def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir) # Writes entries directly to event files in the logdir to be consumed by TensorBoard.
device, args.gpu_ids = util.get_available_devices()
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
args.batch_size *= max(1, len(args.gpu_ids)) # 一个gpu： batch_size=64 看实际情况
# Set random seed
log.info(f'Using random seed {args.seed}...')
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path: # default=None
log.info(f'Loading checkpoint from {args.load_path}...')
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay) # ema_decay = 0.999
# ema core => new_average = (1.0 - decay) * param.data + decay * self.shadow[name]
# Get saver
# metric_name: NLL or EM F1
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints, # max_checkpoints = 5
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(), args.lr,
weight_decay=args.l2_wd) # lr : default=0.5 l2_wd : default=0
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2) # train_record_file = './data/train.npz'
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size, # 64
shuffle=True, # sampler = RandomSampler(dataset) batch_sampler = BatchSampler(sampler, batch_size, drop_last)
num_workers=args.num_workers, # 4
collate_fn=collate_fn) # merges a list of samples to form a mini-batch.
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2) # dev_record_file = './data/dev.npz'
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)# Merge examples of different length by padding all examples to the maximum length in the batch.
# Train
log.info('Training...')
steps_till_eval = args.eval_steps # 50000
epoch = step // len(train_dataset) # len(train_dataset)= 7 epoch=0
while epoch != args.num_epochs: # 30
epoch += 1
log.info(f'Starting epoch {epoch}...')
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0) # 64
optimizer.zero_grad()
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) # max_grad_norm : default=5.0
optimizer.step() # 进行1次optimize
scheduler.step(step // batch_size)# train : step=0
ema(model, step // batch_size) # def __call__(self, model, num_updates):
# Log info
step += batch_size #step: 0 batch_size=64
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch,
NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step) # Add scalar data to summary.
tbx.add_scalar('train/LR',
optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps # 50000
# Evaluate and save checkpoint
log.info(f'Evaluating at step {step}...')
ema.assign(model) ##
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file, # './data/dev_eval.json'
args.max_ans_len, # 15
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'Dev {results_str}')
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar(f'dev/{k}', v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
print("in main")
print("args: ", args)
if True:
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info('Using random seed {}...'.format(args.seed))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
# CHECK IF WE NEED TO USE ALL OF THESE????
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info('Loading checkpoint from {}...'.format(args.load_path))
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(), args.lr,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
print("train dataset!: ", train_dataset)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
while epoch != args.num_epochs:
epoch += 1
log.info('Starting epoch {}...'.format(epoch))
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch,
NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR',
optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info('Evaluating at step {}...'.format(step))
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
log.info('Dev {}'.format(results_str))
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar('dev/{}'.format(k), v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info('Using random seed {}...'.format(args.seed))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
char_vectors = util.torch_from_json(args.char_emb_file)
# ######################################
# tokenizer = BertTokenizer.from_pretrained('bert-large-uncased', do_lower_case=True)
# train_examples = None
# train_examples = read_squad_examples(
# input_file=args.train_file, is_training=True, version_2_with_negative=args.version_2_with_negative)
# train_features = convert_examples_to_features(
# examples=train_examples,
# tokenizer=tokenizer,
# max_seq_length=args.max_seq_length,
# doc_stride=args.doc_stride,
# max_query_length=args.max_query_length,
# is_training=True)
# if args.local_rank == -1 or torch.distributed.get_rank() == 0:
# logger.info(" Saving train features into cached file %s", cached_train_features_file)
# with open(cached_train_features_file, "wb") as writer:
# pickle.dump(train_features, writer)
# all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
# x = all_input_ids
###########################################
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info('Loading checkpoint from {}...'.format(args.load_path))
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(),
args.lr,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
while epoch != args.num_epochs:
epoch += 1
log.info('Starting epoch {}...'.format(epoch))
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
# added_flag
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
optimizer.zero_grad()
# Forward
# log_p1, log_p2 = model(cw_idxs, qw_idxs)
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch, NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR', optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info('Evaluating at step {}...'.format(step))
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
log.info('Dev {}'.format(results_str))
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar('dev/{}'.format(k), v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info(f'Using random seed {args.seed}...')
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info(f'Loading checkpoint from {args.load_path}...')
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(),
args.lr,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
while epoch != args.num_epochs:
epoch += 1
log.info(f'Starting epoch {epoch}...')
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
if step % 1000 == 0 and step > 0:
log.info(f'Step {step}: training loss {loss_val}...')
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info(f'Evaluating at step {step}...')
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}'
for k, v in results.items())
log.info(f'Dev {results_str}')
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar(f'dev/{k}', v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
# Comment out to only use 1 GPU on nv12
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info('Using random seed {}...'.format(args.seed))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
model = None
max_context_len, max_question_len = args.para_limit, args.ques_limit
if (args.model_type == "bidaf" or args.model_type == "bert-bidaf"):
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif (args.model_type == "dcn" or args.model_type == "bert-dcn"):
model = DCN(word_vectors=word_vectors,
hidden_size=args.hidden_size,
max_context_len=max_context_len,
max_question_len=max_question_len,
drop_prob=args.drop_prob)
elif (args.model_type == "bert-basic"):
model = BERT(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
if model is None:
raise ValueError('Model is unassigned. Please ensure --model_type \
chooses between {bidaf, bert-bidaf, dcn, bert-dcn, bert-basic} ')
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info('Loading checkpoint from {}...'.format(args.load_path))
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(),
args.lr,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
count_skip = 0
while epoch != args.num_epochs:
epoch += 1
log.info('Starting epoch {}...'.format(epoch))
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
batch_size = cw_idxs.size(0)
count_skip += 1
if (args.skip_examples == True
and (count_skip % 5 == 1 or count_skip % 5 == 2
or count_skip % 5 == 3 or count_skip % 5 == 4)):
step += batch_size
progress_bar.update(batch_size)
steps_till_eval -= batch_size
continue
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
## Additions for BERT ##
max_context_len, max_question_len = args.para_limit, args.ques_limit
if "bert" in args.model_type:
bert_train_embeddings = get_embeddings(
"train", ids, args.para_limit, args.ques_limit)
else:
bert_train_embeddings = None
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs, bert_train_embeddings, \
max_context_len, max_question_len, device)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch, NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR', optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info('Evaluating at step {}...'.format(step))
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2, args)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
log.info('Dev {}'.format(results_str))
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar('dev/{}'.format(k), v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging and devices
name = "train_exp2"
args.save_dir = util.get_save_dir(args.logging_dir, name, training=True)
log = get_logger(args.save_dir, name)
tbx = SummaryWriter(args.save_dir)
device, gpu_ids = util.get_available_devices()
log.info(f"Args: {dumps(vars(args), indent=4, sort_keys=True)}")
args.batch_size *= max(1, len(gpu_ids))
# Set random seed
log.info(f"Using random seed {args.random_seed}...")
random.seed(args.random_seed)
np.random.seed(args.random_seed)
torch.manual_seed(args.random_seed)
torch.cuda.manual_seed_all(args.random_seed)
# Get embeddings
log.info(f"Loading embeddings from {args.word_emb_file}...")
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info("Building model...")
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
model = nn.DataParallel(model, gpu_ids)
if args.load_path:
log.info(f"Loading checkpoint from {args.load_path}...")
model, step = util.load_model(model, args.load_path, gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(), args.learning_rate,
weight_decay=args.learning_rate_decay)
# scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
scheduler = sched.ReduceLROnPlateau(optimizer=optimizer,
mode="min", factor=0.1,
patience=2, verbose=True, cooldown=0
min_lr=0.0005)
for epoch in range(args.num_epochs):
log.info(f"Starting epoch {epoch}...")
for i in range(args.num_train_chunks):
# Get data loader
train_rec_file = f"{args.train_record_file_exp2}_{i}.npz"
log.info(f'Building dataset from {train_rec_file} ...')
train_dataset = SQuAD(train_rec_file, args.exp2_train_topic_contexts, use_v2=True)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = 0
# torch.set_num_threads(7)
with torch.enable_grad(), tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = qw_idxs.size(0)
optimizer.zero_grad()
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch,
NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR',
optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info(f"Evaluating at step {step}...")
ema.assign(model)
for i in range(args.num_dev_chunks):
# Get data loader
all_pred_dicts = {}
all_results = OrderedDict()
dev_rec_file = f"{args.dev_record_file_exp2}_{i}.npz"
log.info(f'Building evaluating dataset from {dev_rec_file} ...')
dev_dataset = SQuAD(dev_rec_file,
args.exp2_dev_topic_contexts,
use_v2=True)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
use_squad_v2=True)
all_results.update(results)
all_pred_dicts.update(pred_dict)
del dev_dataset
del dev_loader
del results
del pred_dict
torch.cuda.empty_cache()
saver.save(step, model, all_results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in all_results.items())
log.info(f"Dev {results_str}")
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in all_results.items():
tbx.add_scalar(f"dev/{k}", v, step)
util.visualize(tbx,
pred_dict=all_pred_dicts,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
torch.cuda.empty_cache()
del train_dataset
del train_loader
torch.cuda.empty_cache()
def main(args):
save_dir = os.path.join("./save", time.strftime("%m%d%H%M%S"))
if not os.path.exists(save_dir):
os.makedirs(save_dir)
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
if args.all_data:
data_loader = get_ext_data_loader(tokenizer,
"./data/train/",
shuffle=True,
args=args)
else:
data_loader, _, _ = get_data_loader(tokenizer,
"./data/train-v1.1.json",
shuffle=True,
args=args)
vocab_size = len(tokenizer.vocab)
if args.bidaf:
print("train bidaf")
model = BiDAF(embedding_size=args.embedding_size,
vocab_size=vocab_size,
hidden_size=args.hidden_size,
drop_prob=args.dropout)
else:
ntokens = len(tokenizer.vocab)
model = QANet(ntokens,
embedding=args.embedding,
embedding_size=args.embedding_size,
hidden_size=args.hidden_size,
num_head=args.num_head)
if args.load_model:
state_dict = torch.load(args.model_path, map_location="cpu")
model.load_state_dict(state_dict)
print("load pre-trained model")
device = torch.device("cuda")
model = model.to(device)
model.train()
ema = EMA(model, args.decay)
base_lr = 1
parameters = filter(lambda param: param.requires_grad, model.parameters())
optimizer = optim.Adam(lr=base_lr,
betas=(0.9, 0.999),
eps=1e-7,
weight_decay=5e-8,
params=parameters)
cr = args.lr / math.log2(args.lr_warm_up_num)
scheduler = optim.lr_scheduler.LambdaLR(
optimizer,
lr_lambda=lambda ee: cr * math.log2(ee + 1)
if ee < args.lr_warm_up_num else args.lr)
step = 0
num_batches = len(data_loader)
avg_loss = 0
best_f1 = 0
for epoch in range(1, args.num_epochs + 1):
step += 1
start = time.time()
model.train()
for i, batch in enumerate(data_loader, start=1):
c_ids, q_ids, start_positions, end_positions = batch
c_len = torch.sum(torch.sign(c_ids), 1)
max_c_len = torch.max(c_len)
c_ids = c_ids[:, :max_c_len].to(device)
q_len = torch.sum(torch.sign(q_ids), 1)
max_q_len = torch.max(q_len)
q_ids = q_ids[:, :max_q_len].to(device)
start_positions = start_positions.to(device)
end_positions = end_positions.to(device)
optimizer.zero_grad()
loss = model(c_ids,
q_ids,
start_positions=start_positions,
end_positions=end_positions)
loss.backward()
avg_loss = cal_running_avg_loss(loss.item(), avg_loss)
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step(step)
ema(model, step // args.batch_size)
batch_size = c_ids.size(0)
step += batch_size
msg = "{}/{} {} - ETA : {} - qa_loss: {:.4f}" \
.format(i, num_batches, progress_bar(i, num_batches),
eta(start, i, num_batches),
avg_loss)
print(msg, end="\r")
if not args.debug:
metric_dict = eval_qa(args, model)
f1 = metric_dict["f1"]
em = metric_dict["exact_match"]
print("epoch: {}, final loss: {:.4f}, F1:{:.2f}, EM:{:.2f}".format(
epoch, avg_loss, f1, em))
if args.bidaf:
model_name = "bidaf"
else:
model_name = "qanet"
if f1 > best_f1:
best_f1 = f1
state_dict = model.state_dict()
save_file = "{}_{:.2f}_{:.2f}".format(model_name, f1, em)
path = os.path.join(save_dir, save_file)
torch.save(state_dict, path)
def main(data, flags):
# Set up logging and devices
log_dir = data.logging_dir
log = util.get_logger(log_dir, "toy")
tbx = SummaryWriter(data.logging_dir)
device, data.gpu_ids = util.get_available_devices()
log.info('Config: {}'.format(dumps(vars(data), indent=4, sort_keys=True)))
data.batch_size *= max(1, len(data.gpu_ids))
# Set random seed
log.info('Using random seed {}...'.format(data.random_seed))
random.seed(data.random_seed)
np.random.seed(data.random_seed)
torch.manual_seed(data.random_seed)
torch.cuda.manual_seed_all(data.random_seed)
if flags[1] == "toy":
word_emb_file = data.toy_word_emb_file
training_data = data.toy_record_file_exp3
test_data = data.dev_record_file_exp3
eval_file = data.toy_eval_exp3
elif flags[1] == "train":
word_emb_file = data.word_emb_file
training_data = data.train_record_file_exp3
test_data = data.dev_record_file_exp3
eval_file = data.train_eval_exp3
elif flags[1] == "dev":
word_emb_file = data.word_emb_file
training_data = data.dev_record_file_exp3
test_data = data.toy_record_file_exp3
eval_file = data.dev_eval_exp3
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(word_emb_file)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
hidden_size=data.hidden_size,
drop_prob=data.drop_prob)
model = nn.DataParallel(model, data.gpu_ids)
if data.load_path:
log.info('Loading checkpoint from {}...'.format(data.load_path))
model, step = util.load_model(model, data.load_path, data.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, data.ema_decay)
# Get saver
saver = util.CheckpointSaver(data.logging_dir,
max_checkpoints=10,
metric_name=data.metric_name,
maximize_metric=data.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(),
data.learning_rate,
weight_decay=data.learning_weight_decay)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
# np.load(data.toy_record_file_exp3)
train_dataset = SQuAD3(training_data, use_v2=True)
train_loader = torchdata.DataLoader(train_dataset,
batch_size=data.batch_size,
shuffle=True,
num_workers=data.num_workers,
collate_fn=collate_fn)
test_dataset = SQuAD3(test_data, use_v2=True)
test_loader = torchdata.DataLoader(test_dataset,
batch_size=data.batch_size,
shuffle=False,
num_workers=data.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = data.eval_steps
epoch = step // len(test_dataset)
while epoch != data.num_epochs:
epoch += 1
log.info('Starting epoch {}...'.format(epoch))
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
# Forward
log.info("cw_idxs length: {}".format(str(len(cw_idxs))))
log.info("qw_idxs length: {}".format(str(len(qw_idxs))))
log.info("cw_idxs size: {}".format(str(
sys.getsizeof(cw_idxs))))
log.info("qw_idxs size: {}".format(str(
sys.getsizeof(qw_idxs))))
log.info("cw_idxs shape: {}".format(str(cw_idxs.shape)))
log.info("qw_idxs shape: {}".format(str(qw_idxs.shape)))
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(),
data.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch, NLL=loss_val)
tbx.add_scalar('toy/NLL', loss_val, step)
tbx.add_scalar('toy/LR', optimizer.param_groups[0]['lr'], step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = data.eval_steps
# Evaluate and save checkpoint
log.info('Evaluating at step {}...'.format(step))
ema.assign(model)
results, pred_dict = evaluate(model,
test_loader,
device,
eval_path=eval_file,
max_len=sys.maxsize,
use_squad_v2=True)
saver.save(step, model, results[data.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
log.info('Dev {}'.format(results_str))
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar('dev/{}'.format(k), v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=step,
split='dev',
num_visuals=data.num_visuals)
def train(args):
# Set up logging and devices
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info(f"Args: {dumps(vars(args), indent=4, sort_keys=True)}")
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info(f"Using random seed {args.seed}...")
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info("Loading embeddings...")
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info("Building model...")
model = BiDAF(
word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob,
use_glove=args.use_glove,
)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info(f"Loading checkpoint from {args.load_path}...")
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = stats.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(
args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log,
)
# Get optimizer and scheduler
optimizer = optim.Adam(model.parameters(),
args.lr,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.0) # Constant LR
# Get data loader
log.info("Building dataset...")
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn,
)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(
dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn,
)
# Train
log.info("Training...")
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
scaler = amp.GradScaler()
while epoch != args.num_epochs:
epoch += 1
log.info(f"Starting epoch {epoch}...")
with torch.enable_grad(), tqdm(
total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
# Forward
with amp.autocast():
log_p1, log_p2 = model(cw_idxs, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
scaler.step(optimizer)
scaler.update()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch, NLL=loss_val)
tbx.add_scalar("train/NLL", loss_val, step)
tbx.add_scalar("train/LR", optimizer.param_groups[0]["lr"],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info(f"Evaluating at step {step}...")
ema.assign(model)
results, pred_dict = evaluate(
model,
dev_loader,
device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2,
)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ", ".join(f"{k}: {v:05.2f}"
for k, v in results.items())
log.info(f"Dev {results_str}")
# Log to TensorBoard
log.info("Visualizing in TensorBoard...")
for k, v in results.items():
tbx.add_scalar(f"dev/{k}", v, step)
util.visualize(
tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split="dev",
num_visuals=args.num_visuals,
)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info('Using random seed {}...'.format(args.seed))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
char_vectors = util.torch_from_json(args.char_emb_file)
# Get model
log.info('Building model...')
model = BiDAF(vectors=(word_vectors, char_vectors),
hidden_size=args.hidden_size,
drop_prob=args.drop_prob,
p_sdd=args.p_sdd,
char_limit=args.char_limit,
use_transformer=args.use_transformer,
inter_size=args.inter_size,
heads=args.heads,
c2w_size=args.c2w_size,
enc_blocks=args.enc_blocks,
enc_convs=args.enc_convs,
mod_blocks=args.mod_blocks,
mod_convs=args.mod_convs,
use_GRU=args.use_GRU)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info('Loading checkpoint from {}...'.format(args.load_path))
model, step = util.load_model(model, args.load_path,
args.gpu_ids) # uses the saved step num
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
# optimizer = optim.Adadelta(model.parameters(), args.lr,
# weight_decay=args.l2_wd)
# The scheduler MULTIPLIES the base LR, NOT replaces
optimizer = optim.Adam(model.parameters(),
1.,
betas=(.9, .98),
eps=1e-9,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(
optimizer, lambda s: 0.001 * math.log(s + 1) / math.log(1000 - 1)
if s < 1000 else 0.001) # Chute (must use math.log, else TypeError)
# scheduler = sched.LambdaLR(optimizer, lambda s: (args.hidden_size**(-.5)) *
# min((s+1e-9)**(-.5), s*(4000**(-1.5)))
# ) # From Vaswani et. al 2017
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
while epoch != args.num_epochs:
epoch += 1
log.info('Starting epoch {}...'.format(epoch))
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
optimizer.zero_grad()
batch_size = cw_idxs.size(0)
cc_idxs = cc_idxs.to(device) # (batch, c_limit, char_limit)
qc_idxs = qc_idxs.to(device)
cw_idxs = cw_idxs.to(device) # (batch, c_limit)
qw_idxs = qw_idxs.to(device)
c_idxs, q_idxs = (cw_idxs, cc_idxs), (qw_idxs, qc_idxs)
# Forward
log_p1, log_p2 = model(c_idxs, q_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step(
step // batch_size
) # By default, schedules per epoch; pass in step # as "epoch"
ema(model, step // batch_size)
# Log info
step += batch_size # Number of examples. Step is usually the number of (mini)-batches
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch, NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR', optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info('Evaluating at step {}...'.format(step))
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
log.info('Dev {}'.format(results_str))
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar('dev/{}'.format(k), v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info(f'Using random seed {args.seed}...')
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
char_vec = util.torch_from_json(args.char_emb_file)
# Get model
log.info('Building model...')
if args.name == 'baseline':
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif args.name == 'charembeddings':
model = BiDAFChar(word_vectors=word_vectors,
char_vec=char_vec,
word_len=16,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif args.name == 'charembeddings2':
model = BiDAFChar2(word_vectors=word_vectors,
char_vec=char_vec,
word_len=16,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif args.name == 'qanet':
model = QANet(word_vectors=word_vectors,
char_vec=char_vec,
word_len=16,
emb_size=args.hidden_size,
drop_prob=args.drop_prob,
enc_size=args.enc_size,
n_head=args.n_head,
LN_train=args.ln_train,
DP_residual=args.dp_res,
mask_pos=args.mask_pos,
two_pos=args.two_pos,
total_prob=args.total_drop,
final_prob=args.final_prob)
elif args.name == 'qanet2':
model = QANet2(word_vectors=word_vectors,
char_vec=char_vec,
word_len=16,
emb_size=args.hidden_size,
drop_prob=args.drop_prob,
enc_size=args.enc_size,
n_head=args.n_head,
LN_train=args.ln_train,
DP_residual=args.dp_res,
mask_pos=args.mask_pos,
two_pos=args.two_pos,
rel=args.rel_att,
total_prob=args.total_drop,
final_prob=args.final_prob,
freeze=args.freeze_emb)
elif args.name == 'qanet3':
model = QANet3(word_vectors=word_vectors,
char_vec=char_vec,
word_len=16,
emb_size=args.hidden_size,
drop_prob=args.drop_prob,
enc_size=args.enc_size,
n_head=args.n_head,
LN_train=args.ln_train,
DP_residual=args.dp_res,
mask_pos=args.mask_pos,
two_pos=args.two_pos,
rel=args.rel_att,
total_prob=args.total_drop,
final_prob=args.final_prob,
freeze=args.freeze_emb)
elif args.name == 'qanet4':
model = QANet4(word_vectors=word_vectors,
char_vec=char_vec,
word_len=16,
emb_size=args.hidden_size,
drop_prob=args.drop_prob,
enc_size=args.enc_size,
n_head=args.n_head,
LN_train=args.ln_train,
DP_residual=args.dp_res,
mask_pos=args.mask_pos,
two_pos=args.two_pos,
rel=args.rel_att,
total_prob=args.total_drop,
final_prob=args.final_prob,
freeze=args.freeze_emb)
else:
raise ValueError('Wrong model name')
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info(f'Loading checkpoint from {args.load_path}...')
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
if args.name == 'qanet':
optimizer = optim.Adam(model.parameters(),
args.lr,
betas=(0.8, 0.999),
weight_decay=3 * 1e-7,
eps=1e-7)
scheduler = warmup(optimizer, 1, 2000)
elif args.opt == 'adam':
if args.grad_cent:
optimizer = AdamWGC(model.parameters(),
args.lr,
betas=(0.9, 0.999),
weight_decay=3 * 1e-7,
eps=1e-7,
use_gc=True)
else:
optimizer = AdamW(model.parameters(),
args.lr,
betas=(0.8, 0.999),
weight_decay=3 * 1e-7,
eps=1e-7)
scheduler = warmup(optimizer, 1, 2000)
elif args.opt == 'adadelta':
optimizer = optim.Adadelta(model.parameters(),
args.lr,
weight_decay=3 * 1e-7)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
elif args.opt == 'sgd':
optimizer = optim.SGD(model.parameters(),
args.lr,
weight_decay=3 * 1e-7)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
i = 0
while epoch != args.num_epochs:
epoch += 1
log.info(f'Starting epoch {epoch}...')
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
log_p1, log_p2 = model(cw_idxs, cc_idxs, qw_idxs, qc_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
i += 1
loss /= args.acc_step
# Backward
loss.backward()
if i % args.acc_step == 0:
nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step(i // (args.acc_step))
ema(model, i // (args.acc_step))
optimizer.zero_grad()
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch, NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR', optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0 and i % args.acc_step == 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info(f'Evaluating at step {step}...')
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}'
for k, v in results.items())
log.info(f'Dev {results_str}')
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar(f'dev/{k}', v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
args.batch_size *= max(1, len(args.gpu_ids))
max_len = 10
# Set random seed
log.info(f'Using random seed {args.seed}...')
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
ch_vectors = util.torch_from_json(args.char_emb_file)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
ch_vectors=ch_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info(f'Loading checkpoint from {args.load_path}...')
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(),
args.lr,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
while epoch != args.num_epochs:
epoch += 1
log.info(f'Starting epoch {epoch}...')
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
# Forward
log_p = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
y1, y2 = y1.to(device), y2.to(device)
#print("ckpt 1")
ans_lens = y2 - y1
loss = 0
for i in range(max_len):
mask = ((torch.ones_like(y1) * i) == ans_lens).type(
torch.cuda.LongTensor)
y = y1 * mask
loss += F.nll_loss(log_p[:, :, i], y)
#print("ckpt 2")
loss_val = loss.item()
#print("ckpt 3")
# Backward
loss.backward()
#print("ckpt 4")
nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
#print("ckpt 5")
optimizer.step()
#print("ckpt 6")
scheduler.step(step // batch_size)
ema(model, step // batch_size)
#print("ckpt 7")
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch, NLL=loss_val)
if step % (50 * batch_size) == 0:
print(loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR', optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
#print("ckpt 8")
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info(f'Evaluating at step {step}...')
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}'
for k, v in results.items())
log.info(f'Dev {results_str}')
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar(f'dev/{k}', v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
# To make the data generation of every experiment same
log.info(f'Using random seed {args.seed}...')
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
char_vectors = util.torch_from_json(args.char_emb_file)
print(word_vectors.size())
print(char_vectors.size())
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info(f'Loading checkpoint from {args.load_path}...')
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(),
args.lr,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
# (context_idxs(context_len,): Indices of the words in the context.,
# context_char_idx(context_len, max_word_len): Indices of the characters in the context,
# question_idxs(question_len,): Indices of the words in the question,
# question_char_idx(question_len, max_word_len): Indices of the characters in the question,
# y1:start, -1 if no answer: answer start index,
# y2:start, -1 if no answer: answer end index,
# id ID of the example)
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
while epoch != args.num_epochs:
epoch += 1
log.info(f'Starting epoch {epoch}...')
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
# cw_idxs: Indices of the words in the context
# cc_idxs: Indices of the characters in the context
# qw_idxs: Indices of the words in the query
# qc_idxs: Indices of the characters in teh query
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
# cw_idx with shape(context_len, )
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
# Forward
log_p1, log_p2 = model(cw_idxs, cc_idxs, qw_idxs, qc_idxs)
y1, y2 = y1.to(device), y2.to(device)
# L(theta) = - 1/N * sum(log(P1_yi_1) + log(P2_yi_2))
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch, NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR', optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info(f'Evaluating at step {step}...')
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}'
for k, v in results.items())
log.info(f'Dev {results_str}')
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar(f'dev/{k}', v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info(f'Using random seed {args.seed}...')
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
'''
model = QANet(word_vectors, args.hidden_size, args.char_embed_size, args.word_from_char_size,
args.dropout_main,
args.embed_encoder_num_convs, args.embed_encoder_conv_kernel_size,
args.embed_encoder_num_heads, args.embed_encoder_num_blocks,
args.model_encoder_num_convs, args.model_encoder_conv_kernel_size,
args.model_encoder_num_heads, args.model_encoder_num_blocks)
'''
char_vectors = util.torch_from_json(args.char_emb_file)
model = BiDAF(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info(f'Loading checkpoint from {args.load_path}...')
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
# Increase LR from 0 to args.lr in num_warmup_steps steps, then keep constant LR
optimizer = optim.Adam(model.parameters(), args.lr, betas=(0.8, 0.999), eps=1e-07, weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer,
lambda s: math.log(1+s)/math.log(args.num_warmup_steps) if s < args.num_warmup_steps else 1)
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
while epoch != args.num_epochs:
epoch += 1
log.info(f'Starting epoch {epoch}...')
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
# Forward
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch,
NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR',
optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info(f'Evaluating at step {step}...')
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'Dev {results_str}')
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar(f'dev/{k}', v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)
def main(args):
# Set up logging and devices
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
tbx = SummaryWriter(args.save_dir)
device, args.gpu_ids = util.get_available_devices()
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
args.batch_size *= max(1, len(args.gpu_ids))
# Set random seed
log.info('Using random seed {}...'.format(args.seed))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
char_vectors = util.torch_from_json(args.char_emb_file)
# NEW : load the tag embeddings
pos_vectors = util.torch_from_json(args.pos_emb_file)
ner_vectors = util.torch_from_json(args.ner_emb_file)
# Add loss
if 'loss' in args.name:
distance_criterion = DistanceFromAnswerLoss(coefficient=.5, device=device,
normalization=True, penalization_type='quadratic',
reduction='mean')
# Choose model
log.info('Building model {}...'.format(args.name))
if 'baseline' in args.name:
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif args.name == 'BiDAF_char':
model = BiDAF_char(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif (args.name == 'BiDAF_tag') or (args.name == 'BiDAF_tag_loss'):
model = BiDAF_tag(word_vectors=word_vectors,
char_vectors=char_vectors,
pos_vectors=pos_vectors,
ner_vectors=ner_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif (args.name == 'BiDAF_tag_unfrozen') or (args.name == 'BiDAF_tag_unfrozen_loss'):
model = BiDAF_tag(word_vectors=word_vectors,
char_vectors=char_vectors,
pos_vectors=pos_vectors,
ner_vectors=ner_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob,
freeze_tag=False)
elif args.name == 'BiDAF_tag_ext':
model = BiDAF_tag_ext(word_vectors=word_vectors,
char_vectors=char_vectors,
pos_vectors=pos_vectors,
ner_vectors=ner_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif args.name == 'BiDAF_tag_ext_unfrozen':
model = BiDAF_tag_ext(word_vectors=word_vectors,
char_vectors=char_vectors,
pos_vectors=pos_vectors,
ner_vectors=ner_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob,
freeze_tag=False)
elif args.name == 'coattn':
model = CoattentionModel(hidden_dim=args.hidden_size,
embedding_matrix=word_vectors,
train_word_embeddings=False,
dropout=0.35,
pooling_size=16,
number_of_iters=4,
number_of_layers=2,
device=device)
else:
raise NameError('No model named ' + args.name)
model = nn.DataParallel(model, args.gpu_ids)
if args.load_path:
log.info('Loading checkpoint from {}...'.format(args.load_path))
model, step = util.load_model(model, args.load_path, args.gpu_ids)
else:
step = 0
model = model.to(device)
model.train()
ema = util.EMA(model, args.ema_decay)
# Get saver
saver = util.CheckpointSaver(args.save_dir,
max_checkpoints=args.max_checkpoints,
metric_name=args.metric_name,
maximize_metric=args.maximize_metric,
log=log)
# Get optimizer and scheduler
optimizer = optim.Adadelta(model.parameters(), args.lr,
weight_decay=args.l2_wd)
scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Get data loader
log.info('Building dataset...')
train_dataset = SQuAD(args.train_record_file, args.use_squad_v2)
train_loader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn,
drop_last=True)
dev_dataset = SQuAD(args.dev_record_file, args.use_squad_v2)
dev_loader = data.DataLoader(dev_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn,
drop_last=True)
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
while epoch != args.num_epochs:
epoch += 1
log.info('Starting epoch {}...'.format(epoch))
with torch.enable_grad(), \
tqdm(total=len(train_loader.dataset)) as progress_bar:
for cw_idxs, cc_idxs, cpos_idxs, cner_idxs, cw_ems, cw_tfs, qw_idxs, qc_idxs, qpos_idxs, qner_idxs, qw_ems, qw_tfs, y1, y2, ids in train_loader: # NEW
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
# Forward
if 'baseline' in args.name:
log_p1, log_p2 = model(cw_idxs, qw_idxs)
elif args.name == 'BiDAF_char':
# Additional setup for forward
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs)
elif (args.name == 'BiDAF_tag') or (args.name == 'BiDAF_tag_unfrozen') or (args.name == 'BiDAF_tag_loss') or (args.name == 'BiDAF_tag_unfrozen_loss'):
# Additional setup for forward
cc_idxs = cc_idxs.to(device)
cpos_idxs = cpos_idxs.to(device)
cner_idxs = cner_idxs.to(device)
qc_idxs = qc_idxs.to(device)
qpos_idxs = qpos_idxs.to(device)
qner_idxs = qner_idxs.to(device)
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs, cpos_idxs, qpos_idxs, cner_idxs, qner_idxs)
elif (args.name == 'BiDAF_tag_ext') or (args.name == 'BiDAF_tag_ext_unfrozen'):
# Additional setup for forward
cc_idxs = cc_idxs.to(device)
cpos_idxs = cpos_idxs.to(device)
cner_idxs = cner_idxs.to(device)
cw_ems = cw_ems.to(device)
cw_tfs = cw_tfs.to(device)
qc_idxs = qc_idxs.to(device)
qpos_idxs = qpos_idxs.to(device)
qner_idxs = qner_idxs.to(device)
qw_ems = qw_ems.to(device)
qw_tfs = qw_tfs.to(device)
log_p1, log_p2 = model(cw_idxs, qw_idxs, cc_idxs, qc_idxs, cpos_idxs, qpos_idxs, cner_idxs, qner_idxs, cw_ems, qw_ems, cw_tfs, qw_tfs)
elif args.name == 'coattn':
max_c_len = cw_idxs.size(1)
max_q_len = qw_idxs.size(1)
c_len = []
q_len = []
for i in range(cw_idxs.size(0)):
if len((cw_idxs[i] == 0).nonzero()) != 0:
c_len_i = (cw_idxs[i] == 0).nonzero()[0].item()
else:
c_len_i = cw_idxs.size(1)
if len((qw_idxs[i] == 0).nonzero()) != 0:
q_len_i = (qw_idxs[i] == 0).nonzero()[0].item()
else:
q_len_i = qw_idxs.size(1)
c_len.append(int(c_len_i))
q_len.append(int(q_len_i))
c_len = torch.Tensor(c_len).int()
q_len = torch.Tensor(q_len).int()
num_examples = int(cw_idxs.size(0) / len(args.gpu_ids))
log_p1, log_p2 = model(max_c_len, max_q_len, cw_idxs, qw_idxs, c_len, q_len, num_examples, True, True)
else:
raise NameError('No model named ' + args.name)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
# Add distance penalization
if 'loss' in args.name:
loss += distance_criterion(log_p1, y1) + distance_criterion(log_p2, y2)
loss_val = loss.item()
# Backward
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
scheduler.step(step // batch_size)
ema(model, step // batch_size)
# Log info
step += batch_size
progress_bar.update(batch_size)
progress_bar.set_postfix(epoch=epoch,
NLL=loss_val)
tbx.add_scalar('train/NLL', loss_val, step)
tbx.add_scalar('train/LR',
optimizer.param_groups[0]['lr'],
step)
steps_till_eval -= batch_size
if steps_till_eval <= 0:
steps_till_eval = args.eval_steps
# Evaluate and save checkpoint
log.info('Evaluating at step {}...'.format(step))
ema.assign(model)
results, pred_dict = evaluate(model, dev_loader, device,
args.dev_eval_file,
args.max_ans_len,
args.use_squad_v2,
args.name,
args.gpu_ids)
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
log.info('Dev {}'.format(results_str))
# Log to TensorBoard
log.info('Visualizing in TensorBoard...')
for k, v in results.items():
tbx.add_scalar('dev/{}'.format(k), v, step)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.dev_eval_file,
step=step,
split='dev',
num_visuals=args.num_visuals)