def train(local_rank, args):
# debug = False
# print("GPU:", gpu)
# world_size = args.world_size
args.global_rank = args.node_rank * args.gpus_per_node + local_rank
args.local_rank = local_rank
# args.warmup_steps = 20
debug_count = 1000
num_epoch = args.epochs
actual_train_batch_size = args.world_size * args.per_gpu_train_batch_size * args.gradient_accumulation_steps
args.actual_train_batch_size = actual_train_batch_size
set_seed(args.seed)
num_labels = 3 # we are doing NLI so we set num_labels = 3, for other task we can change this value.
max_length = args.max_length
model_class_item = MODEL_CLASSES[args.model_class_name]
model_name = model_class_item['model_name']
do_lower_case = model_class_item[
'do_lower_case'] if 'do_lower_case' in model_class_item else False
tokenizer = model_class_item['tokenizer'].from_pretrained(
model_name,
cache_dir=str(config.PRO_ROOT / "trans_cache"),
do_lower_case=do_lower_case)
model = model_class_item['sequence_classification'].from_pretrained(
model_name,
cache_dir=str(config.PRO_ROOT / "trans_cache"),
num_labels=num_labels)
padding_token_value = tokenizer.convert_tokens_to_ids(
[tokenizer.pad_token])[0]
padding_segement_value = model_class_item["padding_segement_value"]
padding_att_value = model_class_item["padding_att_value"]
left_pad = model_class_item[
'left_pad'] if 'left_pad' in model_class_item else False
batch_size_per_gpu_train = args.per_gpu_train_batch_size
batch_size_per_gpu_eval = args.per_gpu_eval_batch_size
if not args.cpu and not args.single_gpu:
dist.init_process_group(backend='nccl',
init_method='env://',
world_size=args.world_size,
rank=args.global_rank)
train_data_str = args.train_data
train_data_weights_str = args.train_weights
eval_data_str = args.eval_data
train_data_name = []
train_data_path = []
train_data_list = []
train_data_weights = []
eval_data_name = []
eval_data_path = []
eval_data_list = []
train_data_named_path = train_data_str.split(',')
weights_str = train_data_weights_str.split(
',') if train_data_weights_str is not None else None
eval_data_named_path = eval_data_str.split(',')
for named_path in train_data_named_path:
ind = named_path.find(':')
name = named_path[:ind]
path = name[ind + 1:]
if name in registered_path:
d_list = common.load_jsonl(registered_path[name])
else:
d_list = common.load_jsonl(path)
train_data_name.append(name)
train_data_path.append(path)
train_data_list.append(d_list)
if weights_str is not None:
for weights in weights_str:
train_data_weights.append(float(weights))
else:
for i in range(len(train_data_list)):
train_data_weights.append(1)
for named_path in eval_data_named_path:
ind = named_path.find(':')
name = named_path[:ind]
path = name[ind + 1:]
if name in registered_path:
d_list = common.load_jsonl(registered_path[name])
else:
d_list = common.load_jsonl(path)
eval_data_name.append(name)
eval_data_path.append(path)
eval_data_list.append(d_list)
assert len(train_data_weights) == len(train_data_list)
batching_schema = {
'uid':
RawFlintField(),
'y':
LabelFlintField(),
'input_ids':
ArrayIndexFlintField(pad_idx=padding_token_value, left_pad=left_pad),
'token_type_ids':
ArrayIndexFlintField(pad_idx=padding_segement_value,
left_pad=left_pad),
'attention_mask':
ArrayIndexFlintField(pad_idx=padding_att_value, left_pad=left_pad),
}
data_transformer = NLITransform(model_name, tokenizer, max_length)
# data_transformer = NLITransform(model_name, tokenizer, max_length, with_element=True)
eval_data_loaders = []
for eval_d_list in eval_data_list:
d_dataset, d_sampler, d_dataloader = build_eval_dataset_loader_and_sampler(
eval_d_list, data_transformer, batching_schema,
batch_size_per_gpu_eval)
eval_data_loaders.append(d_dataloader)
# Estimate the training size:
training_list = []
for i in range(len(train_data_list)):
print("Build Training Data ...")
train_d_list = train_data_list[i]
train_d_name = train_data_name[i]
train_d_weight = train_data_weights[i]
cur_train_list = sample_data_list(
train_d_list, train_d_weight
) # change later # we can apply different sample strategy here.
print(
f"Data Name:{train_d_name}; Weight: {train_d_weight}; "
f"Original Size: {len(train_d_list)}; Sampled Size: {len(cur_train_list)}"
)
training_list.extend(cur_train_list)
estimated_training_size = len(training_list)
print("Estimated training size:", estimated_training_size)
# Estimate the training size ends:
# t_total = estimated_training_size // args.gradient_accumulation_steps * num_epoch
t_total = estimated_training_size * num_epoch // args.actual_train_batch_size
if args.warmup_steps <= 0: # set the warmup steps to 0.1 * total step if the given warmup step is -1.
args.warmup_steps = int(t_total * 0.1)
if not args.cpu:
torch.cuda.set_device(args.local_rank)
model.cuda(args.local_rank)
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0
},
]
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
if not args.cpu and not args.single_gpu:
model = nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
find_unused_parameters=True)
args_dict = dict(vars(args))
file_path_prefix = '.'
if args.global_rank in [-1, 0]:
print("Total Steps:", t_total)
args.total_step = t_total
print("Warmup Steps:", args.warmup_steps)
print("Actual Training Batch Size:", actual_train_batch_size)
print("Arguments", pp.pprint(args))
# Let build the logger and log everything before the start of the first training epoch.
if args.global_rank in [
-1, 0
]: # only do logging if we use cpu or global_rank=0
if not args.debug_mode:
file_path_prefix, date = save_tool.gen_file_prefix(
f"{args.experiment_name}")
# # # Create Log File
# Save the source code.
script_name = os.path.basename(__file__)
with open(os.path.join(file_path_prefix, script_name),
'w') as out_f, open(__file__, 'r') as it:
out_f.write(it.read())
out_f.flush()
# Save option file
common.save_json(args_dict,
os.path.join(file_path_prefix, "args.json"))
checkpoints_path = Path(file_path_prefix) / "checkpoints"
if not checkpoints_path.exists():
checkpoints_path.mkdir()
prediction_path = Path(file_path_prefix) / "predictions"
if not prediction_path.exists():
prediction_path.mkdir()
global_step = 0
# print(f"Global Rank:{args.global_rank} ### ", 'Init!')
for epoch in tqdm(range(num_epoch),
desc="Epoch",
disable=args.global_rank not in [-1, 0]):
# Let's build up training dataset for this epoch
training_list = []
for i in range(len(train_data_list)):
print("Build Training Data ...")
train_d_list = train_data_list[i]
train_d_name = train_data_name[i]
train_d_weight = train_data_weights[i]
cur_train_list = sample_data_list(
train_d_list, train_d_weight
) # change later # we can apply different sample strategy here.
print(
f"Data Name:{train_d_name}; Weight: {train_d_weight}; "
f"Original Size: {len(train_d_list)}; Sampled Size: {len(cur_train_list)}"
)
training_list.extend(cur_train_list)
random.shuffle(training_list)
train_dataset = NLIDataset(training_list, data_transformer)
train_sampler = SequentialSampler(train_dataset)
if not args.cpu and not args.single_gpu:
print("Use distributed sampler.")
train_sampler = DistributedSampler(train_dataset,
args.world_size,
args.global_rank,
shuffle=True)
train_dataloader = DataLoader(
dataset=train_dataset,
batch_size=batch_size_per_gpu_train,
shuffle=False, #
num_workers=0,
pin_memory=True,
sampler=train_sampler,
collate_fn=BaseBatchBuilder(batching_schema)) #
# training build finished.
print(debug_node_info(args), "epoch: ", epoch)
if not args.cpu and not args.single_gpu:
train_sampler.set_epoch(
epoch
) # setup the epoch to ensure random sampling at each epoch
for forward_step, batch in enumerate(
tqdm(train_dataloader,
desc="Iteration",
disable=args.global_rank not in [-1, 0]), 0):
model.train()
batch = move_to_device(batch, local_rank)
# print(batch['input_ids'], batch['y'])
if args.model_class_name in ["distilbert", "bart-large"]:
outputs = model(batch['input_ids'],
attention_mask=batch['attention_mask'],
labels=batch['y'])
else:
outputs = model(batch['input_ids'],
attention_mask=batch['attention_mask'],
token_type_ids=batch['token_type_ids'],
labels=batch['y'])
loss, logits = outputs[:2]
# print(debug_node_info(args), loss, logits, batch['uid'])
# print(debug_node_info(args), loss, batch['uid'])
# Accumulated loss
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
# if this forward step need model updates
# handle fp16
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# Gradient clip: if max_grad_norm < 0
if (forward_step + 1) % args.gradient_accumulation_steps == 0:
if args.max_grad_norm > 0:
if args.fp16:
torch.nn.utils.clip_grad_norm_(
amp.master_params(optimizer), args.max_grad_norm)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step() # Update learning rate schedule
model.zero_grad()
global_step += 1
if args.global_rank in [
-1, 0
] and args.eval_frequency > 0 and global_step % args.eval_frequency == 0:
r_dict = dict()
# Eval loop:
for i in range(len(eval_data_name)):
cur_eval_data_name = eval_data_name[i]
cur_eval_data_list = eval_data_list[i]
cur_eval_dataloader = eval_data_loaders[i]
# cur_eval_raw_data_list = eval_raw_data_list[i]
evaluation_dataset(args,
cur_eval_dataloader,
cur_eval_data_list,
model,
r_dict,
eval_name=cur_eval_data_name)
# saving checkpoints
current_checkpoint_filename = \
f'e({epoch})|i({global_step})'
for i in range(len(eval_data_name)):
cur_eval_data_name = eval_data_name[i]
current_checkpoint_filename += \
f'|{cur_eval_data_name}#({round(r_dict[cur_eval_data_name]["acc"], 4)})'
if not args.debug_mode:
# save model:
model_output_dir = checkpoints_path / current_checkpoint_filename
if not model_output_dir.exists():
model_output_dir.mkdir()
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
torch.save(model_to_save.state_dict(),
str(model_output_dir / "model.pt"))
torch.save(optimizer.state_dict(),
str(model_output_dir / "optimizer.pt"))
torch.save(scheduler.state_dict(),
str(model_output_dir / "scheduler.pt"))
# save prediction:
if not args.debug_mode and args.save_prediction:
cur_results_path = prediction_path / current_checkpoint_filename
if not cur_results_path.exists():
cur_results_path.mkdir(parents=True)
for key, item in r_dict.items():
common.save_jsonl(
item['predictions'],
cur_results_path / f"{key}.jsonl")
# avoid saving too many things
for key, item in r_dict.items():
del r_dict[key]['predictions']
common.save_json(r_dict,
cur_results_path /
"results_dict.json",
indent=2)
# End of epoch evaluation.
if args.global_rank in [-1, 0]:
r_dict = dict()
# Eval loop:
for i in range(len(eval_data_name)):
cur_eval_data_name = eval_data_name[i]
cur_eval_data_list = eval_data_list[i]
cur_eval_dataloader = eval_data_loaders[i]
# cur_eval_raw_data_list = eval_raw_data_list[i]
evaluation_dataset(args,
cur_eval_dataloader,
cur_eval_data_list,
model,
r_dict,
eval_name=cur_eval_data_name)
# saving checkpoints
current_checkpoint_filename = \
f'e({epoch})|i({global_step})'
for i in range(len(eval_data_name)):
cur_eval_data_name = eval_data_name[i]
current_checkpoint_filename += \
f'|{cur_eval_data_name}#({round(r_dict[cur_eval_data_name]["acc"], 4)})'
if not args.debug_mode:
# save model:
model_output_dir = checkpoints_path / current_checkpoint_filename
if not model_output_dir.exists():
model_output_dir.mkdir()
model_to_save = (
model.module if hasattr(model, "module") else model
) # Take care of distributed/parallel training
torch.save(model_to_save.state_dict(),
str(model_output_dir / "model.pt"))
torch.save(optimizer.state_dict(),
str(model_output_dir / "optimizer.pt"))
torch.save(scheduler.state_dict(),
str(model_output_dir / "scheduler.pt"))
# save prediction:
if not args.debug_mode and args.save_prediction:
cur_results_path = prediction_path / current_checkpoint_filename
if not cur_results_path.exists():
cur_results_path.mkdir(parents=True)
for key, item in r_dict.items():
common.save_jsonl(item['predictions'],
cur_results_path / f"{key}.jsonl")
# avoid saving too many things
for key, item in r_dict.items():
del r_dict[key]['predictions']
common.save_json(r_dict,
cur_results_path / "results_dict.json",
indent=2)
def model_go_with_old_data():
seed = 12
torch.manual_seed(seed)
# bert_model_name = 'bert-large-uncased'
bert_model_name = 'bert-base-uncased'
experiment_name = 'fever_v1_nli'
lazy = False
# lazy = True
forward_size = 16
# batch_size = 64
# batch_size = 192
batch_size = 32
gradient_accumulate_step = int(batch_size / forward_size)
warmup_proportion = 0.1
learning_rate = 5e-5
num_train_epochs = 3
eval_frequency = 2000
do_lower_case = True
pair_order = 'cq'
# debug_mode = True
debug_mode = False
# est_datasize = 900_000
num_class = 3
# num_train_optimization_steps
train_sent_filtering_prob = 0.35
dev_sent_filtering_prob = 0.1
# dev_sent_results_file = config.RESULT_PATH / "doc_retri_results/fever_results/sent_results/4-14-sent_results_v0/i(5000)|e(0)|s01(0.9170917091709171)|s05(0.8842384238423843)|seed(12)_dev_sent_results.json"
# train_sent_results_file = config.RESULT_PATH / "doc_retri_results/fever_results/sent_results/4-14-sent_results_v0/train_sent_results.jsonl"
from utest.utest_format_converter_for_old_sent.tool import format_convert
dev_sent_results_file = format_convert(
config.PRO_ROOT /
"results/doc_retri_results/fever_results/sent_results/old_sent_data_by_NSMN/4-15-dev_sent_pred_scores_old_format.jsonl"
)
train_sent_results_file = format_convert(
config.PRO_ROOT /
"results/doc_retri_results/fever_results/sent_results/old_sent_data_by_NSMN/train_sent_scores_old_format.jsonl"
)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device_num = 0 if torch.cuda.is_available() else -1
n_gpu = torch.cuda.device_count()
unk_token_num = {'tokens': 1} # work around for initiating vocabulary.
vocab = ExVocabulary(unk_token_num=unk_token_num)
vocab.add_token_to_namespace('SUPPORTS', namespace='labels')
vocab.add_token_to_namespace('REFUTES', namespace='labels')
vocab.add_token_to_namespace('NOT ENOUGH INFO', namespace='labels')
vocab.add_token_to_namespace("hidden", namespace="labels")
vocab.change_token_with_index_to_namespace("hidden",
-2,
namespace='labels')
# Load Dataset
# train_fitems_list = get_inference_pair('train', True, train_sent_results_file, debug_mode, train_sent_filtering_prob)
dev_debug_num = 2481 if debug_mode else None
dev_fitems_list, dev_list = get_inference_pair('dev', False,
dev_sent_results_file,
dev_debug_num,
dev_sent_filtering_prob)
# = common.load_jsonl(config.FEVER_DEV)
if debug_mode:
dev_list = dev_list[:50]
eval_frequency = 1
# print(dev_list[-1]['_id'])
# exit(0)
# sampled_train_list = down_sample_neg(train_fitems_list, ratio=pos_ratio)
train_debug_num = 2971 if debug_mode else None
train_fitems_list, _ = get_inference_pair('train', True,
train_sent_results_file,
train_debug_num,
train_sent_filtering_prob)
est_datasize = len(train_fitems_list)
# dev_o_dict = list_dict_data_tool.list_to_dict(dev_list, 'id')
# print(dev_o_dict)
bert_tokenizer = BertTokenizer.from_pretrained(bert_model_name,
do_lower_case=do_lower_case)
bert_cs_reader = BertFeverNLIReader(bert_tokenizer,
lazy,
is_paired=True,
query_l=64,
example_filter=None,
max_l=364,
pair_order=pair_order)
bert_encoder = BertModel.from_pretrained(bert_model_name)
model = BertMultiLayerSeqClassification(bert_encoder,
num_labels=num_class,
num_of_pooling_layer=1,
act_type='tanh',
use_pretrained_pooler=True,
use_sigmoid=False)
#
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
}]
num_train_optimization_steps = int(est_datasize / forward_size / gradient_accumulate_step) * \
num_train_epochs
if debug_mode:
num_train_optimization_steps = 100
print("Estimated training size", est_datasize)
print("Number of optimization steps:", num_train_optimization_steps)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_proportion,
t_total=num_train_optimization_steps)
dev_instances = bert_cs_reader.read(dev_fitems_list)
biterator = BasicIterator(batch_size=forward_size)
biterator.index_with(vocab)
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
forbackward_step = 0
update_step = 0
logging_agent = save_tool.ScoreLogger({})
file_path_prefix = '.'
if not debug_mode:
file_path_prefix, date = save_tool.gen_file_prefix(
f"{experiment_name}")
# # # Create Log File
# Save the source code.
script_name = os.path.basename(__file__)
with open(os.path.join(file_path_prefix, script_name),
'w') as out_f, open(__file__, 'r') as it:
out_f.write(it.read())
out_f.flush()
# # # Log File end
for epoch_i in range(num_train_epochs):
print("Epoch:", epoch_i)
train_fitems_list, _ = get_inference_pair('train', True,
train_sent_results_file,
train_debug_num,
train_sent_filtering_prob)
random.shuffle(train_fitems_list)
train_instance = bert_cs_reader.read(train_fitems_list)
train_iter = biterator(train_instance, num_epochs=1, shuffle=True)
for batch in tqdm(train_iter):
model.train()
batch = move_to_device(batch, device_num)
paired_sequence = batch['paired_sequence']
paired_segments_ids = batch['paired_segments_ids']
labels_ids = batch['label']
att_mask, _ = torch_util.get_length_and_mask(paired_sequence)
s1_span = batch['bert_s1_span']
s2_span = batch['bert_s2_span']
loss = model(
paired_sequence,
token_type_ids=paired_segments_ids,
attention_mask=att_mask,
mode=BertMultiLayerSeqClassification.ForwardMode.TRAIN,
labels=labels_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if gradient_accumulate_step > 1:
loss = loss / gradient_accumulate_step
loss.backward()
forbackward_step += 1
if forbackward_step % gradient_accumulate_step == 0:
optimizer.step()
optimizer.zero_grad()
update_step += 1
if update_step % eval_frequency == 0:
print("Update steps:", update_step)
dev_iter = biterator(dev_instances,
num_epochs=1,
shuffle=False)
cur_eval_results_list = eval_model(model,
dev_iter,
device_num,
with_probs=True,
make_int=True)
results_dict = list_dict_data_tool.list_to_dict(
cur_eval_results_list, 'oid')
copied_dev_list = copy.deepcopy(dev_list)
list_dict_data_tool.append_item_from_dict_to_list(
copied_dev_list, results_dict, 'id', 'predicted_label')
mode = {'standard': True}
strict_score, acc_score, pr, rec, f1 = fever_scorer.fever_score(
copied_dev_list,
dev_fitems_list,
mode=mode,
max_evidence=5)
logging_item = {
'ss': strict_score,
'ac': acc_score,
'pr': pr,
'rec': rec,
'f1': f1,
}
save_file_name = f'i({update_step})|e({epoch_i})' \
f'|ss({strict_score})|ac({acc_score})|pr({pr})|rec({rec})|f1({f1})' \
f'|seed({seed})'
common.save_jsonl(
copied_dev_list,
Path(file_path_prefix) /
f"{save_file_name}_dev_nli_results.json")
# print(save_file_name)
logging_agent.incorporate_results({}, save_file_name,
logging_item)
logging_agent.logging_to_file(
Path(file_path_prefix) / "log.json")
model_to_save = model.module if hasattr(
model, 'module') else model
output_model_file = Path(file_path_prefix) / save_file_name
torch.save(model_to_save.state_dict(),
str(output_model_file))
def model_go_pure_aug():
# for some_params in [0.25, 0.25, 0.25]:
for some_params in [0.25, 0.25, 0.25]:
# bert_model_name = 'bert-large-uncased'
seed = 6
bert_model_name = 'bert-base-uncased'
lazy = False
forward_size = 16
batch_size = 32
gradient_accumulate_step = int(batch_size / forward_size)
warmup_proportion = 0.1
learning_rate = 5e-5
num_train_epochs = 3
do_ema = False
dev_prob_threshold = 0.1
train_prob_threshold = 0.35
debug_mode = False
# experiment_name = f"bert_fever_nli_baseline_on_fulldata"
# experiment_name = f"bert_fever_nli_baseline_on_fulldata_aug_the_same_gt_mrate({some_params})"
# experiment_name = f"bert_fever_nli_baseline_on_10p_aug_ratio({some_params})"
experiment_name = f"bert_fever_nli_baseline_on_fulldata_aug_ratio({some_params})"
# experiment_name = f"bert_fever_nli_baseline_pure_aug"
data_aug = True
# data_aug_file = config.FEVER_DATA_ROOT / "qa_aug/squad_train_turker_groundtruth.json"
# data_aug_size = int(21_015 * some_params) # 10p
# data_aug_size = int(208_346 * some_params)
# training_file = config.FEVER_DATA_ROOT / "fever_1.0/train_10.jsonl"
training_file = config.FEVER_DATA_ROOT / "fever_1.0/train.jsonl"
train_sample_top_k = 8
# est_datasize = 208_346 # full
# est_datasize = 14_544
# est_datasize = 21_015 + data_aug_size # 10p
aug_size = int(208_346 * some_params)
est_datasize = 208_346 + aug_size
# est_datasize = 208_346 + data_aug_size
num_class = 3
# num_train_optimization_steps
torch.manual_seed(seed)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
unk_token_num = {'tokens': 1} # work around for initiating vocabulary.
vocab = ExVocabulary(unk_token_num=unk_token_num)
vocab.add_token_to_namespace('SUPPORTS', namespace='labels')
vocab.add_token_to_namespace('REFUTES', namespace='labels')
vocab.add_token_to_namespace('NOT ENOUGH INFO', namespace='labels')
vocab.add_token_to_namespace("hidden", namespace="labels")
vocab.change_token_with_index_to_namespace("hidden", -2, namespace='labels')
# Finished build vocabulary.
# Load standardized sentence file
dev_upstream_sent_list = common.load_jsonl(config.FEVER_DATA_ROOT /
"upstream_sentence_selection_Feb16/dev_sent_pred_scores.jsonl")
dev_sent_after_threshold_filter = fever_ss_sampler.threshold_sampler_insure_unique(
config.FEVER_DATA_ROOT / "fever_1.0/shared_task_dev.jsonl",
dev_upstream_sent_list,
prob_threshold=dev_prob_threshold, top_n=5)
dev_data_list = fever_nli_sampler.select_sent_with_prob_for_eval(
config.FEVER_DATA_ROOT / "fever_1.0/shared_task_dev.jsonl", dev_sent_after_threshold_filter,
None, tokenized=True)
# print(dev_data_list[0])
# exit(0)
train_upstream_sent_list = common.load_jsonl(config.FEVER_DATA_ROOT /
"upstream_sentence_selection_Feb16/train_sent_scores.jsonl")
# Finished loading standardized sentence file.
bert_tokenizer = BertTokenizer.from_pretrained(bert_model_name, do_lower_case=True)
bert_fever_reader = BertReaderFeverNLI(bert_tokenizer, lazy=lazy)
dev_instances = bert_fever_reader.read(dev_data_list)
biterator = BasicIterator(batch_size=forward_size)
biterator.index_with(vocab)
# print(list(mnli_dev_instances))
# Load training model
# Load training model
model_clf = BertForSequenceClassification.from_pretrained(bert_model_name, num_labels=num_class)
ema_tracker = None
ema_model_copy = None
if do_ema and ema_tracker is None:
ema_tracker = EMA(model_clf.named_parameters(), on_cpu=True)
ema_model_copy = copy.deepcopy(model_clf)
model_clf.to(device)
param_optimizer = list(model_clf.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}
]
num_train_optimization_steps = int(est_datasize / forward_size / gradient_accumulate_step) * \
num_train_epochs
print(num_train_optimization_steps)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_proportion,
t_total=num_train_optimization_steps)
# optimizer = optim.Adam(optimizer_grouped_parameters, lr=learning_rate)
# # # Create Log File
file_path_prefix, date = save_tool.gen_file_prefix(f"{experiment_name}")
# Save the source code.
script_name = os.path.basename(__file__)
with open(os.path.join(file_path_prefix, script_name), 'w') as out_f, open(__file__, 'r') as it:
out_f.write(it.read())
out_f.flush()
# # # Log File end
model_clf.train()
if n_gpu > 1:
model_clf = nn.DataParallel(model_clf)
forbackward_step = 0
update_step = 0
eval_iter_num = 2_000 # Change this to real evaluation.
best_fever_score = -1
for n_epoch in range(num_train_epochs):
print("Resampling...")
train_sent_after_threshold_filter = \
fever_ss_sampler.threshold_sampler_insure_unique(training_file,
train_upstream_sent_list,
train_prob_threshold,
top_n=train_sample_top_k)
#
train_data_list = fever_nli_sampler.adv_simi_sample_with_prob_v1_1(
training_file,
train_sent_after_threshold_filter,
None,
tokenized=True)
aug_d_list = []
if data_aug:
aug_d_list = get_sample_data(-1)
random.shuffle(aug_d_list)
aug_d_list = aug_d_list[:aug_size]
train_data_list = train_data_list + aug_d_list
random.shuffle(train_data_list)
# train_data_list = get_sample_data(-1)
print("Sample data length:", len(train_data_list))
sampled_train_instances = bert_fever_reader.read(train_data_list)
#
train_iter = biterator(sampled_train_instances, shuffle=True, num_epochs=1)
for i, batch in enumerate(tqdm(train_iter)):
paired_sequence = batch['paired_sequence']
paired_segments_ids = batch['paired_segments_ids']
labels_ids = batch['label']
att_mask, _ = torch_util.get_length_and_mask(paired_sequence)
paired_sequence = paired_sequence.to(device)
paired_segments_ids = paired_segments_ids.to(device)
labels_ids = labels_ids.to(device)
att_mask = att_mask.to(device)
loss = model_clf(paired_sequence, token_type_ids=paired_segments_ids, attention_mask=att_mask,
labels=labels_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if gradient_accumulate_step > 1:
loss = loss / gradient_accumulate_step
loss.backward()
forbackward_step += 1
if forbackward_step % gradient_accumulate_step == 0:
optimizer.step()
optimizer.zero_grad()
update_step += 1
if do_ema and ema_tracker is not None:
# if model_clf is DataParallel, then we use model_clf.module
model_to_track = model_clf.module if hasattr(model_clf,
'module') else model_clf
ema_tracker(model_to_track.named_parameters()) # Whenever we do update, the do ema update
if update_step % eval_iter_num == 0:
print("Update steps:", update_step)
dev_iter = biterator(dev_instances, num_epochs=1, shuffle=False)
if do_ema and ema_model_copy is not None and ema_tracker is not None:
print("EMA evaluation.")
EMA.load_ema_to_model(ema_model_copy, ema_tracker)
ema_model_copy.to(device)
if n_gpu > 1:
ema_model_copy = nn.DataParallel(ema_model_copy)
dev_data_list = hidden_eval(ema_model_copy, dev_iter, dev_data_list, device)
else:
dev_data_list = hidden_eval(model_clf, dev_iter, dev_data_list, device)
eval_mode = {'check_sent_id_correct': True, 'standard': True}
fever_score, label_score, pr, rec, f1 = fever_scorer.fever_score(dev_data_list,
common.load_jsonl(config.FEVER_DATA_ROOT / "fever_1.0/shared_task_dev.jsonl"),
mode=eval_mode,
verbose=False)
print("Fever Score(FScore/LScore:/Precision/Recall/F1):", fever_score, label_score, pr, rec, f1)
print(f"Dev:{fever_score}/{label_score}")
if best_fever_score < fever_score:
print("New Best FScore")
best_fever_score = fever_score
save_path = os.path.join(
file_path_prefix,
f'i({update_step})_epoch({n_epoch})_dev({fever_score})_lacc({label_score})_seed({seed})'
)
model_to_save = model_clf.module if hasattr(model_clf,
'module') else model_clf
output_model_file = os.path.join(file_path_prefix, save_path)
torch.save(model_to_save.state_dict(), output_model_file)
print("Update steps:", update_step)
dev_iter = biterator(dev_instances, num_epochs=1, shuffle=False)
if do_ema and ema_model_copy is not None and ema_tracker is not None:
print("EMA evaluation.")
EMA.load_ema_to_model(ema_model_copy, ema_tracker)
ema_model_copy.to(device)
if n_gpu > 1:
ema_model_copy = nn.DataParallel(ema_model_copy)
dev_data_list = hidden_eval(ema_model_copy, dev_iter, dev_data_list, device)
else:
dev_data_list = hidden_eval(model_clf, dev_iter, dev_data_list, device)
eval_mode = {'check_sent_id_correct': True, 'standard': True}
fever_score, label_score, pr, rec, f1 = fever_scorer.fever_score(dev_data_list,
common.load_jsonl(config.FEVER_DATA_ROOT / "fever_1.0/shared_task_dev.jsonl"),
mode=eval_mode,
verbose=False)
print("Fever Score(FScore/LScore:/Precision/Recall/F1):", fever_score, label_score, pr, rec, f1)
print(f"Dev:{fever_score}/{label_score}")
if best_fever_score < fever_score:
print("New Best FScore")
best_fever_score = fever_score
save_path = os.path.join(
file_path_prefix,
f'i({update_step})_epoch({n_epoch})_dev({fever_score})_lacc({label_score})_seed({seed})'
)
model_to_save = model_clf.module if hasattr(model_clf,
'module') else model_clf
output_model_file = os.path.join(file_path_prefix, save_path)
torch.save(model_to_save.state_dict(), output_model_file)
def model_go():
seed = 12
torch.manual_seed(seed)
# bert_model_name = 'bert-large-uncased'
bert_model_name = 'bert-base-uncased'
experiment_name = 'hotpot_v0_cs'
lazy = False
# lazy = True
forward_size = 16
# batch_size = 64
batch_size = 128
gradient_accumulate_step = int(batch_size / forward_size)
warmup_proportion = 0.1
learning_rate = 5e-5
num_train_epochs = 5
eval_frequency = 5000
pos_ratio = 0.2
do_lower_case = True
debug_mode = False
# est_datasize = 900_000
num_class = 1
# num_train_optimization_steps
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device_num = 0 if torch.cuda.is_available() else -1
n_gpu = torch.cuda.device_count()
unk_token_num = {'tokens': 1} # work around for initiating vocabulary.
vocab = ExVocabulary(unk_token_num=unk_token_num)
vocab.add_token_to_namespace("false", namespace="labels") # 0
vocab.add_token_to_namespace("true", namespace="labels") # 1
vocab.add_token_to_namespace("hidden", namespace="labels")
vocab.change_token_with_index_to_namespace("hidden", -2, namespace='labels')
# Load Dataset
train_list = common.load_json(config.TRAIN_FILE)
dev_list = common.load_json(config.DEV_FULLWIKI_FILE)
dev_fitems_list = common.load_jsonl(
config.PDATA_ROOT / "content_selection_forward" / "hotpot_dev_p_level_unlabeled.jsonl")
train_fitems_list = common.load_jsonl(
config.PDATA_ROOT / "content_selection_forward" / "hotpot_train_p_level_labeled.jsonl")
if debug_mode:
dev_list = dev_list[:10]
dev_fitems_list = dev_fitems_list[:296]
train_fitems_list = train_fitems_list[:300]
eval_frequency = 2
# print(dev_list[-1]['_id'])
# exit(0)
sampled_train_list = down_sample_neg(train_fitems_list, ratio=pos_ratio)
est_datasize = len(sampled_train_list)
dev_o_dict = list_dict_data_tool.list_to_dict(dev_list, '_id')
# print(dev_o_dict)
bert_tokenizer = BertTokenizer.from_pretrained(bert_model_name, do_lower_case=do_lower_case)
bert_cs_reader = BertContentSelectionReader(bert_tokenizer, lazy, is_paired=True,
example_filter=lambda x: len(x['context']) == 0, max_l=286)
bert_encoder = BertModel.from_pretrained(bert_model_name)
model = BertMultiLayerSeqClassification(bert_encoder, num_labels=num_class, num_of_pooling_layer=1,
act_type='tanh', use_pretrained_pooler=True, use_sigmoid=True)
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
#
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}
]
num_train_optimization_steps = int(est_datasize / forward_size / gradient_accumulate_step) * \
num_train_epochs
print("Estimated training size", est_datasize)
print("Number of optimization steps:", num_train_optimization_steps)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_proportion,
t_total=num_train_optimization_steps)
dev_instances = bert_cs_reader.read(dev_fitems_list)
biterator = BasicIterator(batch_size=forward_size)
biterator.index_with(vocab)
forbackward_step = 0
update_step = 0
logging_agent = save_tool.ScoreLogger({})
# # # Create Log File
file_path_prefix, date = save_tool.gen_file_prefix(f"{experiment_name}")
# Save the source code.
script_name = os.path.basename(__file__)
with open(os.path.join(file_path_prefix, script_name), 'w') as out_f, open(__file__, 'r') as it:
out_f.write(it.read())
out_f.flush()
# # # Log File end
for epoch_i in range(num_train_epochs):
print("Epoch:", epoch_i)
sampled_train_list = down_sample_neg(train_fitems_list, ratio=pos_ratio)
train_instance = bert_cs_reader.read(sampled_train_list)
train_iter = biterator(train_instance, num_epochs=1, shuffle=True)
for batch in tqdm(train_iter):
model.train()
batch = move_to_device(batch, device_num)
paired_sequence = batch['paired_sequence']
paired_segments_ids = batch['paired_segments_ids']
labels_ids = batch['label']
att_mask, _ = torch_util.get_length_and_mask(paired_sequence)
s1_span = batch['bert_s1_span']
s2_span = batch['bert_s2_span']
loss = model(paired_sequence, token_type_ids=paired_segments_ids, attention_mask=att_mask,
mode=BertMultiLayerSeqClassification.ForwardMode.TRAIN,
labels=labels_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if gradient_accumulate_step > 1:
loss = loss / gradient_accumulate_step
loss.backward()
forbackward_step += 1
if forbackward_step % gradient_accumulate_step == 0:
optimizer.step()
optimizer.zero_grad()
update_step += 1
if update_step % eval_frequency == 0:
print("Update steps:", update_step)
dev_iter = biterator(dev_instances, num_epochs=1, shuffle=False)
cur_eval_results_list = eval_model(model, dev_iter, device_num, with_probs=True)
copied_dev_o_dict = copy.deepcopy(dev_o_dict)
list_dict_data_tool.append_subfield_from_list_to_dict(cur_eval_results_list, copied_dev_o_dict,
'qid', 'fid', check=True)
# Top_5
cur_results_dict_top5 = select_top_k_and_to_results_dict(copied_dev_o_dict, top_k=5)
upperbound_results_dict_top5 = append_gt_downstream_to_get_upperbound_from_doc_retri(
cur_results_dict_top5,
dev_list)
cur_results_dict_top10 = select_top_k_and_to_results_dict(copied_dev_o_dict, top_k=10)
upperbound_results_dict_top10 = append_gt_downstream_to_get_upperbound_from_doc_retri(
cur_results_dict_top10,
dev_list)
_, metrics_top5 = ext_hotpot_eval.eval(cur_results_dict_top5, dev_list, verbose=False)
_, metrics_top5_UB = ext_hotpot_eval.eval(upperbound_results_dict_top5, dev_list, verbose=False)
_, metrics_top10 = ext_hotpot_eval.eval(cur_results_dict_top10, dev_list, verbose=False)
_, metrics_top10_UB = ext_hotpot_eval.eval(upperbound_results_dict_top10, dev_list, verbose=False)
# top5_doc_f1, top5_UB_sp_f1, top10_doc_f1, top10_Ub_sp_f1
# top5_doc_f1 = metrics_top5['doc_f1']
# top5_UB_sp_f1 = metrics_top5_UB['sp_f1']
# top10_doc_f1 = metrics_top10['doc_f1']
# top10_Ub_sp_f1 = metrics_top10_UB['sp_f1']
top5_doc_recall = metrics_top5['doc_recall']
top5_UB_sp_recall = metrics_top5_UB['sp_recall']
top10_doc_recall = metrics_top10['doc_recall']
top10_Ub_sp_recall = metrics_top10_UB['sp_recall']
logging_item = {
'top5': metrics_top5,
'top5_UB': metrics_top5_UB,
'top10': metrics_top10,
'top10_UB': metrics_top10_UB,
}
# print(logging_item)
save_file_name = f'i({update_step})|e({epoch_i})' \
f'|t5_doc_recall({top5_doc_recall})|t5_sp_recall({top5_UB_sp_recall})' \
f'|t10_doc_recall({top10_doc_recall})|t5_sp_recall({top10_Ub_sp_recall})|seed({seed})'
# print(save_file_name)
logging_agent.incorporate_results({}, save_file_name, logging_item)
logging_agent.logging_to_file(Path(file_path_prefix) / "log.json")
model_to_save = model.module if hasattr(model, 'module') else model
output_model_file = Path(file_path_prefix) / save_file_name
torch.save(model_to_save.state_dict(), str(output_model_file))
def model_go(sent_filter_value, sent_top_k=5):
seed = 12
torch.manual_seed(seed)
bert_pretrain_path = config.PRO_ROOT / '.pytorch_pretrained_bert'
bert_model_name = "bert-base-uncased"
lazy = False
forward_size = 32
batch_size = 32
gradient_accumulate_step = int(batch_size / forward_size)
warmup_rate = 0.1
learning_rate = 5e-5
num_train_epochs = 5
eval_frequency = 1000
do_lower_case = True
debug = False
max_pre_context_length = 320
max_query_length = 64
doc_stride = 128
qa_num_of_layer = 2
do_ema = True
ema_device_num = 1
# s_filter_value = 0.5
s_filter_value = sent_filter_value
# s_top_k = 5
s_top_k = sent_top_k
experiment_name = f'hotpot_v0_qa_(s_top_k:{s_top_k},s_fv:{s_filter_value},qa_layer:{qa_num_of_layer})'
print("Potential total length:",
max_pre_context_length + max_query_length + 3)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device_num = 0 if torch.cuda.is_available() else -1
n_gpu = torch.cuda.device_count()
tokenizer = BertTokenizer.from_pretrained(bert_model_name,
do_lower_case=do_lower_case,
cache_dir=bert_pretrain_path)
# Load Dataset.
dev_list = common.load_json(config.DEV_FULLWIKI_FILE)
train_list = common.load_json(config.TRAIN_FILE)
dev_sentence_level_results = common.load_jsonl(
config.PRO_ROOT /
"data/p_hotpotqa/hotpotqa_sentence_level/04-19-02:17:11_hotpot_v0_slevel_retri_(doc_top_k:2)/i(12000)|e(2)|v02_f1(0.7153646038858843)|v02_recall(0.7114645831323757)|v05_f1(0.7153646038858843)|v05_recall(0.7114645831323757)|seed(12)/dev_s_level_bert_v1_results.jsonl"
)
train_sentence_level_results = common.load_jsonl(
config.PRO_ROOT /
"data/p_hotpotqa/hotpotqa_sentence_level/04-19-02:17:11_hotpot_v0_slevel_retri_(doc_top_k:2)/i(12000)|e(2)|v02_f1(0.7153646038858843)|v02_recall(0.7114645831323757)|v05_f1(0.7153646038858843)|v05_recall(0.7114645831323757)|seed(12)/train_s_level_bert_v1_results.jsonl"
)
dev_fitem_dict, dev_fitem_list, dev_sp_results_dict = get_qa_item_with_upstream_sentence(
dev_list,
dev_sentence_level_results,
is_training=False,
tokenizer=tokenizer,
max_context_length=max_pre_context_length,
max_query_length=max_query_length,
filter_value=s_filter_value,
doc_stride=doc_stride,
top_k=s_top_k,
debug_mode=debug)
train_fitem_dict, train_fitem_list, _ = get_qa_item_with_upstream_sentence(
train_list,
train_sentence_level_results,
is_training=True,
tokenizer=tokenizer,
max_context_length=max_pre_context_length,
max_query_length=max_query_length,
filter_value=s_filter_value,
doc_stride=doc_stride,
top_k=s_top_k,
debug_mode=debug)
# print(len(dev_fitem_list))
# print(len(dev_fitem_dict))
dev_o_dict = list_dict_data_tool.list_to_dict(dev_list, '_id')
if debug:
dev_list = dev_list[:100]
eval_frequency = 2
est_datasize = len(train_fitem_list)
span_pred_reader = BertPairedSpanPredReader(bert_tokenizer=tokenizer,
lazy=lazy,
example_filter=None)
bert_encoder = BertModel.from_pretrained(bert_model_name,
cache_dir=bert_pretrain_path)
model = BertSpan(bert_encoder, qa_num_of_layer)
ema = None
if do_ema:
ema = EMA(model, model.named_parameters(), device_num=ema_device_num)
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
iterator = BasicIterator(batch_size=batch_size)
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
}]
print("Total train instances:", len(train_fitem_list))
num_train_optimization_steps = int(est_datasize / forward_size / gradient_accumulate_step) * \
num_train_epochs
if debug:
num_train_optimization_steps = 100
print("Estimated training size", est_datasize)
print("Number of optimization steps:", num_train_optimization_steps)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_rate,
t_total=num_train_optimization_steps)
dev_instances = span_pred_reader.read(dev_fitem_list)
forbackward_step = 0
update_step = 0
logging_agent = save_tool.ScoreLogger({})
# # # Create Log File
file_path_prefix = None
if not debug:
file_path_prefix, date = save_tool.gen_file_prefix(
f"{experiment_name}")
# Save the source code.
script_name = os.path.basename(__file__)
with open(os.path.join(file_path_prefix, script_name),
'w') as out_f, open(__file__, 'r') as it:
out_f.write(it.read())
out_f.flush()
# # # Log File end
for epoch_i in range(num_train_epochs):
print("Epoch:", epoch_i)
print("Resampling:")
train_fitem_dict, train_fitem_list, _ = get_qa_item_with_upstream_sentence(
train_list,
train_sentence_level_results,
is_training=True,
tokenizer=tokenizer,
max_context_length=max_pre_context_length,
max_query_length=max_query_length,
filter_value=s_filter_value,
doc_stride=doc_stride,
top_k=s_top_k,
debug_mode=debug)
random.shuffle(train_fitem_list)
train_instances = span_pred_reader.read(train_fitem_list)
train_iter = iterator(train_instances, num_epochs=1, shuffle=True)
for batch in tqdm(train_iter, desc="Batch Loop"):
model.train()
batch = allen_util.move_to_device(batch, device_num)
paired_sequence = batch['paired_sequence']
paired_segments_ids = batch['paired_segments_ids']
att_mask, _ = torch_util.get_length_and_mask(paired_sequence)
gt_span = batch['gt_span']
loss = model(mode=BertSpan.ForwardMode.TRAIN,
input_ids=paired_sequence,
token_type_ids=paired_segments_ids,
attention_mask=att_mask,
gt_span=gt_span)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if gradient_accumulate_step > 1:
loss = loss / gradient_accumulate_step
loss.backward()
forbackward_step += 1
if forbackward_step % gradient_accumulate_step == 0:
optimizer.step()
if ema is not None and do_ema:
updated_model = model.module if hasattr(
model, 'module') else model
ema(updated_model.named_parameters())
optimizer.zero_grad()
update_step += 1
if update_step % eval_frequency == 0:
# print("Non-EMA EVAL:")
# eval_iter = iterator(dev_instances, num_epochs=1, shuffle=False)
# cur_eitem_list, cur_eval_dict = span_eval(model, eval_iter, do_lower_case, dev_fitem_dict,
# device_num)
# cur_results_dict = dict()
# cur_results_dict['p_answer'] = cur_eval_dict
# cur_results_dict['sp'] = dev_sp_results_dict
#
# _, metrics = ext_hotpot_eval.eval(cur_results_dict, dev_list, verbose=False)
# # print(metrics)
#
# logging_item = {
# 'score': metrics,
# }
#
# joint_f1 = metrics['joint_f1']
# joint_em = metrics['joint_em']
#
# print(logging_item)
#
# if not debug:
# save_file_name = f'i({update_step})|e({epoch_i})' \
# f'|j_f1({joint_f1})|j_em({joint_em})|seed({seed})'
#
# # print(save_file_name)
# logging_agent.incorporate_results({}, save_file_name, logging_item)
# logging_agent.logging_to_file(Path(file_path_prefix) / "log.json")
#
# model_to_save = model.module if hasattr(model, 'module') else model
# output_model_file = Path(file_path_prefix) / save_file_name
# torch.save(model_to_save.state_dict(), str(output_model_file))
if do_ema and ema is not None:
print("EMA EVAL")
ema_model = ema.get_inference_model()
ema_inference_device_ids = get_ema_gpu_id_list(
master_device_num=ema_device_num)
ema_model = ema_model.to(ema_device_num)
ema_model = torch.nn.DataParallel(
ema_model, device_ids=ema_inference_device_ids)
dev_iter = iterator(dev_instances,
num_epochs=1,
shuffle=False)
cur_eitem_list, cur_eval_dict = span_eval(
ema_model,
dev_iter,
do_lower_case,
dev_fitem_dict,
ema_device_num,
show_progress=False)
cur_results_dict = dict()
cur_results_dict['p_answer'] = cur_eval_dict
cur_results_dict['sp'] = dev_sp_results_dict
_, metrics = ext_hotpot_eval.eval(cur_results_dict,
dev_list,
verbose=False)
print(metrics)
print("---------------" * 3)
logging_item = {
'label': 'ema',
'score': metrics,
}
joint_f1 = metrics['joint_f1']
joint_em = metrics['joint_em']
print(logging_item)
if not debug:
save_file_name = f'ema_i({update_step})|e({epoch_i})' \
f'|j_f1({joint_f1})|j_em({joint_em})|seed({seed})'
# print(save_file_name)
logging_agent.incorporate_results({},
save_file_name,
logging_item)
logging_agent.logging_to_file(
Path(file_path_prefix) / "log.json")
model_to_save = ema_model.module if hasattr(
ema_model, 'module') else ema_model
output_model_file = Path(
file_path_prefix) / save_file_name
torch.save(model_to_save.state_dict(),
str(output_model_file))
def model_go():
seed = 12
torch.manual_seed(seed)
# bert_model_name = 'bert-large-uncased'
bert_model_name = 'bert-base-uncased'
lazy = False
# lazy = True
forward_size = 64
# batch_size = 64
batch_size = 128
gradient_accumulate_step = int(batch_size / forward_size)
warmup_proportion = 0.1
learning_rate = 5e-5
num_train_epochs = 5
eval_frequency = 5000
do_lower_case = True
ignore_non_verifiable = True
experiment_name = f'fever_v0_plevel_retri_(ignore_non_verifiable:{ignore_non_verifiable})'
debug_mode = False
max_l = 264
# est_datasize = 900_000
num_class = 1
# num_train_optimization_steps
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device_num = 0 if torch.cuda.is_available() else -1
n_gpu = torch.cuda.device_count()
unk_token_num = {'tokens': 1} # work around for initiating vocabulary.
vocab = ExVocabulary(unk_token_num=unk_token_num)
vocab.add_token_to_namespace("false", namespace="labels") # 0
vocab.add_token_to_namespace("true", namespace="labels") # 1
vocab.add_token_to_namespace("hidden", namespace="labels")
vocab.change_token_with_index_to_namespace("hidden",
-2,
namespace='labels')
# Load Dataset
train_ruleterm_doc_results = common.load_jsonl(
config.PRO_ROOT /
"results/doc_retri_results/fever_results/merged_doc_results/m_doc_train.jsonl"
)
dev_ruleterm_doc_results = common.load_jsonl(
config.PRO_ROOT /
"results/doc_retri_results/fever_results/merged_doc_results/m_doc_dev.jsonl"
)
# train_list = common.load_json(config.TRAIN_FILE)
dev_list = common.load_jsonl(config.FEVER_DEV)
train_fitems = fever_p_level_sampler.get_paragraph_forward_pair(
'train',
train_ruleterm_doc_results,
is_training=True,
debug=debug_mode,
ignore_non_verifiable=True)
dev_fitems = fever_p_level_sampler.get_paragraph_forward_pair(
'dev',
dev_ruleterm_doc_results,
is_training=False,
debug=debug_mode,
ignore_non_verifiable=False)
# Just to show the information
fever_p_level_sampler.down_sample_neg(train_fitems, None)
fever_p_level_sampler.down_sample_neg(dev_fitems, None)
if debug_mode:
dev_list = dev_list[:100]
eval_frequency = 2
# print(dev_list[-1]['_id'])
# exit(0)
# sampled_train_list = down_sample_neg(train_fitems_list, ratio=pos_ratio)
est_datasize = len(train_fitems)
dev_o_dict = list_dict_data_tool.list_to_dict(dev_list, 'id')
# print(dev_o_dict)
bert_tokenizer = BertTokenizer.from_pretrained(bert_model_name,
do_lower_case=do_lower_case)
bert_cs_reader = BertContentSelectionReader(
bert_tokenizer,
lazy,
is_paired=True,
example_filter=lambda x: len(x['context']) == 0,
max_l=max_l,
element_fieldname='element')
bert_encoder = BertModel.from_pretrained(bert_model_name)
model = BertMultiLayerSeqClassification(bert_encoder,
num_labels=num_class,
num_of_pooling_layer=1,
act_type='tanh',
use_pretrained_pooler=True,
use_sigmoid=True)
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
#
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
}]
num_train_optimization_steps = int(est_datasize / forward_size / gradient_accumulate_step) * \
num_train_epochs
if debug_mode:
num_train_optimization_steps = 100
print("Estimated training size", est_datasize)
print("Number of optimization steps:", num_train_optimization_steps)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_proportion,
t_total=num_train_optimization_steps)
dev_instances = bert_cs_reader.read(dev_fitems)
biterator = BasicIterator(batch_size=forward_size)
biterator.index_with(vocab)
forbackward_step = 0
update_step = 0
logging_agent = save_tool.ScoreLogger({})
if not debug_mode:
# # # Create Log File
file_path_prefix, date = save_tool.gen_file_prefix(
f"{experiment_name}")
# Save the source code.
script_name = os.path.basename(__file__)
with open(os.path.join(file_path_prefix, script_name),
'w') as out_f, open(__file__, 'r') as it:
out_f.write(it.read())
out_f.flush()
# # # Log File end
for epoch_i in range(num_train_epochs):
print("Epoch:", epoch_i)
# sampled_train_list = down_sample_neg(train_fitems_list, ratio=pos_ratio)
random.shuffle(train_fitems)
train_instance = bert_cs_reader.read(train_fitems)
train_iter = biterator(train_instance, num_epochs=1, shuffle=True)
for batch in tqdm(train_iter):
model.train()
batch = move_to_device(batch, device_num)
paired_sequence = batch['paired_sequence']
paired_segments_ids = batch['paired_segments_ids']
labels_ids = batch['label']
att_mask, _ = torch_util.get_length_and_mask(paired_sequence)
s1_span = batch['bert_s1_span']
s2_span = batch['bert_s2_span']
loss = model(
paired_sequence,
token_type_ids=paired_segments_ids,
attention_mask=att_mask,
mode=BertMultiLayerSeqClassification.ForwardMode.TRAIN,
labels=labels_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if gradient_accumulate_step > 1:
loss = loss / gradient_accumulate_step
loss.backward()
forbackward_step += 1
if forbackward_step % gradient_accumulate_step == 0:
optimizer.step()
optimizer.zero_grad()
update_step += 1
if update_step % eval_frequency == 0:
print("Update steps:", update_step)
dev_iter = biterator(dev_instances,
num_epochs=1,
shuffle=False)
cur_eval_results_list = eval_model(model,
dev_iter,
device_num,
make_int=True,
with_probs=True)
copied_dev_o_dict = copy.deepcopy(dev_o_dict)
copied_dev_d_list = copy.deepcopy(dev_list)
list_dict_data_tool.append_subfield_from_list_to_dict(
cur_eval_results_list,
copied_dev_o_dict,
'qid',
'fid',
check=True)
cur_results_dict_th0_5 = select_top_k_and_to_results_dict(
copied_dev_o_dict,
score_field_name='prob',
top_k=5,
filter_value=0.5)
list_dict_data_tool.append_item_from_dict_to_list_hotpot_style(
copied_dev_d_list, cur_results_dict_th0_5, 'id',
'predicted_docids')
# mode = {'standard': False, 'check_doc_id_correct': True}
strict_score, pr, rec, f1 = fever_scorer.fever_doc_only(
copied_dev_d_list, dev_list, max_evidence=5)
score_05 = {
'ss': strict_score,
'pr': pr,
'rec': rec,
'f1': f1,
}
list_dict_data_tool.append_subfield_from_list_to_dict(
cur_eval_results_list,
copied_dev_o_dict,
'qid',
'fid',
check=True)
cur_results_dict_th0_2 = select_top_k_and_to_results_dict(
copied_dev_o_dict,
score_field_name='prob',
top_k=5,
filter_value=0.2)
list_dict_data_tool.append_item_from_dict_to_list_hotpot_style(
copied_dev_d_list, cur_results_dict_th0_2, 'id',
'predicted_docids')
# mode = {'standard': False, 'check_doc_id_correct': True}
strict_score, pr, rec, f1 = fever_scorer.fever_doc_only(
copied_dev_d_list, dev_list, max_evidence=5)
score_02 = {
'ss': strict_score,
'pr': pr,
'rec': rec,
'f1': f1,
}
logging_item = {
'score_02': score_02,
'score_05': score_05,
}
print(logging_item)
s02_ss_score = score_02['ss']
s05_ss_score = score_05['ss']
if not debug_mode:
save_file_name = f'i({update_step})|e({epoch_i})' \
f'|v02_ofever({s02_ss_score})' \
f'|v05_ofever({s05_ss_score})|seed({seed})'
# print(save_file_name)
logging_agent.incorporate_results({}, save_file_name,
logging_item)
logging_agent.logging_to_file(
Path(file_path_prefix) / "log.json")
model_to_save = model.module if hasattr(
model, 'module') else model
output_model_file = Path(
file_path_prefix) / save_file_name
torch.save(model_to_save.state_dict(),
str(output_model_file))
def model_go():
seed = 12
torch.manual_seed(seed)
# bert_model_name = 'bert-large-uncased'
bert_pretrain_path = config.PRO_ROOT / '.pytorch_pretrained_bert'
bert_model_name = 'bert-base-uncased'
lazy = False
# lazy = True
forward_size = 128
# batch_size = 64
batch_size = 128
gradient_accumulate_step = int(batch_size / forward_size)
warmup_proportion = 0.1
learning_rate = 5e-5
num_train_epochs = 5
eval_frequency = 2000
pos_ratio = 0.2
do_lower_case = True
document_top_k = 2
experiment_name = f'hotpot_v0_slevel_retri_(doc_top_k:{document_top_k})'
debug_mode = False
do_ema = True
# est_datasize = 900_000
num_class = 1
# num_train_optimization_steps
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device_num = 0 if torch.cuda.is_available() else -1
n_gpu = torch.cuda.device_count()
unk_token_num = {'tokens': 1} # work around for initiating vocabulary.
vocab = ExVocabulary(unk_token_num=unk_token_num)
vocab.add_token_to_namespace("false", namespace="labels") # 0
vocab.add_token_to_namespace("true", namespace="labels") # 1
vocab.add_token_to_namespace("hidden", namespace="labels")
vocab.change_token_with_index_to_namespace("hidden",
-2,
namespace='labels')
# Load Dataset
train_list = common.load_json(config.TRAIN_FILE)
dev_list = common.load_json(config.DEV_FULLWIKI_FILE)
# train_fitems = sentence_level_sampler.get_train_sentence_pair(document_top_k, True, debug_mode)
# dev_fitems = sentence_level_sampler.get_dev_sentence_pair(document_top_k, False, debug_mode)
# Load train eval results list
cur_train_eval_results_list = common.load_jsonl(
config.PRO_ROOT /
"data/p_hotpotqa/hotpotqa_paragraph_level/04-10-17:44:54_hotpot_v0_cs/"
"i(40000)|e(4)|t5_doc_recall(0.8793382849426064)|t5_sp_recall(0.879496479212887)|t10_doc_recall(0.888656313301823)|t5_sp_recall(0.8888325134240054)|seed(12)/train_p_level_bert_v1_results.jsonl"
)
cur_dev_eval_results_list = common.load_jsonl(
config.PRO_ROOT /
"data/p_hotpotqa/hotpotqa_paragraph_level/04-10-17:44:54_hotpot_v0_cs/"
"i(40000)|e(4)|t5_doc_recall(0.8793382849426064)|t5_sp_recall(0.879496479212887)|t10_doc_recall(0.888656313301823)|t5_sp_recall(0.8888325134240054)|seed(12)/dev_p_level_bert_v1_results.jsonl"
)
train_fitems = get_sentence_pair(document_top_k,
train_list,
cur_train_eval_results_list,
is_training=True,
debug_mode=debug_mode)
dev_fitems = get_sentence_pair(document_top_k,
dev_list,
cur_dev_eval_results_list,
is_training=False,
debug_mode=debug_mode)
if debug_mode:
dev_list = dev_list[:100]
eval_frequency = 2
# print(dev_list[-1]['_id'])
# exit(0)
# sampled_train_list = down_sample_neg(train_fitems_list, ratio=pos_ratio)
est_datasize = len(train_fitems)
dev_o_dict = list_dict_data_tool.list_to_dict(dev_list, '_id')
# print(dev_o_dict)
bert_tokenizer = BertTokenizer.from_pretrained(
bert_model_name,
do_lower_case=do_lower_case,
cache_dir=bert_pretrain_path)
bert_cs_reader = BertContentSelectionReader(
bert_tokenizer,
lazy,
is_paired=True,
example_filter=lambda x: len(x['context']) == 0,
max_l=128,
element_fieldname='element')
bert_encoder = BertModel.from_pretrained(bert_model_name,
cache_dir=bert_pretrain_path)
model = BertMultiLayerSeqClassification(bert_encoder,
num_labels=num_class,
num_of_pooling_layer=1,
act_type='tanh',
use_pretrained_pooler=True,
use_sigmoid=True)
ema = None
if do_ema:
ema = EMA(model, model.named_parameters(), device_num=1)
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
#
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
}]
num_train_optimization_steps = int(est_datasize / forward_size / gradient_accumulate_step) * \
num_train_epochs
if debug_mode:
num_train_optimization_steps = 100
print("Estimated training size", est_datasize)
print("Number of optimization steps:", num_train_optimization_steps)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_proportion,
t_total=num_train_optimization_steps)
dev_instances = bert_cs_reader.read(dev_fitems)
biterator = BasicIterator(batch_size=forward_size)
biterator.index_with(vocab)
forbackward_step = 0
update_step = 0
logging_agent = save_tool.ScoreLogger({})
# # # Create Log File
file_path_prefix, date = save_tool.gen_file_prefix(f"{experiment_name}")
# Save the source code.
script_name = os.path.basename(__file__)
with open(os.path.join(file_path_prefix, script_name),
'w') as out_f, open(__file__, 'r') as it:
out_f.write(it.read())
out_f.flush()
# # # Log File end
for epoch_i in range(num_train_epochs):
print("Epoch:", epoch_i)
# sampled_train_list = down_sample_neg(train_fitems_list, ratio=pos_ratio)
random.shuffle(train_fitems)
train_instance = bert_cs_reader.read(train_fitems)
train_iter = biterator(train_instance, num_epochs=1, shuffle=True)
for batch in tqdm(train_iter):
model.train()
batch = move_to_device(batch, device_num)
paired_sequence = batch['paired_sequence']
paired_segments_ids = batch['paired_segments_ids']
labels_ids = batch['label']
att_mask, _ = torch_util.get_length_and_mask(paired_sequence)
s1_span = batch['bert_s1_span']
s2_span = batch['bert_s2_span']
loss = model(
paired_sequence,
token_type_ids=paired_segments_ids,
attention_mask=att_mask,
mode=BertMultiLayerSeqClassification.ForwardMode.TRAIN,
labels=labels_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if gradient_accumulate_step > 1:
loss = loss / gradient_accumulate_step
loss.backward()
forbackward_step += 1
if forbackward_step % gradient_accumulate_step == 0:
optimizer.step()
if ema is not None and do_ema:
updated_model = model.module if hasattr(
model, 'module') else model
ema(updated_model.named_parameters())
optimizer.zero_grad()
update_step += 1
if update_step % eval_frequency == 0:
print("Update steps:", update_step)
dev_iter = biterator(dev_instances,
num_epochs=1,
shuffle=False)
cur_eval_results_list = eval_model(model,
dev_iter,
device_num,
with_probs=True)
copied_dev_o_dict = copy.deepcopy(dev_o_dict)
list_dict_data_tool.append_subfield_from_list_to_dict(
cur_eval_results_list,
copied_dev_o_dict,
'qid',
'fid',
check=True)
# 0.5
cur_results_dict_v05 = select_top_k_and_to_results_dict(
copied_dev_o_dict,
top_k=5,
score_field_name='prob',
filter_value=0.5,
result_field='sp')
cur_results_dict_v02 = select_top_k_and_to_results_dict(
copied_dev_o_dict,
top_k=5,
score_field_name='prob',
filter_value=0.2,
result_field='sp')
_, metrics_v5 = ext_hotpot_eval.eval(cur_results_dict_v05,
dev_list,
verbose=False)
_, metrics_v2 = ext_hotpot_eval.eval(cur_results_dict_v02,
dev_list,
verbose=False)
v02_sp_f1 = metrics_v2['sp_f1']
v02_sp_recall = metrics_v2['sp_recall']
v02_sp_prec = metrics_v2['sp_prec']
v05_sp_f1 = metrics_v5['sp_f1']
v05_sp_recall = metrics_v5['sp_recall']
v05_sp_prec = metrics_v5['sp_prec']
logging_item = {
'v02': metrics_v2,
'v05': metrics_v5,
}
print(logging_item)
# print(logging_item)
if not debug_mode:
save_file_name = f'i({update_step})|e({epoch_i})' \
f'|v02_f1({v02_sp_f1})|v02_recall({v02_sp_recall})' \
f'|v05_f1({v05_sp_f1})|v05_recall({v05_sp_recall})|seed({seed})'
# print(save_file_name)
logging_agent.incorporate_results({}, save_file_name,
logging_item)
logging_agent.logging_to_file(
Path(file_path_prefix) / "log.json")
model_to_save = model.module if hasattr(
model, 'module') else model
output_model_file = Path(
file_path_prefix) / save_file_name
torch.save(model_to_save.state_dict(),
str(output_model_file))
if do_ema and ema is not None:
ema_model = ema.get_inference_model()
master_device_num = 1
ema_inference_device_ids = get_ema_gpu_id_list(
master_device_num=master_device_num)
ema_model = ema_model.to(master_device_num)
ema_model = torch.nn.DataParallel(
ema_model, device_ids=ema_inference_device_ids)
dev_iter = biterator(dev_instances,
num_epochs=1,
shuffle=False)
cur_eval_results_list = eval_model(ema_model,
dev_iter,
master_device_num,
with_probs=True)
copied_dev_o_dict = copy.deepcopy(dev_o_dict)
list_dict_data_tool.append_subfield_from_list_to_dict(
cur_eval_results_list,
copied_dev_o_dict,
'qid',
'fid',
check=True)
# 0.5
cur_results_dict_v05 = select_top_k_and_to_results_dict(
copied_dev_o_dict,
top_k=5,
score_field_name='prob',
filter_value=0.5,
result_field='sp')
cur_results_dict_v02 = select_top_k_and_to_results_dict(
copied_dev_o_dict,
top_k=5,
score_field_name='prob',
filter_value=0.2,
result_field='sp')
_, metrics_v5 = ext_hotpot_eval.eval(
cur_results_dict_v05, dev_list, verbose=False)
_, metrics_v2 = ext_hotpot_eval.eval(
cur_results_dict_v02, dev_list, verbose=False)
v02_sp_f1 = metrics_v2['sp_f1']
v02_sp_recall = metrics_v2['sp_recall']
v02_sp_prec = metrics_v2['sp_prec']
v05_sp_f1 = metrics_v5['sp_f1']
v05_sp_recall = metrics_v5['sp_recall']
v05_sp_prec = metrics_v5['sp_prec']
logging_item = {
'label': 'ema',
'v02': metrics_v2,
'v05': metrics_v5,
}
print(logging_item)
if not debug_mode:
save_file_name = f'ema_i({update_step})|e({epoch_i})' \
f'|v02_f1({v02_sp_f1})|v02_recall({v02_sp_recall})' \
f'|v05_f1({v05_sp_f1})|v05_recall({v05_sp_recall})|seed({seed})'
# print(save_file_name)
logging_agent.incorporate_results({},
save_file_name,
logging_item)
logging_agent.logging_to_file(
Path(file_path_prefix) / "log.json")
model_to_save = ema_model.module if hasattr(
ema_model, 'module') else ema_model
output_model_file = Path(
file_path_prefix) / save_file_name
torch.save(model_to_save.state_dict(),
str(output_model_file))
def multitask_model_go():
seed = 12
torch.manual_seed(seed)
# bert_model_name = 'bert-large-uncased'
bert_pretrain_path = config.PRO_ROOT / '.pytorch_pretrained_bert'
bert_model_name = 'bert-base-uncased'
lazy = False
# lazy = True
forward_size = 64
# batch_size = 64
batch_size = 128
gradient_accumulate_step = int(batch_size / forward_size)
warmup_proportion = 0.1
learning_rate = 5e-5
num_train_epochs = 1
eval_frequency = 5000
hotpot_pos_ratio = 0.2
do_lower_case = True
max_l = 264
experiment_name = f'mtr_p_level_(num_train_epochs:{num_train_epochs})'
debug_mode = False
do_ema = True
# est_datasize = 900_000
num_class = 1
# num_train_optimization_steps
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device_num = 0 if torch.cuda.is_available() else -1
n_gpu = torch.cuda.device_count()
unk_token_num = {'tokens': 1} # work around for initiating vocabulary.
vocab = ExVocabulary(unk_token_num=unk_token_num)
vocab.add_token_to_namespace("false", namespace="labels") # 0
vocab.add_token_to_namespace("true", namespace="labels") # 1
vocab.add_token_to_namespace("hidden", namespace="labels")
vocab.change_token_with_index_to_namespace("hidden",
-2,
namespace='labels')
# Load Hotpot Dataset
hotpot_train_list = common.load_json(config.TRAIN_FILE)
hotpot_dev_list = common.load_json(config.DEV_FULLWIKI_FILE)
hotpot_dev_o_dict = list_dict_data_tool.list_to_dict(
hotpot_dev_list, '_id')
# Load Hotpot upstream paragraph forward item
hotpot_dev_fitems_list = common.load_jsonl(
config.PDATA_ROOT / "content_selection_forward" /
"hotpot_dev_p_level_unlabeled.jsonl")
hotpot_train_fitems_list = common.load_jsonl(
config.PDATA_ROOT / "content_selection_forward" /
"hotpot_train_p_level_labeled.jsonl")
hotpot_train_fitems_list = hotpot_sampler_utils.field_name_convert(
hotpot_train_fitems_list, 'doc_t', 'element')
hotpot_dev_fitems_list = hotpot_sampler_utils.field_name_convert(
hotpot_dev_fitems_list, 'doc_t', 'element')
# Load FEVER Dataset
# fever_train_list = common.load_json(config.FEVER_TRAIN)
fever_dev_list = common.load_jsonl(config.FEVER_DEV)
fever_dev_o_dict = list_dict_data_tool.list_to_dict(fever_dev_list, 'id')
train_ruleterm_doc_results = common.load_jsonl(
config.PRO_ROOT /
"results/doc_retri_results/fever_results/merged_doc_results/m_doc_train.jsonl"
)
dev_ruleterm_doc_results = common.load_jsonl(
config.PRO_ROOT /
"results/doc_retri_results/fever_results/merged_doc_results/m_doc_dev.jsonl"
)
fever_train_fitems_list = fever_p_level_sampler.get_paragraph_forward_pair(
'train',
train_ruleterm_doc_results,
is_training=True,
debug=debug_mode,
ignore_non_verifiable=True)
fever_dev_fitems_list = fever_p_level_sampler.get_paragraph_forward_pair(
'dev',
dev_ruleterm_doc_results,
is_training=False,
debug=debug_mode,
ignore_non_verifiable=False)
if debug_mode:
hotpot_dev_list = hotpot_dev_list[:10]
hotpot_dev_fitems_list = hotpot_dev_fitems_list[:296]
hotpot_train_fitems_list = hotpot_train_fitems_list[:300]
fever_dev_list = fever_dev_list[:100]
eval_frequency = 2
# Down_sample for hotpot.
hotpot_sampled_train_list = down_sample_neg(hotpot_train_fitems_list,
ratio=hotpot_pos_ratio)
hotpot_est_datasize = len(hotpot_sampled_train_list)
fever_est_datasize = len(fever_train_fitems_list)
print("Hotpot Train Size:", hotpot_est_datasize)
print("Fever Train Size:", fever_est_datasize)
est_datasize = hotpot_est_datasize + fever_est_datasize
bert_tokenizer = BertTokenizer.from_pretrained(
bert_model_name,
do_lower_case=do_lower_case,
cache_dir=bert_pretrain_path)
bert_cs_reader = BertContentSelectionReader(
bert_tokenizer,
lazy,
is_paired=True,
example_filter=lambda x: len(x['context']) == 0,
max_l=max_l,
element_fieldname='element')
bert_encoder = BertModel.from_pretrained(bert_model_name,
cache_dir=bert_pretrain_path)
model = BertMultiLayerSeqClassification(bert_encoder,
num_labels=num_class,
num_of_pooling_layer=1,
act_type='tanh',
use_pretrained_pooler=True,
use_sigmoid=True)
ema = None
if do_ema:
ema = EMA(model, model.named_parameters(), device_num=1)
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
#
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
}]
num_train_optimization_steps = int(est_datasize / forward_size / gradient_accumulate_step) * \
num_train_epochs
if debug_mode:
num_train_optimization_steps = 100
print("Estimated training size", est_datasize)
print("Number of optimization steps:", num_train_optimization_steps)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_proportion,
t_total=num_train_optimization_steps)
hotpot_dev_instances = bert_cs_reader.read(hotpot_dev_fitems_list)
fever_dev_instances = bert_cs_reader.read(fever_dev_fitems_list)
biterator = BasicIterator(batch_size=forward_size)
biterator.index_with(vocab)
forbackward_step = 0
update_step = 0
logging_agent = save_tool.ScoreLogger({})
file_path_prefix = '.'
if not debug_mode:
# # # Create Log File
file_path_prefix, date = save_tool.gen_file_prefix(
f"{experiment_name}")
# Save the source code.
script_name = os.path.basename(__file__)
with open(os.path.join(file_path_prefix, script_name),
'w') as out_f, open(__file__, 'r') as it:
out_f.write(it.read())
out_f.flush()
# # # Log File end
for epoch_i in range(num_train_epochs):
print("Epoch:", epoch_i)
# sampled_train_list = down_sample_neg(train_fitems_list, ratio=pos_ratio)
hotpot_sampled_train_list = down_sample_neg(hotpot_train_fitems_list,
ratio=hotpot_pos_ratio)
all_train_data = hotpot_sampled_train_list + fever_train_fitems_list
random.shuffle(all_train_data)
train_instance = bert_cs_reader.read(all_train_data)
train_iter = biterator(train_instance, num_epochs=1, shuffle=True)
for batch in tqdm(train_iter):
model.train()
batch = move_to_device(batch, device_num)
paired_sequence = batch['paired_sequence']
paired_segments_ids = batch['paired_segments_ids']
labels_ids = batch['label']
att_mask, _ = torch_util.get_length_and_mask(paired_sequence)
s1_span = batch['bert_s1_span']
s2_span = batch['bert_s2_span']
loss = model(
paired_sequence,
token_type_ids=paired_segments_ids,
attention_mask=att_mask,
mode=BertMultiLayerSeqClassification.ForwardMode.TRAIN,
labels=labels_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if gradient_accumulate_step > 1:
loss = loss / gradient_accumulate_step
loss.backward()
forbackward_step += 1
if forbackward_step % gradient_accumulate_step == 0:
optimizer.step()
if ema is not None and do_ema:
updated_model = model.module if hasattr(
model, 'module') else model
ema(updated_model.named_parameters())
optimizer.zero_grad()
update_step += 1
if update_step % eval_frequency == 0:
print("Update steps:", update_step)
# Eval FEVER
eval_fever_procedure(biterator, fever_dev_instances, model,
device_num, 1, fever_dev_list,
fever_dev_o_dict, debug_mode,
logging_agent, update_step, epoch_i,
file_path_prefix, do_ema, ema, seed)
eval_hotpot_procedure(biterator, hotpot_dev_instances,
model, device_num, 1,
hotpot_dev_list, hotpot_dev_o_dict,
debug_mode, logging_agent,
update_step, epoch_i,
file_path_prefix, do_ema, ema, seed)
if not debug_mode:
print("Final Saving.")
save_file_name = f'i({update_step})|e({num_train_epochs})_final_model'
model_to_save = model.module if hasattr(model, 'module') else model
output_model_file = Path(file_path_prefix) / save_file_name
torch.save(model_to_save.state_dict(), str(output_model_file))
if do_ema and ema is not None:
print("Final EMA Saving")
ema_model = ema.get_inference_model()
save_file_name = f'i({update_step})|e({num_train_epochs})_final_ema_model'
model_to_save = ema_model.module if hasattr(
ema_model, 'module') else ema_model
output_model_file = Path(file_path_prefix) / save_file_name
torch.save(model_to_save.state_dict(), str(output_model_file))
def model_go(th_filter_prob=0.2, top_k_sent=5):
seed = 12
torch.manual_seed(seed)
# bert_model_name = 'bert-large-uncased'
bert_model_name = 'bert-base-uncased'
bert_pretrain_path = config.PRO_ROOT / '.pytorch_pretrained_bert'
lazy = False
# lazy = True
forward_size = 32
# batch_size = 64
# batch_size = 192
batch_size = 32
gradient_accumulate_step = int(batch_size / forward_size)
warmup_proportion = 0.1
# schedule_type = 'warmup_constant'
# 'warmup_cosine': warmup_cosine,
# 'warmup_constant': warmup_constant,
# 'warmup_linear': warmup_linear,
schedule_type = 'warmup_linear'
learning_rate = 5e-5
num_train_epochs = 5
eval_frequency = 4000
do_lower_case = True
pair_order = 'cq'
# debug_mode = True
# debug_mode = True
debug_mode = False
do_ema = True
maxout_model = False
# est_datasize = 900_000
num_class = 3
# num_train_optimization_steps
top_k = top_k_sent
train_sent_filtering_prob = th_filter_prob
dev_sent_filtering_prob = th_filter_prob
experiment_name = f'fever_v2_nli_th{train_sent_filtering_prob}_tk{top_k}'
# Data dataset and upstream sentence results.
dev_sent_results_list = common.load_jsonl(
config.PRO_ROOT / "data/p_fever/fever_sentence_level/04-24-00-11-19_fever_v0_slevel_retri_(ignore_non_verifiable-True)/fever_s_level_dev_results.jsonl")
train_sent_results_list = common.load_jsonl(
config.PRO_ROOT / "data/p_fever/fever_sentence_level/04-24-00-11-19_fever_v0_slevel_retri_(ignore_non_verifiable-True)/fever_s_level_train_results.jsonl")
dev_fitems, dev_list = get_nli_pair('dev', is_training=False,
sent_level_results_list=dev_sent_results_list, debug=debug_mode,
sent_top_k=top_k_sent, sent_filter_value=dev_sent_filtering_prob)
train_fitems, train_list = get_nli_pair('train', is_training=True,
sent_level_results_list=train_sent_results_list, debug=debug_mode,
sent_top_k=top_k_sent, sent_filter_value=train_sent_filtering_prob)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device_num = 0 if torch.cuda.is_available() else -1
n_gpu = torch.cuda.device_count()
unk_token_num = {'tokens': 1} # work around for initiating vocabulary.
vocab = ExVocabulary(unk_token_num=unk_token_num)
vocab.add_token_to_namespace('SUPPORTS', namespace='labels')
vocab.add_token_to_namespace('REFUTES', namespace='labels')
vocab.add_token_to_namespace('NOT ENOUGH INFO', namespace='labels')
vocab.add_token_to_namespace("hidden", namespace="labels")
vocab.change_token_with_index_to_namespace("hidden", -2, namespace='labels')
if debug_mode:
dev_list = dev_list[:100]
train_list = train_list[:100]
eval_frequency = 2
est_datasize = len(train_fitems)
bert_tokenizer = BertTokenizer.from_pretrained(bert_model_name, do_lower_case=do_lower_case,
cache_dir=bert_pretrain_path)
bert_cs_reader = BertFeverNLIReader(bert_tokenizer, lazy, is_paired=True, query_l=64,
example_filter=None, max_l=384, pair_order=pair_order)
bert_encoder = BertModel.from_pretrained(bert_model_name, cache_dir=bert_pretrain_path)
if not maxout_model:
model = BertMultiLayerSeqClassification(bert_encoder, num_labels=num_class, num_of_pooling_layer=1,
act_type='tanh', use_pretrained_pooler=True, use_sigmoid=False)
else:
model = BertPairMaxOutMatcher(bert_encoder, num_of_class=num_class, act_type="gelu", num_of_out_layers=2)
ema = None
if do_ema:
ema = EMA(model, model.named_parameters(), device_num=1)
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}
]
num_train_optimization_steps = int(est_datasize / forward_size / gradient_accumulate_step) * \
num_train_epochs
if debug_mode:
num_train_optimization_steps = 100
print("Estimated training size", est_datasize)
print("Number of optimization steps:", num_train_optimization_steps)
print("Do EMA:", do_ema)
optimizer = BertAdam(optimizer_grouped_parameters,
lr=learning_rate,
warmup=warmup_proportion,
t_total=num_train_optimization_steps,
schedule=schedule_type)
dev_instances = bert_cs_reader.read(dev_fitems)
biterator = BasicIterator(batch_size=forward_size)
biterator.index_with(vocab)
model.to(device)
if n_gpu > 1:
model = torch.nn.DataParallel(model)
forbackward_step = 0
update_step = 0
logging_agent = save_tool.ScoreLogger({})
file_path_prefix = '.'
if not debug_mode:
file_path_prefix, date = save_tool.gen_file_prefix(f"{experiment_name}")
# # # Create Log File
# Save the source code.
script_name = os.path.basename(__file__)
with open(os.path.join(file_path_prefix, script_name), 'w') as out_f, open(__file__, 'r') as it:
out_f.write(it.read())
out_f.flush()
# # # Log File end
for epoch_i in range(num_train_epochs):
print("Epoch:", epoch_i)
train_fitems_list, _ = get_nli_pair('train', is_training=True,
sent_level_results_list=train_sent_results_list, debug=debug_mode,
sent_top_k=5, sent_filter_value=train_sent_filtering_prob)
random.shuffle(train_fitems_list)
train_instance = bert_cs_reader.read(train_fitems_list)
train_iter = biterator(train_instance, num_epochs=1, shuffle=True)
for batch in tqdm(train_iter):
model.train()
batch = move_to_device(batch, device_num)
paired_sequence = batch['paired_sequence']
paired_segments_ids = batch['paired_segments_ids']
labels_ids = batch['label']
att_mask, _ = torch_util.get_length_and_mask(paired_sequence)
s1_span = batch['bert_s1_span']
s2_span = batch['bert_s2_span']
if not maxout_model:
loss = model(paired_sequence, token_type_ids=paired_segments_ids, attention_mask=att_mask,
mode=BertMultiLayerSeqClassification.ForwardMode.TRAIN,
labels=labels_ids)
else:
loss = model(paired_sequence, token_type_ids=paired_segments_ids, attention_mask=att_mask,
s1_span=s1_span, s2_span=s2_span,
mode=BertPairMaxOutMatcher.ForwardMode.TRAIN,
labels=labels_ids)
if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu.
if gradient_accumulate_step > 1:
loss = loss / gradient_accumulate_step
loss.backward()
forbackward_step += 1
if forbackward_step % gradient_accumulate_step == 0:
optimizer.step()
if ema is not None and do_ema:
updated_model = model.module if hasattr(model, 'module') else model
ema(updated_model.named_parameters())
optimizer.zero_grad()
update_step += 1
if update_step % eval_frequency == 0:
print("Update steps:", update_step)
# dev_iter = biterator(dev_instances, num_epochs=1, shuffle=False)
#
# cur_eval_results_list = eval_model(model, dev_iter, device_num, with_probs=True, make_int=True,
# feed_input_span=maxout_model)
#
# ema_results_dict = list_dict_data_tool.list_to_dict(cur_eval_results_list, 'oid')
# copied_dev_list = copy.deepcopy(dev_list)
# list_dict_data_tool.append_item_from_dict_to_list(copied_dev_list, ema_results_dict,
# 'id', 'predicted_label')
#
# mode = {'standard': True}
# strict_score, acc_score, pr, rec, f1 = fever_scorer.fever_score(copied_dev_list, dev_list,
# mode=mode, max_evidence=5)
# logging_item = {
# 'ss': strict_score, 'ac': acc_score,
# 'pr': pr, 'rec': rec, 'f1': f1,
# }
#
# if not debug_mode:
# save_file_name = f'i({update_step})|e({epoch_i})' \
# f'|ss({strict_score})|ac({acc_score})|pr({pr})|rec({rec})|f1({f1})' \
# f'|seed({seed})'
#
# common.save_jsonl(copied_dev_list, Path(file_path_prefix) /
# f"{save_file_name}_dev_nli_results.json")
#
# # print(save_file_name)
# logging_agent.incorporate_results({}, save_file_name, logging_item)
# logging_agent.logging_to_file(Path(file_path_prefix) / "log.json")
#
# model_to_save = model.module if hasattr(model, 'module') else model
# output_model_file = Path(file_path_prefix) / save_file_name
# torch.save(model_to_save.state_dict(), str(output_model_file))
if do_ema and ema is not None:
ema_model = ema.get_inference_model()
ema_device_num = 0
ema_model = ema_model.to(device)
ema_model = torch.nn.DataParallel(ema_model)
dev_iter = biterator(dev_instances, num_epochs=1, shuffle=False)
cur_ema_eval_results_list = eval_model(ema_model, dev_iter, ema_device_num, with_probs=True,
make_int=True,
feed_input_span=maxout_model)
ema_results_dict = list_dict_data_tool.list_to_dict(cur_ema_eval_results_list, 'oid')
copied_dev_list = copy.deepcopy(dev_list)
list_dict_data_tool.append_item_from_dict_to_list(copied_dev_list, ema_results_dict,
'id', 'predicted_label')
mode = {'standard': True}
strict_score, acc_score, pr, rec, f1 = fever_scorer.fever_score(copied_dev_list, dev_list,
mode=mode, max_evidence=5)
ema_logging_item = {
'label': 'ema',
'ss': strict_score, 'ac': acc_score,
'pr': pr, 'rec': rec, 'f1': f1,
}
if not debug_mode:
save_file_name = f'ema_i({update_step})|e({epoch_i})' \
f'|ss({strict_score})|ac({acc_score})|pr({pr})|rec({rec})|f1({f1})' \
f'|seed({seed})'
common.save_jsonl(copied_dev_list, Path(file_path_prefix) /
f"{save_file_name}_dev_nli_results.json")
# print(save_file_name)
logging_agent.incorporate_results({}, save_file_name, ema_logging_item)
logging_agent.logging_to_file(Path(file_path_prefix) / "log.json")
model_to_save = ema_model.module if hasattr(ema_model, 'module') else ema_model
output_model_file = Path(file_path_prefix) / save_file_name
torch.save(model_to_save.state_dict(), str(output_model_file))
