def create_openai_double_heads(self, config, input_ids, token_type_ids, position_ids,
mc_labels, lm_labels, mc_token_ids):
model = OpenAIGPTDoubleHeadsModel(config)
model.eval()
loss = model(input_ids, mc_token_ids,
lm_labels=lm_labels, mc_labels=mc_labels,
token_type_ids=token_type_ids, position_ids=position_ids)
lm_logits, mc_logits = model(input_ids, mc_token_ids, position_ids=position_ids, token_type_ids=token_type_ids)
outputs = {
"loss": loss,
"lm_logits": lm_logits,
"mc_logits": mc_logits,
}
return outputs
def __init__(self, opt, shared=None):
super(TransformerAgent, self).__init__(opt, shared)
args = AttrDict(
opt) # to keep most commands identical to the interact.py script
self.args = args
logging.basicConfig(level=logging.INFO)
self.logger = logging.getLogger(__file__)
self.logger.info(pformat(args))
random.seed(args.seed)
torch.random.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
if shared is None:
self.logger.info("Get pretrained model and tokenizer")
if args.model_checkpoint == "":
args.model_checkpoint = download_pretrained_model()
self.tokenizer = OpenAIGPTTokenizer.from_pretrained(
args.model_checkpoint)
if self.args.eval_type == "hits@1":
self.model_checkpoint = OpenAIGPTDoubleHeadsModel.from_pretrained(
args.model_checkpoint)
else:
self.model_checkpoint = OpenAIGPTLMHeadModel.from_pretrained(
args.model_checkpoint)
self.model_checkpoint.to(args.device)
self.model_checkpoint.eval()
self.logger.info("Build BPE prefix dictionary")
convai_dict = build_dict()
assert len(convai_dict) == 19304
self.prefix2words = self.get_prefix2words(convai_dict)
else:
self.model_checkpoint = shared['model']
self.tokenizer = shared['tokenizer']
self.prefix2words = shared['prefix2words']
# self.special_tokens_ids = self.tokenizer.convert_tokens_to_ids(SPECIAL_TOKENS)
self.persona = []
self.history = []
self.labels = []
self.reset()
def __init__(self, config: Config, embed_dim: int, *args, **kwargs) -> None:
super().__init__(config)
self.representation_dim = embed_dim
self.gptmode = 'gpt2'
#self.gptmode = 'openai-gpt'
if self.gptmode == 'gpt2':
self.model = GPT2DoubleHeadsModel.from_pretrained('gpt2')
self.model.resize_token_embeddings(self.model.config.vocab_size + config.num_special_tokens)
else:
self.model = OpenAIGPTDoubleHeadsModel.from_pretrained('openai-gpt')
self.model.set_num_special_tokens(len(SPECIAL_TOKENS))
self.temperature = 0.9
self.top_k = 0
self.top_p = 0.7
self.min_length = 1
self.max_length = 300
self.no_sample = True
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model_name',
type=str,
default='openai-gpt',
help='pretrained model name')
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(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument('--train_dataset', type=str, default='')
parser.add_argument('--eval_dataset', type=str, default='')
parser.add_argument('--seed', type=int, default=42)
parser.add_argument('--num_train_epochs', type=int, default=3)
parser.add_argument('--train_batch_size', type=int, default=8)
parser.add_argument('--eval_batch_size', type=int, default=16)
parser.add_argument('--max_grad_norm', type=int, default=1)
parser.add_argument('--learning_rate', type=float, default=6.25e-5)
parser.add_argument('--warmup_proportion', type=float, default=0.002)
parser.add_argument('--lr_schedule', type=str, default='warmup_linear')
parser.add_argument('--weight_decay', type=float, default=0.01)
parser.add_argument('--lm_coef', type=float, default=0.9)
parser.add_argument('--n_valid', type=int, default=374)
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()
print(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()
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("device: {}, n_gpu {}".format(device, n_gpu))
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 not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
# Load tokenizer and model
# This loading functions also add new tokens and embeddings called `special tokens`
# These new embeddings will be fine-tuned on the RocStories dataset
special_tokens = ['_start_', '_delimiter_', '_classify_']
tokenizer = OpenAIGPTTokenizer.from_pretrained(
args.model_name, special_tokens=special_tokens)
special_tokens_ids = list(
tokenizer.convert_tokens_to_ids(token) for token in special_tokens)
model = OpenAIGPTDoubleHeadsModel.from_pretrained(
args.model_name, num_special_tokens=len(special_tokens))
model.to(device)
# Load and encode the datasets
if not args.train_dataset and not args.eval_dataset:
roc_stories = cached_path(ROCSTORIES_URL)
def tokenize_and_encode(obj):
""" Tokenize and encode a nested object """
if isinstance(obj, str):
return tokenizer.convert_tokens_to_ids(tokenizer.tokenize(obj))
elif isinstance(obj, int):
return obj
return list(tokenize_and_encode(o) for o in obj)
logger.info("Encoding dataset...")
train_dataset = load_rocstories_dataset(args.train_dataset)
eval_dataset = load_rocstories_dataset(args.eval_dataset)
datasets = (train_dataset, eval_dataset)
encoded_datasets = tokenize_and_encode(datasets)
# Compute the max input length for the Transformer
max_length = model.config.n_positions // 2 - 2
input_length = max(len(story[:max_length]) + max(len(cont1[:max_length]), len(cont2[:max_length])) + 3 \
for dataset in encoded_datasets for story, cont1, cont2, _ in dataset)
input_length = min(input_length, model.config.n_positions
) # Max size of input for the pre-trained model
# Prepare inputs tensors and dataloaders
tensor_datasets = pre_process_datasets(encoded_datasets, input_length,
max_length, *special_tokens_ids)
train_tensor_dataset, eval_tensor_dataset = tensor_datasets[
0], tensor_datasets[1]
train_data = TensorDataset(*train_tensor_dataset)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
eval_data = TensorDataset(*eval_tensor_dataset)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# 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
}]
num_train_optimization_steps = len(
train_data) * args.num_train_epochs // args.train_batch_size
optimizer = OpenAIAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
max_grad_norm=args.max_grad_norm,
weight_decay=args.weight_decay,
t_total=num_train_optimization_steps)
if args.do_train:
nb_tr_steps, tr_loss, exp_average_loss = 0, 0, None
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_steps = 0
tqdm_bar = tqdm(train_dataloader, desc="Training")
for step, batch in enumerate(tqdm_bar):
batch = tuple(t.to(device) for t in batch)
input_ids, mc_token_ids, lm_labels, mc_labels = batch
losses = model(input_ids, mc_token_ids, lm_labels, mc_labels)
loss = args.lm_coef * losses[0] + losses[1]
loss.backward()
optimizer.step()
optimizer.zero_grad()
tr_loss += loss.item()
exp_average_loss = loss.item(
) if exp_average_loss is None else 0.7 * exp_average_loss + 0.3 * loss.item(
)
nb_tr_steps += 1
tqdm_bar.desc = "Training loss: {:.2e} lr: {:.2e}".format(
exp_average_loss,
optimizer.get_lr()[0])
# Save a trained model
if args.do_train:
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")
config = model.config
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 = OpenAIGPTDoubleHeadsModel(config)
model.load_state_dict(model_state_dict)
model.to(device)
if args.do_eval:
model.eval()
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for batch in tqdm(eval_dataloader, desc="Evaluating"):
batch = tuple(t.to(device) for t in batch)
input_ids, mc_token_ids, lm_labels, mc_labels = batch
with torch.no_grad():
_, mc_loss = model(input_ids, mc_token_ids, lm_labels,
mc_labels)
_, mc_logits = model(input_ids, mc_token_ids)
mc_logits = mc_logits.detach().cpu().numpy()
mc_labels = mc_labels.to('cpu').numpy()
tmp_eval_accuracy = accuracy(mc_logits, mc_labels)
eval_loss += mc_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
train_loss = tr_loss / nb_tr_steps if args.do_train else None
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'train_loss': train_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])))
from pytorch_pretrained_bert import OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer
model = OpenAIGPTDoubleHeadsModel.from_pretrained('openai-gpt')
tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
# We will use 5 special tokens:
# - <bos> to indicate the start of the sequence
# - <eos> to indicate the end of the sequence
# - <speaker1> to indicate the beginning and the tokens of an utterance from the user
# - <speaker2> to indicate the beginning and the tokens of an utterance from the bot
# - <pad> as a padding token to build batches of sequences
SPECIAL_TOKENS = ["<bos>", "<eos>", "<speaker1>", "<speaker2>", "<pad>"]
# We can add these special tokens to the vocabulary and the embeddings of the model:
tokenizer.set_special_tokens(SPECIAL_TOKENS)
model.set_num_special_tokens(len(SPECIAL_TOKENS))
################################################################################
from itertools import chain
# Let's define our contexts and special tokens
persona = [["i", "like", "playing", "football", "."],
["i", "am", "from", "NYC", "."]]
history = [["hello", "how", "are", "you", "?"],
["i", "am", "fine", "thanks", "."]]
reply = ["great", "to", "hear"]
bos, eos, speaker1, speaker2 = "<bos>", "<eos>", "<speaker1>", "<speaker2>"
def build_inputs(persona, history, reply):
def main():
parser = argparse.ArgumentParser()
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('--model_name',
type=str,
default='openai-gpt',
help='pretrained model name')
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(
"--answer_only",
default=False,
action='store_true',
help="Whether to run with answers only (blank out question).")
parser.add_argument(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument(
"--load_model_from",
default=None,
type=str,
help=
"The saved model file to load before doing any training or eval (if both --do_train and --do_eval are specified, the saved model will be loaded, then trained, then the trained model will be evaluated)."
)
parser.add_argument(
'--train_filename',
type=str,
default='train.csv',
help="Filename to load train data from (relative to data_dir)")
parser.add_argument(
'--eval_filename',
type=str,
default='val.csv',
help="File to load eval data from (relative to data_dir)")
parser.add_argument(
'--data_format',
type=str,
choices=['swag', 'codah'],
default='swag',
help=
"Format of the train and eval files (original SWAG CSV format vs our TSV format)"
)
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('--seed', type=int, default=42)
parser.add_argument('--num_train_epochs', type=int, default=3)
parser.add_argument('--train_batch_size', type=int, default=32)
parser.add_argument('--eval_batch_size', type=int, default=8)
parser.add_argument('--max_grad_norm', type=int, default=1)
parser.add_argument('--learning_rate', type=float, default=6.25e-5)
parser.add_argument('--warmup_proportion', type=float, default=0.002)
parser.add_argument('--lr_schedule', type=str, default='warmup_linear')
parser.add_argument('--weight_decay', type=float, default=0.01)
parser.add_argument('--lm_coef', type=float, default=0.5)
parser.add_argument('--n_valid', type=int, default=374)
parser.add_argument(
'--gradient_accumulation_steps',
type=int,
default=8,
help=
"Number of updates steps to accumulate before performing a backward/update pass."
)
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()
print(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()
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)
torch.cuda.manual_seed_all(args.seed)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("device: {}, n_gpu {}".format(device, n_gpu))
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 not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
if args.do_eval and (not args.do_train) and args.load_model_from is None:
args.load_model_from = os.path.join(args.output_dir,
'pytorch_model.bin')
# Load tokenizer and model
# This loading functions also add new tokens and embeddings called `special tokens`
# These new embeddings will be fine-tuned on the RocStories dataset
special_tokens = ['_start_', '_delimiter_', '_classify_']
tokenizer = OpenAIGPTTokenizer.from_pretrained(
args.model_name, special_tokens=special_tokens)
special_tokens_ids = list(
tokenizer.convert_tokens_to_ids(token) for token in special_tokens)
model = OpenAIGPTDoubleHeadsModel.from_pretrained(
args.model_name, num_special_tokens=len(special_tokens))
config = model.config
if args.load_model_from:
model_state_dict = torch.load(args.load_model_from)
model = OpenAIGPTDoubleHeadsModel(config)
model.load_state_dict(model_state_dict)
model.to(device)
# Load and encode the datasets
logger.info("Loading datasets...")
datasets = []
dataset_keys = dict()
if args.do_train:
train_dataset = read_swag_examples(os.path.join(
args.data_dir, args.train_filename),
is_training=True,
answer_only=args.answer_only,
data_format=args.data_format)
train_dataset = [
EncodedSwagExample(ex, tokenizer)
for ex in tqdm(train_dataset, desc='Encoding train')
]
dataset_keys['train'] = len(datasets)
datasets.append(train_dataset)
if args.do_eval:
eval_dataset = read_swag_examples(os.path.join(args.data_dir,
args.eval_filename),
is_training=True,
answer_only=args.answer_only,
data_format=args.data_format)
eval_dataset = [
EncodedSwagExample(ex, tokenizer)
for ex in tqdm(eval_dataset, desc='Encoding eval')
]
dataset_keys['eval'] = len(datasets)
datasets.append(eval_dataset)
# Compute the max input length for the Transformer
max_length = model.config.n_positions // 2 - 2
input_length = max(len(swagex.context_tokens[:max_length]) + len(swagex.start_ending_tokens[:max_length]) + max(len(ending[:max_length]) for ending in swagex.endings_tokens) + 3 \
for dataset in datasets for swagex in dataset)
input_length = min(input_length, model.config.n_positions
) # Max size of input for the pre-trained model
print('---')
print('Input length: {}\n'.format(input_length))
print('---')
# Prepare inputs tensors and dataloaders
tensor_datasets = pre_process_datasets(datasets, input_length, max_length,
*special_tokens_ids)
if args.do_train:
train_tensor_dataset = tensor_datasets[dataset_keys['train']]
if args.do_eval:
eval_tensor_dataset = tensor_datasets[dataset_keys['eval']]
# Prepare optimizer
if args.do_train:
train_data = TensorDataset(*train_tensor_dataset)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_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
}]
#num_train_optimization_steps = len(train_data) * args.num_train_epochs // args.train_batch_size
num_train_optimization_steps = int(
len(train_data) / args.train_batch_size /
args.gradient_accumulation_steps) * args.num_train_epochs
optimizer = OpenAIAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
max_grad_norm=args.max_grad_norm,
weight_decay=args.weight_decay,
t_total=num_train_optimization_steps)
nb_tr_steps, tr_loss, exp_average_loss = 0, 0, None
model.train()
for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0
nb_tr_steps = 0
tqdm_bar = tqdm(train_dataloader, desc="Training")
for step, batch in enumerate(tqdm_bar):
batch = tuple(t.to(device) for t in batch)
input_ids, mc_token_ids, lm_labels, mc_labels = batch
losses = model(input_ids, mc_token_ids, lm_labels, mc_labels)
loss = args.lm_coef * losses[0] + losses[1]
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
tr_loss += loss.item()
nb_tr_steps += 1
exp_average_loss = loss.item(
) if exp_average_loss is None else 0.7 * exp_average_loss + 0.3 * loss.item(
)
tqdm_bar.desc = "Training loss: {:.2e} lr: {:.2e}".format(
exp_average_loss,
optimizer.get_lr()[0])
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
optimizer.zero_grad()
# Save a trained model
output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
if args.do_train:
model_to_save = model.module if hasattr(
model, 'module') else model # Only save the model it-self
torch.save(model_to_save.state_dict(), output_model_file)
if args.do_eval:
eval_data = TensorDataset(*eval_tensor_dataset)
eval_sampler = SequentialSampler(eval_data)
eval_dataloader = DataLoader(eval_data,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# Load a trained model that you have fine-tuned
if args.do_train:
model_state_dict = torch.load(output_model_file)
model = OpenAIGPTDoubleHeadsModel(config)
model.load_state_dict(model_state_dict)
model.to(device)
model.eval()
all_model_outputs = []
data_index = 0
eval_loss, eval_accuracy = 0, 0
nb_eval_steps, nb_eval_examples = 0, 0
for batch in tqdm(eval_dataloader, desc="Evaluating"):
batch = tuple(t.to(device) for t in batch)
input_ids, mc_token_ids, lm_labels, mc_labels = batch
with torch.no_grad():
_, mc_loss = model(input_ids, mc_token_ids, lm_labels,
mc_labels)
_, mc_logits = model(input_ids, mc_token_ids)
mc_logits = mc_logits.detach().cpu().numpy()
mc_labels = mc_labels.to('cpu').numpy()
tmp_eval_accuracy = accuracy(mc_logits, mc_labels)
eval_loss += mc_loss.mean().item()
eval_accuracy += tmp_eval_accuracy
for i in range(input_ids.size(0)):
output_obj = dict()
output_obj['logits'] = [float(x) for x in mc_logits[i]]
output_obj['true_label'] = int(mc_labels[i])
output_obj['model_label'] = int(np.argmax(mc_logits[i]))
output_obj['swag_data'] = datasets[
dataset_keys['eval']][data_index].raw_example.to_dict()
all_model_outputs.append(output_obj)
data_index += 1
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
train_loss = tr_loss / nb_tr_steps if args.do_train else None
result = {
'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy,
'train_loss': train_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])))
with open(
os.path.join(args.output_dir, args.model_labels_save_filename),
'w') as f:
json.dump(all_model_outputs, f)
epochs = 10
max_len = 100
batch_size = 128
learning_rate = 6.25e-5
warmup_proportion = 0.002
max_grad_norm = 1
weight_decay = 0.01
train_path = '/data/users/kyle.shaffer/dialog_data/cornell_movie/cakechat_model/corpora_processed/train_no_tok.txt'
valid_path = '/data/users/kyle.shaffer/dialog_data/cornell_movie/cakechat_model/corpora_processed/valid_no_tok.txt'
model_name = 'openai-gpt'
lm_coef = 0.9
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
special_tokens = ['_start_', '_delimiter_', '_classify_']
tokenizer = OpenAIGPTTokenizer.from_pretrained(model_name, special_tokens=special_tokens)
special_tokens_ids = list(tokenizer.convert_tokens_to_ids(token) for token in special_tokens)
model = OpenAIGPTDoubleHeadsModel.from_pretrained(model_name, num_special_tokens=len(special_tokens))
x_train_arr, y_train_arr = load_data_with_tok(train_path, tokenizer, max_len)
x_valid_arr, y_valid_arr = load_data_with_tok(valid_path, tokenizer, max_len)
train_dataloader, valid_dataloader = data_to_torch(x_train_arr, y_train_arr, x_valid_arr, y_valid_arr, batch_size)
opt = prep_optimizer(model=model, epochs=epochs, learning_rate=learning_rate, warmup_proportion=warmup_proportion,
max_grad_norm=max_grad_norm, weight_decay=weight_decay)
train_model(model=model, train_dataloader=train_dataloader, opt=opt, lm_coef=lm_coef, epochs=epochs, valid_dataloader=valid_dataloader)
def train():
EPOCHS = 3
SAVE_ITR = 3
LM_COEF = 1.0
MC_COEF = 1.0
DEVICE = 0
FP16 = True
MAX_NORM = 1.0 # Clipping Gradient Norm
GRAD_ACCUM_STEPS = 6 #4
train_batch_size = 3 #4
tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
model = OpenAIGPTDoubleHeadsModel.from_pretrained('openai-gpt')
add_special_tokens_(model, tokenizer)
model = model.cuda(DEVICE)
optimizer = AdamW(model.parameters(), lr=6.25e-5, correct_bias=True)
if FP16:
model, optimizer = amp.initialize(
model, optimizer,
opt_level='O1') #O1/O2 #https://nvidia.github.io/apex/amp.html
train_dataset = torch.load('train_dataset.pyobj')
train_loader = DataLoader(train_dataset,
batch_size=train_batch_size,
shuffle=True)
def update(b, batch):
model.train()
batch = [input_tensor.to(DEVICE) for input_tensor in batch]
input_ids, mc_token_ids, lm_labels, mc_labels, token_type_ids = batch
lm_loss, mc_loss = model(input_ids, mc_token_ids, lm_labels, mc_labels,
token_type_ids)
loss = (lm_loss * LM_COEF + mc_loss * MC_COEF) / GRAD_ACCUM_STEPS
if FP16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer),
MAX_NORM)
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), MAX_NORM)
if b % GRAD_ACCUM_STEPS == 0:
optimizer.step()
optimizer.zero_grad()
return loss.item()
E, B = EPOCHS, len(train_loader)
for e in range(EPOCHS):
for b, batch in enumerate(train_loader):
loss = update(b, batch)
if b % (B // 300) == 0:
print(e, str(b) + '/' + str(B), loss)
torch.cuda.empty_cache()
if (e + 1) % SAVE_ITR == 0:
torch.save(model.state_dict(),
'/media/sec/conv_ai_weights/' + str(e + 1) + '.pth')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model_name',
type=str,
default='openai-gpt',
help='pretrained model name')
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(
"--output_dir",
default=None,
type=str,
required=True,
help=
"The output directory where the model predictions and checkpoints will be written."
)
parser.add_argument('--train_dataset', type=str, default='')
parser.add_argument('--eval_dataset', type=str, default='')
parser.add_argument('--seed', type=int, default=42)
parser.add_argument('--num_train_epochs', type=int, default=3)
parser.add_argument('--train_batch_size', type=int, default=8)
parser.add_argument('--eval_batch_size', type=int, default=16)
parser.add_argument('--max_grad_norm', type=int, default=1)
parser.add_argument('--learning_rate', type=float, default=6.25e-5)
parser.add_argument('--warmup_proportion', type=float, default=0.002)
parser.add_argument('--lr_schedule', type=str, default='warmup_linear')
parser.add_argument('--weight_decay', type=float, default=0.01)
parser.add_argument('--lm_coef', type=float, default=0.9)
parser.add_argument('--n_valid', type=int, default=374)
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()
print(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()
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
n_gpu = torch.cuda.device_count()
logger.info("device: {}, n_gpu {}".format(device, n_gpu))
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 not os.path.exists(args.output_dir):
os.makedirs(args.output_dir)
# Load tokenizer and model
# This loading functions also add new tokens and embeddings called `special tokens`
# These new embeddings will be fine-tuned on the RocStories dataset
special_tokens = ['_start_', '_delimiter_', '_classify_']
tokenizer = OpenAIGPTTokenizer.from_pretrained(
args.model_name, special_tokens=special_tokens)
special_tokens_ids = list(
tokenizer.convert_tokens_to_ids(token) for token in special_tokens)
model = OpenAIGPTDoubleHeadsModel.from_pretrained(
args.model_name, num_special_tokens=len(special_tokens))
model.to(device)
# Load and encode the datasets
if not args.train_dataset and not args.eval_dataset:
roc_stories = cached_path(ROCSTORIES_URL)
def tokenize_and_encode(obj):
""" Tokenize and encode a nested object """
if isinstance(obj, str):
return tokenizer.convert_tokens_to_ids(tokenizer.tokenize(obj))
elif isinstance(obj, int):
return obj
return list(tokenize_and_encode(o) for o in obj)
logger.info("Encoding dataset...")
train_dataset = load_csqa_dataset(args.train_dataset)
print("Splitting train 90-10 into train-dev.")
dev_dataset = train_dataset[int(len(train_dataset) * 0.9):]
train_dataset = train_dataset[:int(len(train_dataset) * 0.9)]
test_dataset = load_csqa_dataset(args.eval_dataset)
datasets = (train_dataset, dev_dataset, test_dataset)
encoded_datasets = tokenize_and_encode(datasets)
# Compute the mex input length for the Transformer
max_length = model.config.n_positions // 2 - 2
input_length = max(
len(question[:max_length]) +
max(len(answer1[:max_length]), len(answer2[:max_length]),
len(answer3[:max_length])) + 3 for dataset in encoded_datasets
for question, answer1, answer2, answer3, _ in dataset)
input_length = min(input_length, model.config.n_positions
) # Max size of input for the pre-trained model
# Prepare inputs tensors and dataloaders
tensor_datasets = pre_process_datasets(encoded_datasets, input_length,
max_length, *special_tokens_ids)
train_tensor_dataset = tensor_datasets[0]
dev_tensor_dataset = tensor_datasets[1]
test_tensor_dataset = tensor_datasets[2]
train_data = TensorDataset(*train_tensor_dataset)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data,
sampler=train_sampler,
batch_size=args.train_batch_size)
dev_data = TensorDataset(*dev_tensor_dataset)
dev_sampler = RandomSampler(dev_data)
dev_dataloader = DataLoader(dev_data,
sampler=dev_sampler,
batch_size=args.train_batch_size)
test_data = TensorDataset(*test_tensor_dataset)
test_sampler = SequentialSampler(test_data)
test_dataloader = DataLoader(test_data,
sampler=test_sampler,
batch_size=args.eval_batch_size)
# 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
}]
num_train_optimization_steps = len(
train_data) * args.num_train_epochs // args.train_batch_size
optimizer = OpenAIAdam(optimizer_grouped_parameters,
lr=args.learning_rate,
warmup=args.warmup_proportion,
max_grad_norm=args.max_grad_norm,
weight_decay=args.weight_decay,
t_total=num_train_optimization_steps)
if args.do_train:
nb_tr_steps, tr_loss, exp_average_loss = 0, 0, None
best_dev_accuracy = 0
test_acc_best_dev = 0
best_dev_epoch = 0
no_up = 0
tqdm_epoch = tqdm(range(args.num_train_epochs), desc="Epoch")
for epoch in tqdm_epoch:
model.train()
tr_loss = 0
nb_tr_steps = 0
tqdm_bar = tqdm(train_dataloader, desc="Training")
for step, batch in enumerate(tqdm_bar):
batch = tuple(t.to(device) for t in batch)
input_ids, mc_token_ids, lm_labels, mc_labels = batch
losses = model(input_ids, mc_token_ids, lm_labels, mc_labels)
loss = args.lm_coef * losses[0] + losses[1]
loss.backward()
optimizer.step()
optimizer.zero_grad()
tr_loss += loss.item()
exp_average_loss = loss.item(
) if exp_average_loss is None else 0.7 * exp_average_loss + 0.3 * loss.item(
)
nb_tr_steps += 1
tqdm_bar.desc = "Training loss: {:.2e} lr: {:.2e}".format(
exp_average_loss,
optimizer.get_lr()[0])
# train_loss, train_accuracy = evaluate(model, device, train_dataloader, desc="Evaluate Train")
dev_loss, dev_accuracy = evaluate(model,
device,
dev_dataloader,
desc="Evaluate Dev")
test_loss, test_accuracy = evaluate(model,
device,
test_dataloader,
desc="Evaluate Test")
train_loss = tr_loss / nb_tr_steps if args.do_train else None
if dev_accuracy >= best_dev_accuracy:
# New best model.
best_dev_accuracy = dev_accuracy
test_acc_best_dev = test_accuracy
best_dev_epoch = epoch + 1
no_up = 0
# Save the new best 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")
torch.save(model_to_save.state_dict(), output_model_file)
else:
no_up += 1
tqdm.write("\t ***** Eval results (Epoch %s) *****" %
str(epoch + 1))
# tqdm.write("\t train_accuracy = %s" % str(train_accuracy))
tqdm.write("\t dev_accuracy = %s" % str(dev_accuracy))
tqdm.write("")
tqdm.write("\t best_dev_accuracy = %s" % str(best_dev_accuracy))
tqdm.write("\t test_acc_best_dev = %s" % str(test_acc_best_dev))
tqdm.write("\t best_dev_epoch = %s" % str(best_dev_epoch))
tqdm.write("\t no_up = %s" % str(no_up))
tqdm.write("")
if no_up >= 10:
tqdm_epoch.close()
break
def run():
parser = ArgumentParser()
parser.add_argument(
"--dataset_path",
type=str,
default="",
help="Path or url of the dataset. If empty download from S3.")
parser.add_argument("--dataset_cache",
type=str,
default='./dataset_cache',
help="Path or url of the dataset cache")
parser.add_argument("--model_checkpoint",
type=str,
default="",
help="Path, url or short name of the model")
parser.add_argument(
"--max_history",
type=int,
default=2,
help="Number of previous utterances to keep in history")
parser.add_argument("--device",
type=str,
default="cuda" if torch.cuda.is_available() else "cpu",
help="Device (cuda or cpu)")
parser.add_argument("--no_sample",
action='store_true',
help="Set to use greedy decoding instead of sampling")
parser.add_argument("--max_length",
type=int,
default=50,
help="Maximum length of the output utterances")
parser.add_argument("--min_length",
type=int,
default=1,
help="Minimum length of the output utterances")
parser.add_argument("--seed", type=int, default=42, help="Seed")
parser.add_argument("--temperature",
type=int,
default=0.7,
help="Sampling softmax temperature")
parser.add_argument(
"--top_k",
type=int,
default=0,
help="Filter top-k tokens before sampling (<=0: no filtering)")
parser.add_argument(
"--top_p",
type=float,
default=0.9,
help="Nucleus filtering (top-p) before sampling (<=0.0: no filtering)")
args = parser.parse_args()
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__file__)
logger.info(pformat(args))
# if args.model_checkpoint == "":
# args.model_checkpoint = download_pretrained_model()
random.seed(args.seed)
torch.random.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
logger.info("Get pretrained model and tokenizer")
tokenizer = OpenAIGPTTokenizer.from_pretrained(args.model_checkpoint)
model = OpenAIGPTDoubleHeadsModel.from_pretrained(args.model_checkpoint)
model.to(args.device)
model.eval()
# logger.info("Sample a personality")
# personalities = get_dataset_personalities(tokenizer, args.dataset_path, args.dataset_cache)
# personality = random.choice(personalities)
# logger.info("Selected personality: %s", tokenizer.decode(chain(*personality)))
history = []
#examplepara = "Evidence of prehistoric activity in the area comes from Ashton Moss – a 107-hectare (260-acre) peat bog – and is the only one of Tameside's 22 Mesolithic sites not located in the hilly uplands in the north east of the borough. A single Mesolithic flint tool has been discovered in the bog,[6][7] along with a collection of nine Neolithic flints.[8] There was further activity in or around the bog in the Bronze Age. In about 1911, an adult male skull was found in the moss; it was thought to belong to the Romano-British period – similar to the Lindow Man bog body – until radiocarbon dating revealed that it dated from 1,320–970 BC"
#examplepara = tokenizer.encode(examplepara)
search = Searcher()
raw_text = input(">>> ")
start_time = time.time()
while not raw_text:
print('Prompt should not be empty!')
raw_text = input(">>> ")
start_time = time.time()
articlelist = search.searchandsplit(raw_text)
query = tokenizer.encode(raw_text)
toplist = []
topresults = 3
topmcs = [0.01] * topresults
threshold = 0.01
with torch.no_grad():
for arti in articlelist:
for para in arti:
txtpara = para
para = tokenizer.encode(para)
out_ids, mc = sample_sequence(query,
para,
tokenizer,
model,
args,
threshold=threshold)
out_text = tokenizer.decode(out_ids, skip_special_tokens=True)
mcs = mc.item()
if mcs > topmcs[0]:
toplist.append([mcs, out_text, txtpara])
print(f"Answer propability: {mcs}\n")
print(out_text)
topmcs.append(mcs)
topmcs.sort()
del topmcs[0]
sortedresults = sorted(toplist, key=lambda x: x[0], reverse=True)
toprange = min([topresults, len(sortedresults)])
for i in range(toprange):
print("\n\n")
print(f"Top {i}\n")
print(f"Answer propability: {sortedresults[i][0]}\n")
print("Answer: " + sortedresults[i][1] + "\n")
print("Paragraph for this answer: " + sortedresults[i][2])
print("Number of paragraphs searched")
print(len(sortedresults))
finaltime = time.time() - start_time
print(f"Processing finished after {finaltime}")
