False)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
query_templates=query_templates,
nth_query=args.nth_query,
is_training=False)
logger.info("***** Test *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor(
[f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in eval_features], dtype=torch.long)
all_if_trigger_ids = torch.tensor(
[f.if_trigger_ids for f in eval_features], dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0),
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_if_trigger_ids,
all_example_index)
eval_dataloader = DataLoader(eval_data,
batch_size=args.eval_batch_size)
if not args.add_if_trigger_embedding:
model = BertForQuestionAnswering.from_pretrained(args.output_dir)
else:
model = BertForQuestionAnswering_withIfTriggerEmbedding.from_pretrained(
args.output_dir)
if args.fp16:
model.half()
model.to(device)
result, preds = evaluate(args, model, device, eval_dataloader,
eval_examples, gold_examples, eval_features)
with open(os.path.join(args.output_dir, "test_results.txt"),
"w") as writer:
for key in result:
writer.write("%s = %s\n" % (key, str(result[key])))
with open(os.path.join(args.output_dir, "arg_predictions.json"),
"w") as writer:
for key in preds:
writer.write(json.dumps(preds[key], default=int) + "\n")
def main():
parser = argparse.ArgumentParser()
# Required parameters
parser.add_argument(
"--train_file",
default="data/conceptual_caption/training",
type=str,
# required=True,
help="The input train corpus.",
)
parser.add_argument(
"--validation_file",
default="data/conceptual_caption/validation",
type=str,
# required=True,
help="The input train corpus.",
)
parser.add_argument(
"--from_pretrained",
default="",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--bert_model",
default="bert-base-uncased",
type=str,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.",
)
parser.add_argument(
"--output_dir",
default="save",
type=str,
# required=True,
help=
"The output directory where the model checkpoints will be written.",
)
parser.add_argument(
"--config_file",
default="config/bert_config.json",
type=str,
# required=True,
help="The config file which specified the model details.",
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=36,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.",
)
parser.add_argument("--predict_feature",
action="store_true",
help="visual target.")
parser.add_argument(
"--train_batch_size",
default=512,
type=int,
help="Total batch size for training.",
)
parser.add_argument(
"--learning_rate",
default=1e-4,
type=float,
help="The initial learning rate for Adam.",
)
parser.add_argument(
"--num_train_epochs",
default=10.0,
type=float,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--start_epoch",
default=0,
type=float,
help="Total number of training epochs to perform.",
)
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.",
)
parser.add_argument("--img_weight",
default=1,
type=float,
help="weight for image loss")
parser.add_argument("--no_cuda",
action="store_true",
help="Whether not to use CUDA when available")
parser.add_argument(
"--on_memory",
action="store_true",
help="Whether to load train samples into memory or use disk",
)
parser.add_argument(
"--do_lower_case",
type=bool,
default=True,
help=
"Whether to lower case the input text. True for uncased models, False for cased models.",
)
parser.add_argument(
"--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus",
)
parser.add_argument("--seed",
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass.",
)
parser.add_argument(
"--fp16",
action="store_true",
help="Whether to use 16-bit float precision instead of 32-bit",
)
parser.add_argument(
"--loss_scale",
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n",
)
parser.add_argument(
"--num_workers",
type=int,
default=3,
help="Number of workers in the dataloader.",
)
parser.add_argument(
"--save_name",
default='',
type=str,
help="save name for training.",
)
parser.add_argument("--baseline",
action="store_true",
help="Wheter to use the baseline model (single bert).")
parser.add_argument(
"--freeze",
default=-1,
type=int,
help="till which layer of textual stream of vilbert need to fixed.")
parser.add_argument("--use_chuncks",
default=0,
type=float,
help="whether use chunck for parallel training.")
parser.add_argument("--distributed",
action="store_true",
help="whether use chunck for parallel training.")
parser.add_argument("--without_coattention",
action="store_true",
help="whether pair loss.")
args = parser.parse_args()
if args.baseline:
from pytorch_pretrained_bert.modeling import BertConfig
from vilbert.basebert import BertForMultiModalPreTraining
else:
from vilbert.vilbert import BertForMultiModalPreTraining, BertConfig
print(args)
if args.save_name is not '':
timeStamp = args.save_name
else:
timeStamp = strftime("%d-%b-%y-%X-%a", gmtime())
timeStamp += "_{:0>6d}".format(random.randint(0, 10e6))
savePath = os.path.join(args.output_dir, timeStamp)
if not os.path.exists(savePath):
os.makedirs(savePath)
config = BertConfig.from_json_file(args.config_file)
if args.freeze > config.t_biattention_id[0]:
config.fixed_t_layer = config.t_biattention_id[0]
if args.without_coattention:
config.with_coattention = False
# save all the hidden parameters.
with open(os.path.join(savePath, 'command.txt'), 'w') as f:
print(args, file=f) # Python 3.x
print('\n', file=f)
print(config, file=f)
bert_weight_name = json.load(
open("config/" + args.from_pretrained + "_weight_name.json", "r"))
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend="nccl")
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
num_train_optimization_steps = None
viz = TBlogger("logs", timeStamp)
train_dataset = ConceptCapLoaderTrain(
args.train_file,
tokenizer,
seq_len=args.max_seq_length,
batch_size=args.train_batch_size,
predict_feature=args.predict_feature,
num_workers=args.num_workers,
distributed=args.distributed,
)
validation_dataset = ConceptCapLoaderVal(
args.validation_file,
tokenizer,
seq_len=args.max_seq_length,
batch_size=args.train_batch_size,
predict_feature=args.predict_feature,
num_workers=2,
distributed=args.distributed,
)
num_train_optimization_steps = (
int(train_dataset.num_dataset / args.train_batch_size /
args.gradient_accumulation_steps) *
(args.num_train_epochs - args.start_epoch))
# if args.local_rank != -1:
# num_train_optimization_steps = (
# num_train_optimization_steps // torch.distributed.get_world_size()
# )
default_gpu = False
if dist.is_available() and args.distributed:
rank = dist.get_rank()
if rank == 0:
default_gpu = True
else:
default_gpu = True
# pdb.set_trace()
if args.predict_feature:
config.v_target_size = 2048
config.predict_feature = True
else:
config.v_target_size = 1601
config.predict_feature = False
if args.from_pretrained:
model = BertForMultiModalPreTraining.from_pretrained(
args.from_pretrained, config)
else:
model = BertForMultiModalPreTraining(config)
model.cuda()
if args.fp16:
model.half()
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
no_decay = ["bias", "LayerNorm.bias", "LayerNorm.weight"]
if args.freeze != -1:
bert_weight_name_filtered = []
for name in bert_weight_name:
if 'embeddings' in name:
bert_weight_name_filtered.append(name)
elif 'encoder' in name:
layer_num = name.split('.')[2]
if int(layer_num) <= args.freeze:
bert_weight_name_filtered.append(name)
optimizer_grouped_parameters = []
for key, value in dict(model.named_parameters()).items():
if key[12:] in bert_weight_name_filtered:
value.requires_grad = False
if default_gpu:
print("filtered weight")
print(bert_weight_name_filtered)
if not args.from_pretrained:
param_optimizer = list(model.named_parameters())
optimizer_grouped_parameters = [
{
"params": [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
"weight_decay":
0.01,
},
{
"params": [
p for n, p in param_optimizer
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0,
},
]
else:
optimizer_grouped_parameters = []
for key, value in dict(model.named_parameters()).items():
if value.requires_grad:
if key[12:] in bert_weight_name:
lr = args.learning_rate * 0.1
else:
lr = args.learning_rate
if any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [{
"params": [value],
"lr": lr,
"weight_decay": 0.01
}]
if not any(nd in key for nd in no_decay):
optimizer_grouped_parameters += [{
"params": [value],
"lr": lr,
"weight_decay": 0.0
}]
if default_gpu:
print(len(list(model.named_parameters())),
len(optimizer_grouped_parameters))
# set different parameters for vision branch and lanugage branch.
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(
optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0,
)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
if args.from_pretrained:
optimizer = BertAdam(
optimizer_grouped_parameters,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps,
)
else:
optimizer = BertAdam(
optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps,
)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", train_dataset.num_dataset)
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
startIterID = 0
global_step = 0
masked_loss_v_tmp = 0
masked_loss_t_tmp = 0
next_sentence_loss_tmp = 0
loss_tmp = 0
start_t = timer()
# t1 = timer()
for epochId in range(int(args.start_epoch), int(args.num_train_epochs)):
model.train()
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
# iter_dataloader = iter(train_dataloader)
for step, batch in enumerate(train_dataset):
iterId = startIterID + step + (epochId * len(train_dataset))
# batch = iter_dataloader.next()
batch = tuple(
t.cuda(device=device, non_blocking=True) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids, is_next, image_feat, image_loc, image_target, image_label, image_mask, image_ids = (
batch)
masked_loss_t, masked_loss_v, next_sentence_loss = model(
input_ids,
image_feat,
image_loc,
segment_ids,
input_mask,
image_mask,
lm_label_ids,
image_label,
image_target,
is_next,
)
if args.without_coattention:
next_sentence_loss = next_sentence_loss * 0
masked_loss_v = masked_loss_v * args.img_weight
loss = masked_loss_t + masked_loss_v + next_sentence_loss
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
masked_loss_t = masked_loss_t.mean()
masked_loss_v = masked_loss_v.mean()
next_sentence_loss = next_sentence_loss.mean()
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if math.isnan(loss.item()):
pdb.set_trace()
tr_loss += loss.item()
rank = 0
if dist.is_available() and args.distributed:
rank = dist.get_rank()
else:
rank = 0
viz.linePlot(iterId, loss.item(), "loss_" + str(rank), "train")
viz.linePlot(iterId, masked_loss_t.item(),
"masked_loss_t_" + str(rank), "train")
viz.linePlot(iterId, masked_loss_v.item(),
"masked_loss_v_" + str(rank), "train")
viz.linePlot(iterId, next_sentence_loss.item(),
"next_sentence_loss_" + str(rank), "train")
# viz.linePlot(iterId, optimizer.get_lr()[0], 'learning_rate', 'train')
loss_tmp += loss.item()
masked_loss_v_tmp += masked_loss_v.item()
masked_loss_t_tmp += masked_loss_t.item()
next_sentence_loss_tmp += next_sentence_loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion,
)
for param_group in optimizer.param_groups:
param_group["lr"] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if step % 20 == 0 and step != 0:
masked_loss_t_tmp = masked_loss_t_tmp / 20.0
masked_loss_v_tmp = masked_loss_v_tmp / 20.0
next_sentence_loss_tmp = next_sentence_loss_tmp / 20.0
loss_tmp = loss_tmp / 20.0
end_t = timer()
timeStamp = strftime("%a %d %b %y %X", gmtime())
Ep = epochId + nb_tr_steps / float(len(train_dataset))
printFormat = "[%s][Ep: %.2f][Iter: %d][Time: %5.2fs][Loss: %.5g][Loss_v: %.5g][Loss_t: %.5g][Loss_n: %.5g][LR: %.8g]"
printInfo = [
timeStamp,
Ep,
nb_tr_steps,
end_t - start_t,
loss_tmp,
masked_loss_v_tmp,
masked_loss_t_tmp,
next_sentence_loss_tmp,
optimizer.get_lr()[0],
]
start_t = end_t
print(printFormat % tuple(printInfo))
masked_loss_v_tmp = 0
masked_loss_t_tmp = 0
next_sentence_loss_tmp = 0
loss_tmp = 0
# Do the evaluation
torch.set_grad_enabled(False)
start_t = timer()
numBatches = len(validation_dataset)
eval_masked_loss_t = 0
eval_masked_loss_v = 0
eval_next_sentence_loss = 0
eval_total_loss = 0
model.eval()
for step, batch in enumerate(validation_dataset):
batch = tuple(
t.cuda(device=device, non_blocking=True) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids, is_next, image_feat, image_loc, image_target, image_label, image_mask, image_ids = (
batch)
masked_loss_t, masked_loss_v, next_sentence_loss = model(
input_ids,
image_feat,
image_loc,
segment_ids,
input_mask,
image_mask,
lm_label_ids,
image_label,
image_target,
is_next,
)
masked_loss_v = masked_loss_v * args.img_weight
loss = masked_loss_t + masked_loss_v + next_sentence_loss
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
masked_loss_t = masked_loss_t.mean()
masked_loss_v = masked_loss_v.mean()
next_sentence_loss = next_sentence_loss.mean()
eval_masked_loss_t += masked_loss_t.item()
eval_masked_loss_v += masked_loss_v.item()
eval_next_sentence_loss += next_sentence_loss.item()
eval_total_loss += loss.item()
end_t = timer()
delta_t = " Time: %5.2fs" % (end_t - start_t)
start_t = end_t
progressString = "\r Evaluating split '%s' [%d/%d]\t" + delta_t
sys.stdout.write(progressString % ('val', step + 1, numBatches))
sys.stdout.flush()
eval_masked_loss_t = eval_masked_loss_t / float(numBatches)
eval_masked_loss_v = eval_masked_loss_v / float(numBatches)
eval_next_sentence_loss = eval_next_sentence_loss / float(numBatches)
eval_total_loss = eval_total_loss / float(numBatches)
printFormat = "Evaluation: [Loss: %.5g][Loss_v: %.5g][Loss_t: %.5g][Loss_n: %.5g]"
printInfo = [
eval_total_loss, eval_masked_loss_v, eval_masked_loss_t,
eval_next_sentence_loss
]
print(printFormat % tuple(printInfo))
torch.set_grad_enabled(True)
viz.linePlot(epochId, eval_total_loss, "loss_" + str(rank), "val")
viz.linePlot(epochId, eval_masked_loss_t, "masked_loss_t_" + str(rank),
"val")
viz.linePlot(epochId, eval_masked_loss_v, "masked_loss_v_" + str(rank),
"val")
viz.linePlot(epochId, eval_next_sentence_loss,
"next_sentence_loss_" + str(rank), "val")
if default_gpu:
# Save a trained model
logger.info("** ** * Saving fine - tuned model ** ** * ")
model_to_save = (
model.module if hasattr(model, "module") else model
) # Only save the model it-self
output_model_file = os.path.join(
savePath, "pytorch_model_" + str(epochId) + ".bin")
torch.save(model_to_save.state_dict(), output_model_file)
input_ids = torch.cat(input_ids, dim=0)
attention_masks = torch.cat(attention_masks, dim=0)
labels = torch.tensor(labels.to_numpy())
dataset = TensorDataset(input_ids, attention_masks, labels)
with open(
f'{get_current_project_directory()}/../poem_clf_dataset/dataset.pkl',
'wb') as f:
pickle.dump(dataset, f)
return dataset
if __name__ == "__main__":
print('Loading BERT tokenizer...')
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased',
do_lower_case=True)
dataAll = load_data()
dataAll.Poem = preprocess_poem(dataAll.Poem)
filtered_category = preprocess_tag(dataAll)
# Get top tags
top_category = list(filtered_category.keys())
top_category = [
'Activities', 'Arts & Sciences', 'Living', 'Love', 'Nature',
'Relationships', 'Religion', 'Social Commentaries'
]
print(top_category)
dataAll.drop(columns=["Unnamed: 0", "Title", 'Poet'], inplace=True)
dataAll.dropna(inplace=True)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--config_dir',
type=str,
default='',
help="Config file for Bert")
add_quantize_arguments(parser)
args = parser.parse_args()
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mnli-mm": MnliMismatchedProcessor,
"mrpc": MrpcProcessor,
"sst-2": Sst2Processor,
"sts-b": StsbProcessor,
"qqp": QqpProcessor,
"qnli": QnliProcessor,
"rte": RteProcessor,
"wnli": WnliProcessor,
}
output_modes = {
"cola": "classification",
"mnli": "classification",
"mrpc": "classification",
"sst-2": "classification",
"sts-b": "regression",
"qqp": "classification",
"qnli": "classification",
"rte": "classification",
"wnli": "classification",
}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logging.basicConfig(
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s',
datefmt='%m/%d/%Y %H:%M:%S',
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
output_mode = output_modes[task_name]
label_list = processor.get_labels()
num_labels = len(label_list)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(
args.local_rank))
model = BertForSequenceClassification.from_pretrained(
args.bert_model,
cache_dir=cache_dir,
num_labels=num_labels,
config_dir=args.config_dir)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
if args.do_train:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
warmup_linear = WarmupLinearSchedule(
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer, output_mode)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.float)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
# define a new function to compute loss values for both output_modes
logits = model(input_ids, segment_ids, input_mask, labels=None)
if output_mode == "classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
loss = loss_fct(logits.view(-1), label_ids.view(-1))
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear.get_lr(
global_step, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if args.do_train and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
# Save a trained model, configuration and tokenizer
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
# If we save using the predefined names, we can load using `from_pretrained`
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
model_to_save.config.to_json_file(output_config_file)
tokenizer.save_vocabulary(args.output_dir)
# Load a trained model and vocabulary that you have fine-tuned
model = BertForSequenceClassification.from_pretrained(
args.output_dir, num_labels=num_labels)
tokenizer = BertTokenizer.from_pretrained(
args.output_dir, do_lower_case=args.do_lower_case)
else:
model = BertForSequenceClassification.from_pretrained(
args.bert_model, num_labels=num_labels)
model.to(device)
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer, output_mode)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
if output_mode == "classification":
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
elif output_mode == "regression":
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.float)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss = 0
nb_eval_steps = 0
preds = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
logits = model(input_ids, segment_ids, input_mask, labels=None)
# create eval loss and other metric required by the task
if output_mode == "classification":
loss_fct = CrossEntropyLoss()
tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
elif output_mode == "regression":
loss_fct = MSELoss()
tmp_eval_loss = loss_fct(logits.view(-1), label_ids.view(-1))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
preds = preds[0]
if output_mode == "classification":
preds = np.argmax(preds, axis=1)
elif output_mode == "regression":
preds = np.squeeze(preds)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
loss = tr_loss / global_step if args.do_train else None
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
# hack for MNLI-MM
if task_name == "mnli":
task_name = "mnli-mm"
processor = processors[task_name]()
if os.path.exists(args.output_dir +
'-MM') and os.listdir(args.output_dir +
'-MM') and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".
format(args.output_dir))
if not os.path.exists(args.output_dir + '-MM'):
os.makedirs(args.output_dir + '-MM')
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(
eval_examples, label_list, args.max_seq_length, tokenizer,
output_mode)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor(
[f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in eval_features], dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss = 0
nb_eval_steps = 0
preds = []
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
logits = model(input_ids,
segment_ids,
input_mask,
labels=None)
loss_fct = CrossEntropyLoss()
tmp_eval_loss = loss_fct(logits.view(-1, num_labels),
label_ids.view(-1))
eval_loss += tmp_eval_loss.mean().item()
nb_eval_steps += 1
if len(preds) == 0:
preds.append(logits.detach().cpu().numpy())
else:
preds[0] = np.append(preds[0],
logits.detach().cpu().numpy(),
axis=0)
eval_loss = eval_loss / nb_eval_steps
preds = preds[0]
preds = np.argmax(preds, axis=1)
result = compute_metrics(task_name, preds, all_label_ids.numpy())
loss = tr_loss / global_step if args.do_train else None
result['eval_loss'] = eval_loss
result['global_step'] = global_step
result['loss'] = loss
output_eval_file = os.path.join(args.output_dir + '-MM',
"eval_results.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--paragraph", default=None, type=str)
parser.add_argument("--question", default=None, type=str)
parser.add_argument("--model", default=None, type=str)
parser.add_argument("--max_seq_length", default=384, type=int)
parser.add_argument("--doc_stride", default=128, type=int)
parser.add_argument("--max_query_length", default=64, type=int)
parser.add_argument("--config_file", default=None, type=str)
parser.add_argument("--max_answer_length", default=30, type=int)
args = parser.parse_args()
para_file = args.paragraph
question_file = args.question
model_path = args.model
device = torch.device("cuda")
### Raeding paragraph
# f = open(para_file, 'r')
# para = f.read()
# f.close()
## Reading question
# f = open(ques_file, 'r')
# ques = f.read()
# f.close()
# para_list = para.split('\n\n')
f = open(para_file, "rb")
para = f.read()
para = para.decode('windows-1252')
para = para.strip("\n").replace("\r", " ").replace("\n", "")
#print(para)
# print(para)
f.close()
f_ = open(question_file, "r")
question = f_.read()
question = question.split("\n")
while "" in question:
question.remove("")
for q in question:
q = q.strip("\n")
f_.close()
input_data = []
pfinder = ParaFinder(para)
i = 0
for q in question:
closest_para = pfinder.closestParagraph(q)
paragraphs = {}
paragraphs["id"] = i
paragraphs["text"] = closest_para
paragraphs["ques"] = [q]
i += 1
input_data.append(paragraphs)
# print(input_data)
## input_data is a list of dictionary which has a paragraph and questions
examples = read_squad_examples(input_data)
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True)
eval_features = convert_examples_to_features(
examples=examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length)
all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
### Loading Pretrained model for QnA
config = BertConfig(args.config_file)
model = BertForQuestionAnswering(config)
model.load_state_dict(torch.load(model_path, map_location=device))
model.to(device)
pred_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index)
# Run prediction for full data
pred_sampler = SequentialSampler(pred_data)
pred_dataloader = DataLoader(pred_data, sampler=pred_sampler, batch_size=9)
predictions = []
for input_ids, input_mask, segment_ids, example_indices in pred_dataloader:
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
batch_start_logits, batch_end_logits = model(input_ids, segment_ids, input_mask)
features = []
example = []
all_results = []
for i, example_index in enumerate(example_indices):
start_logits = batch_start_logits[i].detach().cpu().tolist()
end_logits = batch_end_logits[i].detach().cpu().tolist()
feature = eval_features[example_index.item()]
unique_id = int(feature.unique_id)
features.append(feature)
all_results.append(RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits))
output = predict(examples, features, all_results, args.max_answer_length)
predictions.append(output)
### For printing the results ####
index = None
for example in examples:
if index != example.example_id:
# print(example.para_text)
index = example.example_id
# print('\n')
# print(colored('***********Question and Answers *************', 'red'))
ques_text = example.question_text
print(ques_text)
prediction, prob = predictions[math.floor(example.unique_id / 12)][example]
if prob > 0.35:
print(prediction)
#print(type(prediction))
else:
print("No result found")
def main():
parser = argparse.ArgumentParser()
# General
parser.add_argument("--config_path",
default=None,
type=str,
help="Bert config file path.")
parser.add_argument(
"--bert_model",
default="bert-base-cased",
type=str,
help=
"Bert pre-trained model selected in the list: bert-base-cased, bert-large-cased"
)
parser.add_argument("--model_recover_path",
default=None,
type=str,
help="The file of fine-tuned pretraining model.")
parser.add_argument('--max_position_embeddings',
type=int,
default=512,
help="max position embeddings")
# For decoding
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--amp',
action='store_true',
help="Whether to use amp for fp16")
parser.add_argument('--seed',
type=int,
default=123,
help="random seed for initialization")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument('--new_segment_ids',
action='store_true',
help="Use new segment ids for bi-uni-directional LM.")
parser.add_argument('--batch_size',
type=int,
default=4,
help="Batch size for decoding.")
parser.add_argument('--beam_size',
type=int,
default=1,
help="Beam size for searching")
parser.add_argument('--length_penalty',
type=float,
default=0,
help="Length penalty for beam search")
parser.add_argument('--forbid_duplicate_ngrams', action='store_true')
parser.add_argument('--forbid_ignore_word',
type=str,
default=None,
help="Forbid the word during forbid_duplicate_ngrams")
parser.add_argument("--min_len", default=None, type=int)
parser.add_argument('--ngram_size', type=int, default=3)
parser.add_argument('--max_tgt_length',
type=int,
default=20,
help="maximum length of target sequence")
# Others for VLP
parser.add_argument("--src_file",
default='/mnt/dat/COCO/annotations/dataset_coco.json',
type=str,
help="The input data file name.")
parser.add_argument('--dataset',
default='coco',
type=str,
help='coco | flickr30k | cc')
parser.add_argument('--len_vis_input', type=int, default=100)
# parser.add_argument('--resnet_model', type=str, default='imagenet_weights/resnet101.pth')
parser.add_argument('--image_root',
type=str,
default='/mnt/dat/COCO/images')
parser.add_argument('--split', type=str, default='val')
parser.add_argument('--drop_prob', default=0.1, type=float)
parser.add_argument('--enable_butd',
action='store_true',
help='set to take in region features')
parser.add_argument(
'--region_bbox_file',
default=
'coco_detection_vg_thresh0.2_feat_gvd_checkpoint_trainvaltest.h5',
type=str)
parser.add_argument(
'--region_det_file_prefix',
default=
'feat_cls_1000/coco_detection_vg_100dets_gvd_checkpoint_trainval',
type=str)
parser.add_argument('--file_valid_jpgs', default='', type=str)
args = parser.parse_args()
if args.enable_butd:
assert (args.len_vis_input == 100)
args.region_bbox_file = os.path.join(args.image_root,
args.region_bbox_file)
args.region_det_file_prefix = os.path.join(
args.image_root, args.region_det_file_prefix) if args.dataset in (
'cc', 'coco') and args.region_det_file_prefix != '' else ''
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
# fix random seed
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
args.max_seq_length = args.max_tgt_length + args.len_vis_input + 3 # +3 for 2x[SEP] and [CLS]
tokenizer.max_len = args.max_seq_length
bi_uni_pipeline = []
bi_uni_pipeline.append(
seq2seq_loader.Preprocess4Seq2seqDecoder(
list(tokenizer.vocab.keys()),
tokenizer.convert_tokens_to_ids,
args.max_seq_length,
max_tgt_length=args.max_tgt_length,
new_segment_ids=args.new_segment_ids,
mode='s2s',
len_vis_input=args.len_vis_input,
enable_butd=args.enable_butd,
region_bbox_file=args.region_bbox_file,
region_det_file_prefix=args.region_det_file_prefix))
amp_handle = None
if args.fp16 and args.amp:
from apex import amp
amp_handle = amp.init(enable_caching=True)
logger.info("enable fp16 with amp")
# Prepare model
cls_num_labels = 2
type_vocab_size = 6 if args.new_segment_ids else 2
mask_word_id, eos_word_ids = tokenizer.convert_tokens_to_ids(
["[MASK]", "[SEP]"])
forbid_ignore_set = None
if args.forbid_ignore_word:
w_list = []
for w in args.forbid_ignore_word.split('|'):
if w.startswith('[') and w.endswith(']'):
w_list.append(w.upper())
else:
w_list.append(w)
forbid_ignore_set = set(tokenizer.convert_tokens_to_ids(w_list))
print(args.model_recover_path)
for model_recover_path in glob.glob(args.model_recover_path.strip()):
logger.info("***** Recover model: %s *****", model_recover_path)
model_recover = torch.load(model_recover_path)
model = BertForSeq2SeqDecoder.from_pretrained(
args.bert_model,
max_position_embeddings=args.max_position_embeddings,
config_path=args.config_path,
state_dict=model_recover,
num_labels=cls_num_labels,
type_vocab_size=type_vocab_size,
task_idx=3,
mask_word_id=mask_word_id,
search_beam_size=args.beam_size,
length_penalty=args.length_penalty,
eos_id=eos_word_ids,
forbid_duplicate_ngrams=args.forbid_duplicate_ngrams,
forbid_ignore_set=forbid_ignore_set,
ngram_size=args.ngram_size,
min_len=args.min_len,
enable_butd=args.enable_butd,
len_vis_input=args.len_vis_input)
del model_recover
# from vlp.resnet import resnet
# cnn = resnet(args.resnet_model, _num_layers=101, _fixed_block=4, pretrained=True) # no finetuning
if args.fp16:
model.half()
# cnn.half()
model.to(device)
# cnn.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
# cnn = torch.nn.DataParallel(cnn)
torch.cuda.empty_cache()
model.eval()
# cnn.eval()
eval_lst = []
#pdb.set_trace()
gpv_image_ids = json.load(
open(
'/home/amitak/data/learning_phase_data/coco_captions/gpv_split/image_ids_val.json',
'r'))[:100]
gpv_image_ids = {imid: 1 for imid in gpv_image_ids}
img_to_cap = {}
with open(args.src_file, "r", encoding='utf-8') as f_src:
img_dat = json.load(f_src)['images']
img_idx = 0
valid_jpgs = None #if (args.file_valid_jpgs == '' or args.dataset in \
# ('coco', 'flickr30k')) else json.load(open(args.file_valid_jpgs))
# You're just reading the image, not each caption
for src in img_dat:
#if src['split'] == args.split and (valid_jpgs is None or src['filename'] in valid_jpgs):
image_id = int(src['filename'].split('_')[2][:-4])
if image_id in gpv_image_ids:
img_to_cap[image_id] = src['sentids']
if args.enable_butd:
src_tk = os.path.join(args.image_root,
src.get('filepath', 'trainval'),
src['filename'][:-4] + '.npy')
else:
src_tk = os.path.join(args.image_root,
src.get('filepath', 'trainval'),
src['filename'])
if args.dataset == 'coco':
imgid = int(src['filename'].split('_')[2][:-4])
elif args.dataset == 'cc':
imgid = int(src['imgid'])
elif args.dataset == 'flickr30k':
imgid = int(src['filename'].split('.')[0])
eval_lst.append(
(img_idx, imgid, src_tk)) # id and path for COCO
img_idx += 1
input_lines = eval_lst
next_i = 0
output_lines = [""] * len(input_lines)
total_batch = math.ceil(len(input_lines) / args.batch_size)
print('start the caption evaluation...')
with tqdm(total=total_batch) as pbar:
while next_i < len(input_lines):
_chunk = input_lines[next_i:next_i + args.batch_size]
buf_id = [x[0] for x in _chunk]
buf = [x[2] for x in _chunk]
next_i += args.batch_size
instances = []
for instance in [(x, args.len_vis_input) for x in buf]:
for proc in bi_uni_pipeline:
instances.append(proc(instance))
with torch.no_grad():
batch = batch_list_to_batch_tensors(instances)
batch = [t.to(device) for t in batch]
input_ids, token_type_ids, position_ids, input_mask, task_idx, img, vis_pe = batch
if args.fp16:
img = img.half()
vis_pe = vis_pe.half()
if args.enable_butd:
conv_feats = img.data # Bx100x2048
vis_pe = vis_pe.data
else:
conv_feats, _ = cnn(img.data) # Bx2048x7x7
conv_feats = conv_feats.view(conv_feats.size(0),
conv_feats.size(1),
-1).permute(
0, 2, 1).contiguous()
traces = model(conv_feats,
vis_pe,
input_ids,
token_type_ids,
position_ids,
input_mask,
task_idx=task_idx)
if args.beam_size > 1:
traces = {k: v.tolist() for k, v in traces.items()}
output_ids = traces['pred_seq']
else:
output_ids = traces[0].tolist()
for i in range(len(buf)):
w_ids = output_ids[i]
output_buf = tokenizer.convert_ids_to_tokens(w_ids)
output_tokens = []
for t in output_buf:
if t in ("[SEP]", "[PAD]"):
break
output_tokens.append(t)
output_sequence = ' '.join(detokenize(output_tokens))
output_lines[buf_id[i]] = output_sequence
pbar.update(1)
# This is all I need:
pdb.set_trace()
predictions = [{
'image_id': tup[1],
'caption': output_lines[img_idx]
} for img_idx, tup in enumerate(input_lines)]
gpv_caption_ids = json.load(
open(
'/home/amitak/data/learning_phase_data/coco_captions/gpv_split/cap_ids_val.json',
'r'))
final_predictions = {
c: {
'answer': p['caption']
}
for p in predictions for c in img_to_cap[p['image_id']]
if c in gpv_caption_ids
}
json.dump(final_predictions, open('vlp_preds.json', 'w'))
"""
new_predictions = []
for p in predictions:
num_caps = len([c for c in img_to_cap[p['image_id']] if c in gpv_caption_ids])
new_predictions.extend([p]*num_caps)
"""
lang_stats = language_eval(
args.dataset, predictions,
args.model_recover_path.split('/')[-2] + '-' + args.split + '-' +
args.model_recover_path.split('/')[-1].split('.')[-2], args.split)
device = torch.device('cuda:0')
sep = ' [SEP] '
samples = []
with open('../samples/QLP/SampleData.tsv', 'r', encoding='utf8') as f:
for line in f:
line = line.strip()
index, title1, title2, description, _, _, url, doc = line.split(
'\t')
samples.append(
[index, title1 + sep + title2 + sep + description, url, doc])
# json_dump(samples, '../samples/QLP/ValidationData_after.json')
# samples = json_load('../samples/QLP/ValidationData_after.json')
epoch_output_dir = '/home/work/waka/projects/pytorch-pretrained-BERT/samples/QLP/result/base/fullAd/Random/2parts/epoch2'
model = BertForSequenceClassification.from_pretrained(
epoch_output_dir, num_labels=2).to(device)
tokenizer = BertTokenizer.from_pretrained(epoch_output_dir,
do_lower_case=True)
def pre_in_batch_eval(raw):
index = []
title = []
url = []
doc = []
input_ids = []
input_mask = []
segment_ids = []
for sample in raw:
index.append(sample[0])
title.append(sample[1])
url.append(sample[2])
doc.append(sample[3])
one_input_ids, one_input_mask, one_segment_ids = preprocess(
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
default=False,
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--optimize_on_cpu',
default=False,
action='store_true',
help=
"Whether to perform optimization and keep the optimizer averages on CPU"
)
parser.add_argument(
'--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=128,
help=
'Loss scaling, positive power of 2 values can improve fp16 convergence.'
)
args = parser.parse_args()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"temporal": TemporalProcessor,
}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
if args.fp16:
logger.info(
"16-bits training currently not supported in distributed training"
)
args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info("device %s n_gpu %d distributed training %r", device, n_gpu,
bool(args.local_rank != -1))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = BertForSequenceClassification.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
if args.fp16:
param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
for n, param in model.named_parameters()]
elif args.optimize_on_cpu:
param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay_rate':
0.0
}]
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=t_total)
global_step = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16 or args.optimize_on_cpu:
if args.fp16 and args.loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
if param.grad is not None:
param.grad.data = param.grad.data / args.loss_scale
is_nan = set_optimizer_params_grad(
param_optimizer,
model.named_parameters(),
test_nan=True)
if is_nan:
logger.info(
"FP16 TRAINING: Nan in gradients, reducing loss scaling"
)
args.loss_scale = args.loss_scale / 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(
model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad()
global_step += 1
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
total_prints = []
for input_ids, input_mask, segment_ids, label_ids in eval_dataloader:
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy, prints = accuracy(logits, label_ids)
total_prints += prints
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': tr_loss / nb_tr_steps
}
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
output_print_file = os.path.join(args.output_dir, "eval_outputs.txt")
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results *****")
for key in sorted(result.keys()):
logger.info(" %s = %s", key, str(result[key]))
writer.write("%s = %s\n" % (key, str(result[key])))
with open(output_print_file, "w") as writer:
for l in total_prints:
writer.write(l + "\n")
def main():
parser = ArgumentParser()
parser.add_argument('--pregenerated_data', type=Path, required=True)
parser.add_argument('--output_dir', type=Path, required=True)
parser.add_argument(
"--bert_model",
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument("--do_lower_case", action="store_true")
parser.add_argument(
"--reduce_memory",
action="store_true",
help=
"Store training data as on-disc memmaps to massively reduce memory usage"
)
parser.add_argument("--epochs",
type=int,
default=3,
help="Number of epochs to train for")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
args = parser.parse_args()
assert args.pregenerated_data.is_dir(), \
"--pregenerated_data should point to the folder of files made by pregenerate_training_data.py!"
samples_per_epoch = []
for i in range(args.epochs):
epoch_file = args.pregenerated_data / f"epoch_{i}.json"
metrics_file = args.pregenerated_data / f"epoch_{i}_metrics.json"
if epoch_file.is_file() and metrics_file.is_file():
metrics = json.loads(metrics_file.read_text())
samples_per_epoch.append(metrics['num_training_examples'])
else:
if i == 0:
exit("No training data was found!")
print(
f"Warning! There are fewer epochs of pregenerated data ({i}) than training epochs ({args.epochs})."
)
print(
"This script will loop over the available data, but training diversity may be negatively impacted."
)
num_data_epochs = i
break
else:
num_data_epochs = args.epochs
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logging.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if args.output_dir.is_dir() and list(args.output_dir.iterdir()):
logging.warning(
f"Output directory ({args.output_dir}) already exists and is not empty!"
)
args.output_dir.mkdir(parents=True, exist_ok=True)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
total_train_examples = 0
for i in range(args.epochs):
# The modulo takes into account the fact that we may loop over limited epochs of data
total_train_examples += samples_per_epoch[i % len(samples_per_epoch)]
num_train_optimization_steps = int(total_train_examples /
args.train_batch_size /
args.gradient_accumulation_steps)
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
model = BertForPreTraining.from_pretrained(args.bert_model)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
logging.info("***** Running training *****")
logging.info(f" Num examples = {total_train_examples}")
logging.info(" Batch size = %d", args.train_batch_size)
logging.info(" Num steps = %d", num_train_optimization_steps)
model.train()
for epoch in range(args.epochs):
epoch_dataset = PregeneratedDataset(
epoch=epoch,
training_path=args.pregenerated_data,
tokenizer=tokenizer,
num_data_epochs=num_data_epochs)
if args.local_rank == -1:
train_sampler = RandomSampler(epoch_dataset)
else:
train_sampler = DistributedSampler(epoch_dataset)
train_dataloader = DataLoader(epoch_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
with tqdm(total=len(train_dataloader), desc=f"Epoch {epoch}") as pbar:
for step, batch in enumerate(train_dataloader):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
loss = model(input_ids, segment_ids, input_mask, lm_label_ids,
is_next)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
pbar.update(1)
mean_loss = tr_loss * args.gradient_accumulation_steps / nb_tr_steps
pbar.set_postfix_str(f"Loss: {mean_loss:.5f}")
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
try:
optimizer.step()
optimizer.zero_grad()
global_step += 1
except:
pass
# Save a trained model
logging.info("** ** * Saving fine-tuned model ** ** * ")
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = args.output_dir / "pytorch_model.bin"
torch.save(model_to_save.state_dict(), str(output_model_file))
def main(args):
# set up logging and device
args.save_dir = utils.get_save_dir(args.save_dir, args.name, training=True)
logger = utils.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError(
"At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified."
)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory () already exists and is not empty.")
os.makedirs(args.output_dir, exist_ok=True)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
# Generating the dictionaries
dep_dict, pos_dict, ent_dict, total_features = generate_dictionary(
args.train_ling_features_file, args.eval_ling_features_file,
args.test_ling_features_file)
# from IPython import embed; embed()
# Generating total_dictionary
total_dict = convert_string_features_to_array(total_features, dep_dict,
pos_dict, ent_dict)
# from IPython import embed; embed()
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = read_squad_examples(
input_file=args.train_file,
is_training=True,
version_2_with_negative=args.version_2_with_negative,
total_dictionary=total_dict)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
# Prepare model
model = BertForQuestionAnsweringLing.from_pretrained(
args.bert_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(args.local_rank))
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizer import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
# load training features
cached_train_features_file = args.train_file + '_{0}_{1}_{2}_{3}'.format(
list(filter(None, args.bert_model.split('/'))).pop(),
str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length))
train_features = None
print(cached_train_features_file)
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info(" Saving train features into cached file %s",
cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
# load eval features
eval_examples = read_squad_examples(
input_file=args.predict_file,
is_training=False,
version_2_with_negative=args.version_2_with_negative,
total_dictionary=total_dict)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
test_examples = read_squad_examples(
input_file=args.test_file,
is_training=False,
version_2_with_negative=args.version_2_with_negative,
total_dictionary=total_dict)
test_features = convert_examples_to_features(
examples=test_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_start_positions = torch.tensor(
[f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor(
[f.end_position for f in train_features], dtype=torch.long)
# from IPython import embed; embed()
all_ling_features = torch.tensor(
[f.ling_features for f in train_features], dtype=torch.float)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_ling_features,
all_start_positions, all_end_positions)
steps_till_eval = args.eval_steps
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
best_F1 = 0
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
if n_gpu == 1:
batch = tuple(
t.to(device)
for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, ling_features, start_positions, end_positions = batch
# from IPython import embed; embed()
loss = model(input_ids, segment_ids, input_mask, ling_features,
start_positions, end_positions)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used and handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
# add to tensorboard
loss_val = loss.item()
tbx.add_scalar('train/NLL', loss_val, global_step)
tbx.add_scalar('train/LR', optimizer.param_groups[0]['lr'],
global_step)
steps_till_eval -= args.train_batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
logger.info('Evaluating at step {}...'.format(step))
# ema.assign(model)
results, _ = evaluate(model, eval_examples, eval_features,
device, args, logger,
args.version_2_with_negative,
args.dev_eval_file)
# saver.save(step, model, results[args.metric_name], device)
# ema.resume(model)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
logger.info('Dev {}'.format(results_str))
# Log to TensorBoard
logger.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar('dev/{}'.format(k), v, global_step)
"""
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
"""
if results['F1'] > best_F1:
best_F1 = results['F1']
model_to_save = model.module if hasattr(
model,
'module') else model # Only save the model it-self
output_model_file = os.path.join(
args.output_dir, "pytorch_model_best.bin")
torch.save(model_to_save.state_dict(),
output_model_file)
#model.to(device)
# Save a trained model
"""
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
if args.do_train:
torch.save(model_to_save.state_dict(), output_model_file)
# Load a trained model that you have fine-tuned
model_state_dict = torch.load(output_model_file)
model = BertForQuestionAnsweringLing.from_pretrained(args.bert_model, state_dict=model_state_dict)
else:
model = BertForQuestionAnsweringLing.from_pretrained(args.bert_model)
model.to(device)
"""
# load the best trained model and eval on the eval set and test set
best_model_file = os.path.join(args.output_dir, "pytorch_model_best.bin")
model_state_dict = torch.load(best_model_file)
model = BertForQuestionAnsweringLing.from_pretrained(
args.bert_model, state_dict=model_state_dict)
model.to(device)
if args.do_predict and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
logger.info('Evaluating at the best model')
results, all_results = evaluate(model, eval_examples, eval_features,
device, args, logger,
args.version_2_with_negative,
args.dev_eval_file)
logger.info('Write the best eval results')
output_prediction_file = os.path.join(args.output_dir,
"predictions.json")
output_nbest_file = os.path.join(args.output_dir,
"nbest_predictions.json")
output_null_log_odds_file = os.path.join(args.output_dir,
"null_odds.json")
write_predictions(eval_examples, eval_features, all_results,
args.n_best_size, args.max_answer_length,
args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file,
args.verbose_logging, args.version_2_with_negative,
args.null_score_diff_threshold, 'dev')
logger.info('Test set at the best model')
results, all_results = evaluate(model, test_examples, test_features,
device, args, logger,
args.version_2_with_negative,
args.test_eval_file)
logger.info('Write the best test set results')
output_prediction_file = os.path.join(args.output_dir,
"predictions_test.json")
output_nbest_file = os.path.join(args.output_dir,
"nbest_predictions_test.json")
output_null_log_odds_file = os.path.join(args.output_dir,
"null_odds_test.json")
write_predictions(test_examples, test_features, all_results,
args.n_best_size, args.max_answer_length,
args.do_lower_case, output_prediction_file,
output_nbest_file, output_null_log_odds_file,
args.verbose_logging, args.version_2_with_negative,
args.null_score_diff_threshold, 'test')
"""
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--eval_file",
default=None,
type=str,
required=True,
help="The input eval corpus.")
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument(
"--output_file",
default=None,
type=str,
required=True,
help="The output file where the scores will be written.")
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=3e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument(
"--on_memory",
action='store_true',
default=True,
help="Whether to load train samples into memory or use disk")
parser.add_argument(
"--do_lower_case",
action='store_true',
help=
"Whether to lower case the input text. True for uncased models, False for cased models."
)
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument("--test_run",
action='store_true',
default=False,
help="If true, shortcut the input data.")
args = parser.parse_args()
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased",
do_lower_case=True)
# Load eval_data
eval_dataset_answerable = BERTDataset(args.eval_file,
tokenizer,
seq_len=args.max_seq_length,
on_memory=args.on_memory,
answerable=True)
eval_dataset_unanswerable = BERTDataset(args.eval_file,
tokenizer,
seq_len=args.max_seq_length,
on_memory=args.on_memory,
answerable=False)
# Prepare model
if n_gpu > 0:
model_state_dict = torch.load(args.bert_model)
else:
model_state_dict = torch.load(args.bert_model, map_location='cpu')
context_model = BertModel.from_pretrained(
"bert-base-uncased") #, state_dict=model_state_dict)
question_model = BertModel.from_pretrained(
"bert-base-uncased") #, state_dict=model_state_dict)
context_model.to(device)
question_model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
context_model = DDP(context_model)
question_model = DDP(question_model)
elif n_gpu > 1:
context_model = torch.nn.DataParallel(context_model)
question_model = torch.nn.DataParallel(question_model)
# Prepare optimizer
print("Checking the vocab size:", len(tokenizer.vocab))
# 768 is bert hidden size, 256 is GRU hidden size, 1 is the layers in the GRU
model = RNNModel("GRU",
len(tokenizer.vocab),
768,
768,
1,
context_model,
question_model,
ngpu=n_gpu)
model.load_state_dict(model_state_dict)
model.to(device)
# eval loader
eval_sampler_ans = SequentialSampler(eval_dataset_answerable)
eval_dataloader_ans = DataLoader(eval_dataset_answerable,
sampler=eval_sampler_ans,
batch_size=args.train_batch_size)
eval_sampler_unans = SequentialSampler(eval_dataset_unanswerable)
eval_dataloader_unans = DataLoader(eval_dataset_unanswerable,
sampler=eval_sampler_unans,
batch_size=args.train_batch_size)
criterion = nn.CrossEntropyLoss()
model.init_hidden(args.train_batch_size)
with torch.no_grad():
model.eval()
with open(args.output_file, "w") as handle:
loss_writer = csv.writer(handle, delimiter=',')
eval_loss_ans = 0
for batch_i, eval_batch in enumerate(eval_dataloader_ans):
if batch_i % 1000 == 0:
print("#### DANITER completed answerable", batch_i)
eids = eval_batch[-1]
eval_batch = tuple(t.to(device) for t in eval_batch[:-1])
question_ids, question_mask, context_ids, context_mask, targets = eval_batch
output, _ = model(context_ids, context_mask, question_ids,
question_mask)
loss = criterion(output.view(-1, len(tokenizer.vocab)),
question_ids.view(-1))
loss_writer.writerow([eids[0], loss.item(), "ANS"])
eval_loss_ans += loss.item()
print("##### DANITER EVAL LOSS IS (ANSWERABLE) : ", eval_loss_ans)
eval_loss_unans = 0
for batch_i, eval_batch in enumerate(eval_dataloader_unans):
if batch_i % 1000 == 0:
print("#### DANITER completed unanswerable", batch_i)
eids = eval_batch[-1]
eval_batch = tuple(t.to(device) for t in eval_batch[:-1])
question_ids, question_mask, context_ids, context_mask, targets = eval_batch
output, _ = model(context_ids, context_mask, question_ids,
question_mask)
loss = criterion(output.view(-1, len(tokenizer.vocab)),
question_ids.view(-1))
loss_writer.writerow([eids[0], loss.item(), "UNANS"])
eval_loss_unans += loss.item()
print("##### DANITER EVAL LOSS IS (UNANSWERABLE) : ",
eval_loss_unans)
if len(args.train_from) > 0:
print("loading model from", args.train_from)
saved_stuff = torch.load(args.train_from)
saved_args = saved_stuff["opt"]
model = TagWordModel(args)
model.load_state_dict(saved_stuff["state_dict"])
else:
model = TagWordModel(args)
model = model.to(device)
nosplits = ("[UNK]", "[SEP]", "[PAD]", "[CLS]", "[MASK]")
if args.nosplit_parenth:
nosplits = nosplits + ("-LPR-", "-RPR-")
tokenizer = BertTokenizer.from_pretrained(
args.bert_model, do_lower_case=args.lower, never_split=nosplits,
cache_dir=CACHEDIR)
print("make sure you agree with never_splits!")
if args.load_saved_db and args.db_fi is not None:
print("loading db from", args.db_fi)
sentdb = data.SentDB(None, None, None, path=args.db_fi, align_strat=args.align_strat,
subsample_all=args.subsample_all)
else:
sentdb = data.SentDB(args.sent_fi, args.tag_fi, tokenizer, args.val_sent_fi,
args.val_tag_fi, lower=args.lower,
align_strat=args.align_strat, subsample_all=args.subsample_all)
nebert = model.bert
if args.zero_shot and "newne" not in args.just_eval:
nebert = BertModel.from_pretrained(args.bert_model, cache_dir=CACHEDIR)
def main():
parser = get_argument_parser()
deepspeed.init_distributed(dist_backend='nccl')
# Include DeepSpeed configuration arguments
parser = deepspeed.add_config_arguments(parser)
args = parser.parse_args()
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = read_squad_examples(input_file=args.train_file,
is_training=True)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
# model = BertForQuestionAnswering.from_pretrained(args.bert_model,
# cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank))
# Support for word embedding padding checkpoints
# Prepare model
bert_model_config = {
"vocab_size_or_config_json_file": 119547,
"hidden_size": 1024,
"num_hidden_layers": 24,
"num_attention_heads": 16,
"intermediate_size": 4096,
"hidden_act": "gelu",
"hidden_dropout_prob": args.dropout,
"attention_probs_dropout_prob": args.dropout,
"hidden_dropout_prob": 0.1,
"attention_probs_dropout_prob": 0.1,
"max_position_embeddings": 512,
"type_vocab_size": 2,
"initializer_range": 0.02
}
if args.ckpt_type == "DS":
if args.preln:
bert_config = BertConfigPreLN(**bert_model_config)
else:
bert_config = BertConfig(**bert_model_config)
else:
# Models from Tensorflow and Huggingface are post-LN.
if args.preln:
raise ValueError("Should NOT use --preln if the loading checkpoint doesn't use pre-layer-norm.")
# Use the original bert config if want to load from non-DeepSpeed checkpoint.
if args.origin_bert_config_file is None:
raise ValueError("--origin_bert_config_file is required for loading non-DeepSpeed checkpoint.")
bert_config = BertConfig.from_json_file(args.origin_bert_config_file)
if bert_config.vocab_size != len(tokenizer.vocab):
raise ValueError("vocab size from original checkpoint mismatch.")
bert_config.vocab_size = len(tokenizer.vocab)
# Padding for divisibility by 8
if bert_config.vocab_size % 8 != 0:
vocab_diff = 8 - (bert_config.vocab_size % 8)
bert_config.vocab_size += vocab_diff
if args.preln:
model = BertForQuestionAnsweringPreLN(bert_config, args)
else:
model = BertForQuestionAnswering(bert_config, args)
print("VOCAB SIZE:", bert_config.vocab_size)
if args.model_file is not "0":
logger.info(f"Loading Pretrained Bert Encoder from: {args.model_file}")
if args.ckpt_type == "DS":
checkpoint_state_dict = torch.load(args.model_file,
map_location=torch.device("cpu"))
if 'module' in checkpoint_state_dict:
logger.info('Loading DeepSpeed v2.0 style checkpoint')
model.load_state_dict(checkpoint_state_dict['module'],
strict=False)
elif 'model_state_dict' in checkpoint_state_dict:
model.load_state_dict(checkpoint_state_dict['model_state_dict'],
strict=False)
else:
raise ValueError("Unable to find model state in checkpoint")
else:
from convert_bert_ckpt_to_deepspeed import convert_ckpt_to_deepspeed
convert_ckpt_to_deepspeed(model, args.ckpt_type, args.model_file, vocab_diff, args.deepspeed_transformer_kernel)
logger.info(f"Pretrained Bert Encoder Loaded from: {args.model_file}")
# Prepare optimizer
param_optimizer = list(model.named_parameters())
# hack to remove pooler, which is not used
# thus it produce None grad that break apex
param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
},{
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
model, optimizer, _, _ = deepspeed.initialize(
args=args,
model=model,
model_parameters=optimizer_grouped_parameters,
dist_init_required=True)
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
#torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_predict:
raise ValueError(
"At least one of `do_train` or `do_predict` must be True.")
if args.do_train:
if not args.train_file:
raise ValueError(
"If `do_train` is True, then `train_file` must be specified.")
if args.do_predict:
if not args.predict_file:
raise ValueError(
"If `do_predict` is True, then `predict_file` must be specified."
)
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
os.makedirs(args.output_dir, exist_ok=True)
# Prepare Summary writer
if torch.distributed.get_rank() == 0 and args.job_name is not None:
args.summary_writer = get_summary_writer(name=args.job_name,
base=args.output_dir)
else:
args.summary_writer = None
logger.info("propagate deepspeed-config settings to client settings")
args.train_batch_size = model.train_micro_batch_size_per_gpu()
args.gradient_accumulation_steps = model.gradient_accumulation_steps()
args.fp16 = model.fp16_enabled()
args.print_steps = model.steps_per_print()
args.learning_rate = model.get_lr()[0]
args.wall_clock_breakdown = model.wall_clock_breakdown()
t_total = num_train_steps
if args.local_rank != -1:
t_total = t_total // torch.distributed.get_world_size()
global_step = 0
if args.do_train:
cached_train_features_file = args.train_file + '_{0}_{1}_{2}_{3}'.format(
list(filter(None, args.bert_model.split('/'))).pop(),
str(args.max_seq_length), str(args.doc_stride),
str(args.max_query_length))
train_features = None
try:
with open(cached_train_features_file, "rb") as reader:
train_features = pickle.load(reader)
except:
train_features = convert_examples_to_features(
examples=train_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=True)
if args.local_rank == -1 or torch.distributed.get_rank() == 0:
logger.info(" Saving train features into cached file %s",
cached_train_features_file)
with open(cached_train_features_file, "wb") as writer:
pickle.dump(train_features, writer)
logger.info("***** Running training *****")
logger.info(" Num orig examples = %d", len(train_examples))
logger.info(" Num split examples = %d", len(train_features))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_start_positions = torch.tensor(
[f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor(
[f.end_position for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_start_positions,
all_end_positions)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
ema_loss = 0.
sample_count = 0
num_epoch = 0
all_step_time = 0.0
ave_rounds = 20
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
num_epoch += 1
epoch_step = 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration", smoothing=0)):
start_time = time.time()
bs_size = batch[0].size()[0]
if n_gpu == 1:
batch = tuple(
t.to(device)
for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, start_positions, end_positions = batch
loss = model(input_ids, segment_ids, input_mask,
start_positions, end_positions)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
ema_loss = args.loss_plot_alpha * ema_loss + (
1 - args.loss_plot_alpha) * loss.item()
model.backward(loss)
loss_item = loss.item() * args.gradient_accumulation_steps
loss = None
sample_count += (args.train_batch_size *
torch.distributed.get_world_size())
if (step + 1) % args.gradient_accumulation_steps == 0:
# modify learning rate with special warm up BERT uses
lr_this_step = args.learning_rate * warmup_linear(
global_step / t_total, args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
model.step()
global_step += 1
epoch_step += 1
if torch.distributed.get_rank(
) == 0 and args.summary_writer:
summary_events = [
(f'Train/Steps/lr', lr_this_step, global_step),
(f'Train/Samples/train_loss', loss_item,
sample_count),
(f'Train/Samples/lr', lr_this_step, sample_count),
(f'Train/Samples/train_ema_loss', ema_loss,
sample_count)
]
if args.fp16 and hasattr(optimizer, 'cur_scale'):
summary_events.append(
(f'Train/Samples/scale', optimizer.cur_scale,
sample_count))
write_summary_events(args.summary_writer,
summary_events)
args.summary_writer.flush()
if torch.distributed.get_rank() == 0 and (
step + 1) % args.print_steps == 0:
logger.info(
f"bert_squad_progress: step={global_step} lr={lr_this_step} loss={ema_loss}"
)
else:
model.step()
if is_time_to_exit(args=args,
epoch_steps=epoch_step,
global_steps=global_step):
logger.info(
f'Warning: Early epoch termination due to max steps limit, epoch step ={epoch_step}, global step = {global_step}, epoch = {num_epoch}'
)
break
one_step_time = time.time() -start_time
all_step_time += one_step_time
if (step + 1)%(ave_rounds) == 0 and torch.distributed.get_rank() == 0:
print('At Step {}, Averaged Throughput for {} rounds is: {} Samples/s'.format(step, ave_rounds, bs_size * ave_rounds * torch.distributed.get_world_size() / all_step_time ), flush=True )
all_step_time = 0.0
# Save a trained model
# model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
#output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
# if args.do_train:
# torch.save(model_to_save.state_dict(), output_model_file)
# Load a trained model that you have fine-tuned
#model_state_dict = torch.load(output_model_file)
#model = BertForQuestionAnswering.from_pretrained(args.bert_model, state_dict=model_state_dict)
# model.to(device)
if args.do_predict and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = read_squad_examples(input_file=args.predict_file,
is_training=False)
eval_features = convert_examples_to_features(
examples=eval_examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
doc_stride=args.doc_stride,
max_query_length=args.max_query_length,
is_training=False)
logger.info("***** Running predictions *****")
logger.info(" Num orig examples = %d", len(eval_examples))
logger.info(" Num split examples = %d", len(eval_features))
logger.info(" Batch size = %d", args.predict_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0),
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_example_index)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.predict_batch_size)
model.eval()
all_results = []
logger.info("Start evaluating")
for input_ids, input_mask, segment_ids, example_indices in tqdm(
eval_dataloader, desc="Evaluating"):
if len(all_results) % 1000 == 0:
logger.info("Processing example: %d" % (len(all_results)))
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
with torch.no_grad():
batch_start_logits, batch_end_logits = model(
input_ids, segment_ids, input_mask)
for i, example_index in enumerate(example_indices):
start_logits = batch_start_logits[i].detach().cpu().tolist()
end_logits = batch_end_logits[i].detach().cpu().tolist()
eval_feature = eval_features[example_index.item()]
unique_id = int(eval_feature.unique_id)
all_results.append(
RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits))
output_prediction_file = os.path.join(args.output_dir,
"predictions.json")
output_nbest_file = os.path.join(args.output_dir,
"nbest_predictions.json")
write_predictions(eval_examples, eval_features, all_results,
args.n_best_size, args.max_answer_length,
args.do_lower_case, output_prediction_file,
output_nbest_file, args.verbose_logging)
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese."
)
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model checkpoints will be written."
)
parser.add_argument("--saved_model_path",
default=None,
type=str,
help="Saved model path")
parser.add_argument("--model_architecture",
default="basic",
type=str,
help="What model architecture to use on top of BERT")
## Other parameters
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
default=False,
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
default=False,
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_test",
default=False,
action='store_true',
help="Whether to run eval on the test set.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=64,
type=int,
help="Total batch size for eval.")
parser.add_argument("--eval_every",
default=1000,
type=int,
help="evaluate at every eval_every number of steps ")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
default=False,
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumualte before performing a backward/update pass."
)
parser.add_argument(
'--optimize_on_cpu',
default=False,
action='store_true',
help=
"Whether to perform optimization and keep the optimizer averages on CPU"
)
parser.add_argument(
'--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--weight_loss',
default=False,
action='store_true',
help="Whether to weight class losses")
parser.add_argument(
'--loss_scale',
type=float,
default=128,
help=
'Loss scaling, positive power of 2 values can improve fp16 convergence.'
)
parser.add_argument("--metric",
default=None,
type=str,
help="Which metric to use.")
parser.add_argument(
"--model_type",
default='ch',
type=str,
help="en, ch or en-ch depending on which language(s) to be used")
args = parser.parse_args()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"wsdm": WSDMFakeNewsProcessor,
}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
if args.fp16:
logger.info(
"16-bits training currently not supported in distributed training"
)
args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info("device %s n_gpu %d distributed training %r", device, n_gpu,
bool(args.local_rank != -1))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size /
args.gradient_accumulation_steps)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if args.do_train and os.path.exists(args.output_dir) and os.listdir(
args.output_dir):
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
os.makedirs(args.output_dir, exist_ok=True)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
label_list = processor.get_labels()
bert_model_ch = 'bert-base-chinese'
bert_model_en = 'bert-base-uncased'
en_tokenizer = BertTokenizer.from_pretrained(bert_model_en)
ch_tokenizer = BertTokenizer.from_pretrained(bert_model_ch)
train_examples = None
num_train_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
if args.model_type == 'en':
bert_model = bert_model_en
elif args.model_type == 'ch':
bert_model = bert_model_ch
if args.model_type == 'ch-en':
en_model = BertForSequenceClassification.from_pretrained(
bert_model_en, len(label_list))
ch_model = BertForSequenceClassification.from_pretrained(
bert_model_ch, len(label_list))
model = BertDualForWSDMFakeNews(ch_model, en_model, len(label_list))
else:
if args.model_architecture == "basic":
model = BertForSequenceClassification.from_pretrained(
args.bert_model, len(label_list))
elif args.model_architecture == "wordpooling":
model = BertForSequenceClassificationWithPooledWordEmbeddings.from_pretrained(
args.bert_model, len(label_list))
else:
raise NotImplementedError
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
if args.fp16:
param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
for n, param in model.named_parameters()]
elif args.optimize_on_cpu:
param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta']
optimizer_grouped_parameters = [{
'params': [p for n, p in param_optimizer if n not in no_decay],
'weight_decay_rate':
0.01
}, {
'params': [p for n, p in param_optimizer if n in no_decay],
'weight_decay_rate':
0.0
}]
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_steps)
metadata = {}
global_step = 0
if args.do_train:
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps)
train_data = create_dataset_from_examples(train_examples, label_list,
en_tokenizer, ch_tokenizer,
args)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
weight_loss = args.weight_loss
model.train()
global_nb_tr_steps = 0
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
guids, \
en_input_ids, en_input_mask, en_segment_ids, en_label_ids, \
ch_input_ids, ch_input_mask, ch_segment_ids, ch_label_ids = batch
if args.model_type == 'ch':
loss, _ = model(ch_input_ids, ch_segment_ids,
ch_input_mask, ch_label_ids, weight_loss)
elif args.model_type == 'en':
loss, _ = model(en_input_ids, en_segment_ids,
en_input_mask, en_label_ids, weight_loss)
if args.model_type == 'ch-en':
loss, _ = model(ch_input_ids, ch_segment_ids,
ch_input_mask, ch_label_ids, en_input_ids,
en_segment_ids, en_input_mask,
en_label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
tr_loss += loss.item()
nb_tr_examples += ch_input_ids.size(0)
nb_tr_steps += 1
global_nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16 or args.optimize_on_cpu:
if args.fp16 and args.loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
param.grad.data = param.grad.data / args.loss_scale
is_nan = set_optimizer_params_grad(
param_optimizer,
model.named_parameters(),
test_nan=True)
if is_nan:
logger.info(
"FP16 TRAINING: Nan in gradients, reducing loss scaling"
)
args.loss_scale = args.loss_scale / 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(
model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad()
global_step += 1
metadata['global_step'] = global_step,
metadata['tr_loss'] = tr_loss
metadata['global_nb_tr_steps'] = global_nb_tr_steps
metadata['nb_tr_steps'] = nb_tr_steps
if nb_tr_steps % args.eval_every == 0:
evaluate(args, model, processor, en_tokenizer,
ch_tokenizer, metadata, device)
evaluate(args,
model,
processor,
en_tokenizer,
ch_tokenizer,
metadata,
device,
mode="test")
if not args.do_train and args.do_eval:
model = torch.load(args.saved_model_path)
evaluate(args, model, processor, en_tokenizer, ch_tokenizer, metadata,
device)
if not args.do_train and args.do_test:
model = torch.load(args.saved_model_path)
evaluate(args,
model,
processor,
en_tokenizer,
ch_tokenizer,
metadata,
device,
mode="test")
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument(
"--data_dir",
default=None,
type=str,
required=True,
help=
"The input data dir. Should contain the .tsv files (or other data files) for the task."
)
parser.add_argument(
"--bert_model",
default=None,
type=str,
required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
## Other parameters
parser.add_argument(
"--cache_dir",
default="",
type=str,
help=
"Where do you want to store the pre-trained models downloaded from s3")
parser.add_argument(
"--max_seq_length",
default=128,
type=int,
help=
"The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument(
"--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=64,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument(
"--warmup_proportion",
default=0.1,
type=float,
help=
"Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=1,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
parser.add_argument(
'--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument(
'--loss_scale',
type=float,
default=0,
help=
"Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--server_ip',
type=str,
default='',
help="Can be used for distant debugging.")
parser.add_argument('--server_port',
type=str,
default='',
help="Can be used for distant debugging.")
args = parser.parse_args()
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port),
redirect_output=True)
ptvsd.wait_for_attach()
processors = {
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
"arg": ArgProcessor,
}
num_labels_task = {
"cola": 2,
"mnli": 3,
"mrpc": 2,
"arg": 2,
}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available()
and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".
format(device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError(
"At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(
args.output_dir) and args.do_train:
raise ValueError(
"Output directory ({}) already exists and is not empty.".format(
args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
num_labels = num_labels_task[task_name]
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.bert_model,
do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size(
)
model_state_dict = torch.load('./models/pytorch_model.bin')
cache_dir = args.cache_dir if args.cache_dir else os.path.join(
PYTORCH_PRETRAINED_BERT_CACHE, 'distributed_{}'.format(
args.local_rank))
model = BertForSequenceClassification.from_pretrained(
args.bert_model, state_dict=model_state_dict, num_labels=num_labels)
#model.load_state_dict(torch.load('./models/pytorch_model1.bin'))
# cache_dir=cache_dir,
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training."
)
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
global_step = 0
nb_tr_steps = 0
tr_loss = 0
if args.do_train:
train_features = convert_examples_to_features(train_examples,
label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features],
dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
if args.local_rank == -1:
train_sampler = RandomSampler(train_data)
else:
train_sampler = DistributedSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, input_mask, segment_ids, label_ids = batch
loss = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
if args.fp16:
optimizer.backward(loss)
else:
loss.backward()
tr_loss += loss.item()
nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1
if (step + 1) % args.gradient_accumulation_steps == 0:
if args.fp16:
# modify learning rate with special warm up BERT uses
# if args.fp16 is False, BertAdam is used that handles this automatically
lr_this_step = args.learning_rate * warmup_linear(
global_step / num_train_optimization_steps,
args.warmup_proportion)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_this_step
optimizer.step()
optimizer.zero_grad()
global_step += 1
if args.do_train:
# Save a trained model and the associated configuration
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
torch.save(model_to_save.state_dict(), output_model_file)
output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
with open(output_config_file, 'w') as f:
f.write(model_to_save.config.to_json_string())
# Load a trained model and config that you have fine-tuned
config = BertConfig(output_config_file)
model = BertForSequenceClassification(config, num_labels=num_labels)
model.load_state_dict(torch.load(output_model_file))
else:
model = BertForSequenceClassification.from_pretrained(
args.bert_model, num_labels=num_labels)
model.to(device)
pred, prob = [], []
gold = []
if args.do_eval and (args.local_rank == -1
or torch.distributed.get_rank() == 0):
eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features(eval_examples, label_list,
args.max_seq_length,
tokenizer)
logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features],
dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features],
dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_label_ids)
# Run prediction for full data
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for input_ids, input_mask, segment_ids, label_ids in tqdm(
eval_dataloader, desc="Evaluating"):
input_ids = input_ids.to(device)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss = model(input_ids, segment_ids, input_mask,
label_ids)
logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
label_ids = label_ids.to('cpu').numpy()
tmp_eval_accuracy = accuracy(logits, label_ids)
for a, b in zip(logits, label_ids):
pred.append(np.argmax(a))
gold.append(b)
#prob.append(a)
eval_loss += tmp_eval_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples
loss = tr_loss / nb_tr_steps if args.do_train else None
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'global_step': global_step,
'loss': loss
}
print(classification_report(gold, pred))
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer:
f = open('predictions.txt', 'w')
for line1 in pred:
f.write(str(line1) + '\n')
def main():
parser = argparse.ArgumentParser()
## Required parameters
parser.add_argument("--data_dir",
default=None,
type=str,
required=True,
help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
parser.add_argument("--bert_model", default=None, type=str, required=True,
help="Bert pre-trained model selected in the list: bert-base-uncased, "
"bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
"bert-base-multilingual-cased, bert-base-chinese.")
parser.add_argument("--task_name",
default=None,
type=str,
required=True,
help="The name of the task to train.")
parser.add_argument("--output_dir",
default=None,
type=str,
required=True,
help="The output directory where the model predictions and checkpoints will be written.")
## Other parameters
parser.add_argument("--cache_dir",
default="",
type=str,
help="Where do you want to store the pre-trained models downloaded from s3")
parser.add_argument("--max_seq_length",
default=128,
type=int,
help="The maximum total input sequence length after WordPiece tokenization. \n"
"Sequences longer than this will be truncated, and sequences shorter \n"
"than this will be padded.")
parser.add_argument("--do_train",
action='store_true',
help="Whether to run training.")
parser.add_argument("--do_eval",
action='store_true',
help="Whether to run eval on the dev set.")
parser.add_argument("--do_lower_case",
action='store_true',
help="Set this flag if you are using an uncased model.")
parser.add_argument("--train_batch_size",
default=32,
type=int,
help="Total batch size for training.")
parser.add_argument("--eval_batch_size",
default=8,
type=int,
help="Total batch size for eval.")
parser.add_argument("--learning_rate",
default=5e-5,
type=float,
help="The initial learning rate for Adam.")
parser.add_argument("--num_train_epochs",
default=3.0,
type=float,
help="Total number of training epochs to perform.")
parser.add_argument("--warmup_proportion",
default=0.1,
type=float,
help="Proportion of training to perform linear learning rate warmup for. "
"E.g., 0.1 = 10%% of training.")
parser.add_argument("--no_cuda",
action='store_true',
help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--seed',
type=int,
default=42,
help="random seed for initialization")
parser.add_argument('--gradient_accumulation_steps',
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.")
parser.add_argument('--fp16',
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--loss_scale',
type=float, default=0,
help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
"0 (default value): dynamic loss scaling.\n"
"Positive power of 2: static loss scaling value.\n")
parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
args = parser.parse_args()
if args.server_ip and args.server_port:
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
import ptvsd
print("Waiting for debugger attach")
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
ptvsd.wait_for_attach()
processors = {
"aus" : OOCLAUSProcessor,
"cola": ColaProcessor,
"mnli": MnliProcessor,
"mrpc": MrpcProcessor,
}
num_labels_task = {
"aus": 33,
"cola": 2,
"mnli": 3,
"mrpc": 2,
}
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
torch.cuda.set_device(args.local_rank)
device = torch.device("cuda", args.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.gradient_accumulation_steps))
args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(args.seed)
if not args.do_train and not args.do_eval:
raise ValueError("At least one of `do_train` or `do_eval` must be True.")
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train:
raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
if not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
task_name = args.task_name.lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]()
num_labels = num_labels_task[task_name]
label_list = processor.get_labels()
tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
train_examples = None
num_train_optimization_steps = None
if args.do_train:
train_examples = processor.get_train_examples(args.data_dir)
num_train_optimization_steps = int(
len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
if args.local_rank != -1:
num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
# Prepare model
cache_dir = args.cache_dir if args.cache_dir else os.path.join(PYTORCH_PRETRAINED_BERT_CACHE, 'distributed_{}'.format(args.local_rank))
model = BertForDocMultiClassification.from_pretrained(args.bert_model,
cache_dir=cache_dir,
num_labels = num_labels)
if args.fp16:
model.half()
model.to(device)
if args.local_rank != -1:
try:
from apex.parallel import DistributedDataParallel as DDP
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
model = DDP(model)
elif n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [
{'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
if args.fp16:
try:
from apex.optimizers import FP16_Optimizer
from apex.optimizers import FusedAdam
except ImportError:
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
optimizer = FusedAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
bias_correction=False,
max_grad_norm=1.0)
if args.loss_scale == 0:
optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
else:
optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
else:
optimizer = BertAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
t_total=num_train_optimization_steps)
if args.do_train:
train_features = convert_examples_to_features(
train_examples, label_list, args.max_seq_length, tokenizer)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_optimization_steps)
all_input_ids = np.array([f.input_ids for f in train_features])
all_input_mask = np.array([f.input_mask for f in train_features])
all_segment_ids = np.array([f.segment_ids for f in train_features])
all_label_ids = np.array([f.label_id for f in train_features])
initial_labeled_samples = 10000
max_queried = 10000
trainset_size = 10000
queried = initial_labeled_samples
samplecount = [initial_labeled_samples]
selection_function = RandomSelection()
train_data, val_data = split(all_input_ids, all_input_mask, all_segment_ids, all_label_ids, trainset_size)
(input_ids, input_mask, segment_ids, label_ids) = train_data
# initial process by applying base learner to labeled training data set to obtain Classifier
permutation, input_ids_train, input_mask_train, segment_ids_train, label_ids_train = \
get_k_random_samples(input_ids, input_mask, segment_ids, label_ids, initial_labeled_samples, trainset_size)
# assgin the val set the rest of the "unlabelled" traning data
input_ids_val = np.copy(input_ids)
input_mask_val = np.copy(input_mask)
segment_ids_val = np.copy(segment_ids)
label_ids_val = np.copy(label_ids)
input_ids_val = np.delete(input_ids_val, permutation, axis=0)
input_mask_val = np.delete(input_mask_val, permutation, axis=0)
segment_ids_val = np.delete(segment_ids_val, permutation, axis=0)
label_ids_val = np.delete(label_ids_val, permutation, axis=0)
print('val set:',input_ids_val.shape, label_ids_val.shape, permutation.shape)
train_dataloader = load_train_data(args, input_ids_train, input_mask_train, segment_ids_train, label_ids_train)
model = train(model, args, n_gpu, optimizer, num_train_optimization_steps, num_labels, train_dataloader, device)
model.to(device)
active_iteration = 1
while queried < max_queried:
print()
print("active_iteration:", active_iteration)
eval_dataloader = load_eval_data(args, input_ids_val, input_mask_val, segment_ids_val, label_ids_val)
probas_val = predict(model, args, eval_dataloader, device)
print("val predicted:", probas_val.shape)
uncertain_samples = selection_function.select(probas_val, initial_labeled_samples)
print("trainset before:", input_ids_train.shape)
input_ids_train = np.concatenate((input_ids_train, input_ids_val[uncertain_samples]))
input_mask_train = np.concatenate((input_mask_train, input_mask_val[uncertain_samples]))
segment_ids_train = np.concatenate((segment_ids_train, segment_ids_val[uncertain_samples]))
label_ids_train = np.concatenate((label_ids_train, label_ids_val[uncertain_samples]))
samplecount.append(input_ids_train.shape[0])
print("trainset after:", input_ids_train.shape)
input_ids_val = np.delete(input_ids_val, uncertain_samples, axis=0)
input_mask_val = np.delete(input_mask_val, uncertain_samples, axis=0)
segment_ids_val= np.delete(segment_ids_val, uncertain_samples, axis=0)
label_ids_val = np.delete(label_ids_val, uncertain_samples, axis=0)
print("val set:", input_ids_val.shape)
queried += initial_labeled_samples
train_dataloader = load_train_data(args, input_ids_train, input_mask_train, segment_ids_train, label_ids_train)
model = train(model, args, n_gpu, optimizer, num_train_optimization_steps, num_labels, train_dataloader, device)
model.to(device)
# logger.info("***** Running evaluation *****")
report_path = os.path.join(args.output_dir, "result" + str(queried) + ".csv")
# eval(model, args, processor, device, global_step, task_name, label_list, tokenizer, report_path)
logger.info("results in"+report_path)
print("results in", report_path)
active_iteration += 1
if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
logger.info("***** Running evaluation *****")
report_path = os.path.join(args.output_dir, "final_report.csv")
eval(model, args, processor, device, global_step, task_name, label_list, tokenizer, report_path)
def __init__(self):
self.bert_model = "bert-base-uncased"
self.pretrained_model = "./pretrained"
self.output_dir = "./tmp"
self.max_seq_length = 384
self.doc_stride = 128
self.max_query_length = 64
self.do_predict = True
self.predict_batch_size = 8
self.n_best_size = 20
self.max_query_length = 30
self.verbose_logging = False
self.no_cuda = False
self.seed = 42
self.fp16 = False
self.local_rank = -1
self.max_answer_length = 30
self.do_lower_case = True
if self.local_rank == -1 or self.no_cuda:
self.device = torch.device("cuda" if torch.cuda.is_available()
and not self.no_cuda else "cpu")
n_gpu = torch.cuda.device_count()
else:
self.device = torch.device("cuda", self.local_rank)
n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl')
if self.fp16:
logger.info(
"16-bits training currently not supported in distributed training"
)
self.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info(
"device: {} n_gpu: {}, distributed training: {}, 16-bits trainiing: {}"
.format(self.device, n_gpu, bool(self.local_rank != -1),
self.fp16))
random.seed(self.seed)
np.random.seed(self.seed)
torch.manual_seed(self.seed)
if n_gpu > 0:
torch.cuda.manual_seed_all(self.seed)
if os.path.exists(self.output_dir) and os.listdir(self.output_dir):
raise ValueError(
"Output directory () already exists and is not empty.")
os.makedirs(self.output_dir, exist_ok=True)
self.tokenizer = BertTokenizer.from_pretrained(self.bert_model)
# Prepare model
self.model = BertForQuestionAnswering.from_pretrained(
self.pretrained_model,
cache_dir=PYTORCH_PRETRAINED_BERT_CACHE /
'distributed_{}'.format(self.local_rank))
if self.fp16:
self.model.half()
self.model.to(self.device)
if self.local_rank != -1:
self.model = torch.nn.parallel.DistributedDataParallel(
self.model,
device_ids=[self.local_rank],
output_device=self.local_rank)
elif n_gpu > 1:
self.model = torch.nn.DataParallel(self.model)
self.model.eval()
def main(yaml_file):
with open(yaml_file) as f:
config = yaml.load(f.read())
if not config['train']['do'] and not config['predict']['do']:
raise ValueError("At least do training or do predicting in a run.")
if config['use_cuda'] and torch.cuda.is_available():
device = torch.device("cuda", torch.cuda.current_device())
use_gpu = True
else:
device = torch.device("cpu")
use_gpu = False
logger.info("device: {}".format(device))
task_name = config['task']['task_name'].lower()
if task_name not in processors:
raise ValueError("Task not found: %s" % task_name)
processor = processors[task_name]()
label_list = processor.get_labels()
os.makedirs(config['task']['output_dir'], exist_ok=True)
ckpts = [
filename for filename in os.listdir(config['task']['output_dir'])
if re.fullmatch('checkpoint-\d+', filename)
]
if config['task']['checkpoint'] or ckpts:
if config['task']['checkpoint']:
model_file = config['task']['checkpoint']
else:
model_file = os.path.join(
config['task']['output_dir'],
sorted(ckpts, key=lambda x: int(x[len('checkpoint-'):]))[-1])
logging.info('Load %s' % model_file)
checkpoint = torch.load(model_file, map_location='cpu')
start_epoch = checkpoint['epoch'] + 1
max_seq_length = checkpoint['max_seq_length']
lower_case = checkpoint['lower_case']
model = BertForNER.from_pretrained(
config['task']['bert_model_dir'],
state_dict=checkpoint['model_state'],
num_labels=len(label_list))
else:
start_epoch = 0
max_seq_length = config['task']['max_seq_length']
lower_case = config['task']['lower_case']
model = BertForNER.from_pretrained(config['task']['bert_model_dir'],
num_labels=len(label_list))
tokenizer = BertTokenizer.from_pretrained(config['task']['bert_model_dir'],
do_lower_case=lower_case)
model.to(device)
if config['train']['do']:
if config['train']['gradient_accumulation_steps'] < 1:
raise ValueError(
"Invalid gradient_accumulation_steps parameter: {}, should be >= 1"
.format(config['train']['gradient_accumulation_steps']))
config['train']['batch_size'] = int(
config['train']['batch_size'] /
config['train']['gradient_accumulation_steps'])
random.seed(config['train']['seed'])
np.random.seed(config['train']['seed'])
torch.manual_seed(config['train']['seed'])
if use_gpu:
torch.cuda.manual_seed_all(config['train']['seed'])
train_examples = processor.get_train_examples(
config['task']['data_dir'])
num_train_steps = int(
len(train_examples) / config['train']['batch_size'] /
config['train']['gradient_accumulation_steps'] *
config['train']['epochs'])
# Prepare optimizer
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = BertAdam(optimizer_grouped_parameters,
lr=config['train']['learning_rate'],
warmup=config['train']['warmup_proportion'],
t_total=num_train_steps)
train_features = convert_examples_to_features(
train_examples, max_seq_length, tokenizer,
config['task']['standard'], label_list)
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", config['train']['batch_size'])
logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features],
dtype=torch.long)
all_predict_mask = torch.ByteTensor(
[f.predict_mask for f in train_features])
all_one_hot_labels = torch.tensor(
[f.one_hot_labels for f in train_features], dtype=torch.float)
train_data = TensorDataset(all_input_ids, all_segment_ids,
all_input_mask, all_predict_mask,
all_one_hot_labels)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=config['train']['batch_size'])
model.train()
global_step = int(
len(train_examples) / config['train']['batch_size'] /
config['train']['gradient_accumulation_steps'] * start_epoch)
for epoch in trange(start_epoch,
config['train']['epochs'],
desc="Epoch"):
for step, batch in enumerate(
tqdm(train_dataloader, desc="Iteration")):
batch = tuple(t.to(device) for t in batch)
input_ids, segment_ids, input_mask, predict_mask, one_hot_labels = batch
loss = model(input_ids, segment_ids, input_mask, predict_mask,
one_hot_labels)
if config['train']['gradient_accumulation_steps'] > 1:
loss = loss / config['train']['gradient_accumulation_steps']
loss.backward()
if (step + 1
) % config['train']['gradient_accumulation_steps'] == 0:
optimizer.step()
optimizer.zero_grad()
global_step += 1
# Save a checkpoint
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
torch.save(
{
'epoch': epoch,
'model_state': model_to_save.state_dict(),
'max_seq_length': max_seq_length,
'lower_case': lower_case,
'label_list': label_list
},
os.path.join(config['task']['output_dir'],
'checkpoint-%d' % epoch))
if config['predict']['do']:
if config['predict']['dataset'] == 'train':
predict_examples = processor.get_train_examples(
config['task']['data_dir'])
elif config['predict']['dataset'] == 'dev':
predict_examples = processor.get_dev_examples(
config['task']['data_dir'])
elif config['predict']['dataset'] == 'test':
predict_examples = processor.get_test_examples(
config['task']['data_dir'])
else:
raise ValueError("The dataset %s cannot be predicted." %
config['predict']['dataset'])
predict_features = convert_examples_to_features(
predict_examples, max_seq_length, tokenizer,
config['task']['standard'], label_list)
logger.info("***** Running prediction *****")
logger.info(" Num examples = %d", len(predict_examples))
logger.info(" Batch size = %d", config['predict']['batch_size'])
all_input_ids = torch.tensor([f.input_ids for f in predict_features],
dtype=torch.long)
all_segment_ids = torch.tensor(
[f.segment_ids for f in predict_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in predict_features],
dtype=torch.long)
predict_data = TensorDataset(all_input_ids, all_segment_ids,
all_input_mask)
# Run prediction for full data
predict_sampler = SequentialSampler(predict_data)
predict_dataloader = DataLoader(
predict_data,
sampler=predict_sampler,
batch_size=config['predict']['batch_size'])
model.eval()
predictions = []
for batch in predict_dataloader:
batch = tuple(t.to(device) for t in batch)
input_ids, segment_ids, input_mask = batch
logits = model(input_ids, segment_ids, input_mask)
logits = logits.detach().cpu().numpy()
predictions.extend(np.argmax(logits, -1).tolist())
writer = codecs.open(os.path.join(
config['task']['output_dir'],
"%s.predict" % config['predict']['dataset']),
'w',
encoding='utf-8')
for predict_line, feature in zip(predictions, predict_features):
predict_labels = []
for index, label_id in enumerate(
predict_line[:sum(feature.input_mask)]):
if feature.predict_mask[index] == 1:
predict_labels.append(label_list[label_id])
writer.write(' '.join(predict_labels) + '\n')
writer.close()
