def train_entry(config):
from models import BiDAF
with open(config.word_emb_file, "rb") as fh:
word_mat = np.array(json.load(fh), dtype=np.float32)
with open(config.char_emb_file, "rb") as fh:
char_mat = np.array(json.load(fh), dtype=np.float32)
with open(config.dev_eval_file, "r") as fh:
dev_eval_file = json.load(fh)
print("Building model...")
train_dataset = get_loader(config.train_record_file, config.batch_size)
dev_dataset = get_loader(config.dev_record_file, config.batch_size)
c_vocab_size, c_emb_size = char_mat.shape
model = BiDAF(word_mat,
w_embedding_size=300,
c_embeding_size=c_emb_size,
c_vocab_size=c_vocab_size,
hidden_size=100,
drop_prob=0.2).to(device)
if config.pretrained:
print("load pre-trained model")
state_dict = torch.load(config.save_path, map_location="cpu")
model.load_state_dict(state_dict)
ema = EMA(config.decay)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
parameters = filter(lambda param: param.requires_grad, model.parameters())
optimizer = optim.Adadelta(lr=0.5, params=parameters)
best_f1 = 0
best_em = 0
patience = 0
for iter in range(config.num_epoch):
train(model, optimizer, train_dataset, dev_dataset, dev_eval_file,
iter, ema)
ema.assign(model)
metrics = test(model, dev_dataset, dev_eval_file,
(iter + 1) * len(train_dataset))
dev_f1 = metrics["f1"]
dev_em = metrics["exact_match"]
if dev_f1 < best_f1 and dev_em < best_em:
patience += 1
if patience > config.early_stop:
break
else:
patience = 0
best_f1 = max(best_f1, dev_f1)
best_em = max(best_em, dev_em)
fn = os.path.join(
config.save_dir,
"model_{}_{:.2f}_{:.2f}.pt".format(iter, best_f1, best_em))
torch.save(model.state_dict(), fn)
ema.resume(model)
def main(args):
log = util.get_logger(args.save_dir, args.name)
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Prepare BiDAF model (must already trained)
log.info('Building BiDAF model (should be pretrained)')
bidaf_model = BiDAF(word_vectors=word_vectors, # todo: these word vectors shouldn't matter?
hidden_size=args.hidden_size) # since they will be loaded in during load_model?
#drop_prob=args.drop_prob) # no drop probability since we are not training
#bidaf_model = nn.DataParallel(bidaf_model, args.gpu_ids)
#log.info(f'Loading checkpoint from {args.load_path}...')
#bidaf_model = util.load_model(bidaf_model, args.load_path, args.gpu_ids, return_step=False) # don't need step since we aren't training
#bidaf_model = bidaf_model.to(device)
bidaf_model.eval() # we eval only (vs train)
# Setup the Paraphraser model
#ema = util.EMA(bidaf_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
a = np.array([[5, 0, 3, 0],
[4, 0, 1, 0]])
c_idx = torch.from_numpy(a).long()
pp(c_idx[c_idx.nonzero()])
def predict(args, cw_idxs, qn_idxs):
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# 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)
model = nn.DataParallel(model, gpu_ids)
log.info('Loading checkpoint from {}...'.format(args.load_path))
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
y_pred = model(cw_idxs, qn_idxs)
return y_pred
def get_model(word_vectors, char_vectors, log, args):
if args.model_name == "BiDAF":
model = BiDAF_char(word_vectors=word_vectors,
char_vectors=char_vectors,
char_len=args.char_len,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif args.model_name == "BiDAF_nochar":
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif args.model_name == "KnowGQA":
model = KnowGQA(word_vectors=word_vectors,
char_vectors=char_vectors,
char_len=args.char_len,
hidden_size=args.hidden_size,
h=args.h,
drop_prob=args.drop_prob)
elif args.model_name == "QANet":
model = QANet(word_vectors=word_vectors,
char_vectors=char_vectors,
char_len=args.char_len,
hidden_size=args.hidden_size,
h=args.h,
drop_prob=args.drop_prob)
else:
raise ValueError("Model name doesn't exist.")
model = nn.DataParallel(model, args.gpu_ids)
'''
for p in model.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
'''
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
return model, step
def main(args):
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# 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)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
log.info('Building dataset...')
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
log.info(f'Evaluating on {args.split} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_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)
# 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)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
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):
# 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(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 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):
args.save_dir = util.get_save_dir(args.save_dir,
"exp1_training",
training=False)
log = get_logger(args.logging_dir, "exp1_training")
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# 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)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, c.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
log.info('Building dataset...')
dataset = SQuAD(args.test_record_file, True)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
log.info(f'Evaluating on {args.datasplit} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
with open(args.test_eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_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, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, c.max_ans_len, True)
# Log info
progress_bar.update(batch_size)
# Not using the unlabeled test set
# if args.split != 'test':
# # No labels for the test set, so NLL would be invalid
# progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(), True)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
results = util.eval_dicts(gold_dict, pred_dict, True)
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.datasplit} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(c.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=args.test_eval_file,
step=0,
split=args.datasplit,
num_visuals=args.num_visuals)
def main(args):
# Load TF-IDF from pickle
scorer = TFIDF([])
scorer.get_from_pickle()
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
log.info('Args: {}'.format(dumps(vars(args), indent=4, sort_keys=True)))
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get data loader
log.info('Building dataset...')
record_file = vars(args)['{}_record_file'.format(args.split)]
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Get model
log.info('Building model...')
model = BiDAF(word_vectors=word_vectors,
char_vocab_size= 1376,
hidden_size=args.hidden_size)
model = nn.DataParallel(model, gpu_ids)
log.info('Loading checkpoint from {}...'.format(args.load_path))
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Evaluate
log.info('Evaluating on {} split...'.format(args.split))
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)['{}_eval_file'.format(args.split)]
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_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, 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)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len, args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict,
ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
if (args.use_tfidf):
# Apply TF-IDF filtering to pred_dict
tf_idf_threshold = 2
tf_idf_common_threshold = 1
for key, value in pred_dict.items():
if value != "":
tf_idf_score = scorer.normalized_additive_idf_ignore_common_words(
value, threshold_frequency=tf_idf_common_threshold)
if tf_idf_score < tf_idf_threshold:
pred_dict[key] = ''
pass
# print ("pred_dict: {}, pruned".format(tf_idf_score))
else:
pass
# print ("pred_dict: {}, kept".format(tf_idf_score))
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('NLL', nll_meter.avg),
('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join('{}: {:05.2f}'.format(k, v)
for k, v in results.items())
log.info('{} {}'.format(args.split.title(), results_str))
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info('Writing submission file to {}...'.format(sub_path))
with open(sub_path, 'w') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
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)
def main(args):
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# 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...')
nbr_model = 0
if (args.load_path_baseline):
model_baseline = Baseline(word_vectors=word_vectors, hidden_size=100)
model_baseline = nn.DataParallel(model_baseline, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_baseline}...')
model_baseline = util.load_model(model_baseline,
args.load_path_baseline,
gpu_ids,
return_step=False)
model_baseline = model_baseline.to(device)
model_baseline.eval()
nll_meter_baseline = util.AverageMeter()
nbr_model += 1
save_prob_baseline_start = []
save_prob_baseline_end = []
if (args.load_path_bidaf):
model_bidaf = BiDAF(word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size)
model_bidaf = nn.DataParallel(model_bidaf, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_bidaf}...')
model_bidaf = util.load_model(model_bidaf,
args.load_path_bidaf,
gpu_ids,
return_step=False)
model_bidaf = model_bidaf.to(device)
model_bidaf.eval()
nll_meter_bidaf = util.AverageMeter()
nbr_model += 1
save_prob_bidaf_start = []
save_prob_bidaf_end = []
if (args.load_path_bidaf_fusion):
model_bidaf_fu = BiDAF_fus(word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size)
model_bidaf_fu = nn.DataParallel(model_bidaf_fu, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_bidaf_fusion}...')
model_bidaf_fu = util.load_model(model_bidaf_fu,
args.load_path_bidaf_fusion,
gpu_ids,
return_step=False)
model_bidaf_fu = model_bidaf_fu.to(device)
model_bidaf_fu.eval()
nll_meter_bidaf_fu = util.AverageMeter()
nbr_model += 1
save_prob_bidaf_fu_start = []
save_prob_bidaf_fu_end = []
if (args.load_path_qanet):
model_qanet = QANet(word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size,
n_heads=args.n_heads,
n_conv_emb_enc=args.n_conv_emb,
n_conv_mod_enc=args.n_conv_mod,
n_emb_enc_blocks=args.n_emb_blocks,
n_mod_enc_blocks=args.n_mod_blocks,
divisor_dim_kqv=args.divisor_dim_kqv)
model_qanet = nn.DataParallel(model_qanet, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_qanet}...')
model_qanet = util.load_model(model_qanet,
args.load_path_qanet,
gpu_ids,
return_step=False)
model_qanet = model_qanet.to(device)
model_qanet.eval()
nll_meter_qanet = util.AverageMeter()
nbr_model += 1
save_prob_qanet_start = []
save_prob_qanet_end = []
if (args.load_path_qanet_old):
model_qanet_old = QANet_old(word_vectors=word_vectors,
char_vectors=char_vectors,
device=device,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size,
n_heads=args.n_heads,
n_conv_emb_enc=args.n_conv_emb,
n_conv_mod_enc=args.n_conv_mod,
n_emb_enc_blocks=args.n_emb_blocks,
n_mod_enc_blocks=args.n_mod_blocks)
model_qanet_old = nn.DataParallel(model_qanet_old, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_qanet_old}...')
model_qanet_old = util.load_model(model_qanet_old,
args.load_path_qanet_old,
gpu_ids,
return_step=False)
model_qanet_old = model_qanet_old.to(device)
model_qanet_old.eval()
nll_meter_qanet_old = util.AverageMeter()
nbr_model += 1
save_prob_qanet_old_start = []
save_prob_qanet_old_end = []
if (args.load_path_qanet_inde):
model_qanet_inde = QANet_independant_encoder(
word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size,
n_heads=args.n_heads,
n_conv_emb_enc=args.n_conv_emb,
n_conv_mod_enc=args.n_conv_mod,
n_emb_enc_blocks=args.n_emb_blocks,
n_mod_enc_blocks=args.n_mod_blocks,
divisor_dim_kqv=args.divisor_dim_kqv)
model_qanet_inde = nn.DataParallel(model_qanet_inde, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_qanet_inde}...')
model_qanet_inde = util.load_model(model_qanet_inde,
args.load_path_qanet_inde,
gpu_ids,
return_step=False)
model_qanet_inde = model_qanet_inde.to(device)
model_qanet_inde.eval()
nll_meter_qanet_inde = util.AverageMeter()
nbr_model += 1
save_prob_qanet_inde_start = []
save_prob_qanet_inde_end = []
if (args.load_path_qanet_s_e):
model_qanet_s_e = QANet_S_E(word_vectors=word_vectors,
char_vectors=char_vectors,
char_emb_dim=args.char_emb_dim,
hidden_size=args.hidden_size,
n_heads=args.n_heads,
n_conv_emb_enc=args.n_conv_emb,
n_conv_mod_enc=args.n_conv_mod,
n_emb_enc_blocks=args.n_emb_blocks,
n_mod_enc_blocks=args.n_mod_blocks,
divisor_dim_kqv=args.divisor_dim_kqv)
model_qanet_s_e = nn.DataParallel(model_qanet_s_e, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path_qanet_s_e}...')
model_qanet_s_e = util.load_model(model_qanet_s_e,
args.load_path_qanet_s_e,
gpu_ids,
return_step=False)
model_qanet_s_e = model_qanet_s_e.to(device)
model_qanet_s_e.eval()
nll_meter_qanet_s_e = util.AverageMeter()
nbr_model += 1
save_prob_qanet_s_e_start = []
save_prob_qanet_s_e_end = []
# Get data loader
log.info('Building dataset...')
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
log.info(f'Evaluating on {args.split} split...')
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_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)
y1, y2 = y1.to(device), y2.to(device)
l_p1, l_p2 = [], []
# Forward
if (args.load_path_baseline):
log_p1_baseline, log_p2_baseline = model_baseline(
cw_idxs, cc_idxs)
loss_baseline = F.nll_loss(log_p1_baseline, y1) + F.nll_loss(
log_p2_baseline, y2)
nll_meter_baseline.update(loss_baseline.item(), batch_size)
l_p1 += [log_p1_baseline.exp()]
l_p2 += [log_p2_baseline.exp()]
if (args.save_probabilities):
save_prob_baseline_start += [
log_p1_baseline.exp().detach().cpu().numpy()
]
save_prob_baseline_end += [
log_p2_baseline.exp().detach().cpu().numpy()
]
if (args.load_path_qanet):
log_p1_qanet, log_p2_qanet = model_qanet(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_qanet = F.nll_loss(log_p1_qanet, y1) + F.nll_loss(
log_p2_qanet, y2)
nll_meter_qanet.update(loss_qanet.item(), batch_size)
# Get F1 and EM scores
l_p1 += [log_p1_qanet.exp()]
l_p2 += [log_p2_qanet.exp()]
if (args.save_probabilities):
save_prob_qanet_start += [
log_p1_qanet.exp().detach().cpu().numpy()
]
save_prob_qanet_end += [
log_p2_qanet.exp().detach().cpu().numpy()
]
if (args.load_path_qanet_old):
log_p1_qanet_old, log_p2_qanet_old = model_qanet_old(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_qanet_old = F.nll_loss(log_p1_qanet_old, y1) + F.nll_loss(
log_p2_qanet_old, y2)
nll_meter_qanet_old.update(loss_qanet_old.item(), batch_size)
# Get F1 and EM scores
l_p1 += [log_p1_qanet_old.exp()]
l_p2 += [log_p2_qanet_old.exp()]
if (args.save_probabilities):
save_prob_qanet_old_start += [
log_p1_qanet_old.exp().detach().cpu().numpy()
]
save_prob_qanet_old_end += [
log_p2_qanet_old.exp().detach().cpu().numpy()
]
if (args.load_path_qanet_inde):
log_p1_qanet_inde, log_p2_qanet_inde = model_qanet_inde(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_qanet_inde = F.nll_loss(
log_p1_qanet_inde, y1) + F.nll_loss(log_p2_qanet_inde, y2)
nll_meter_qanet_inde.update(loss_qanet_inde.item(), batch_size)
# Get F1 and EM scores
l_p1 += [log_p1_qanet_inde.exp()]
l_p2 += [log_p2_qanet_inde.exp()]
if (args.save_probabilities):
save_prob_qanet_inde_start += [
log_p1_qanet_inde.exp().detach().cpu().numpy()
]
save_prob_qanet_inde_end += [
log_p2_qanet_inde.exp().detach().cpu().numpy()
]
if (args.load_path_qanet_s_e):
log_p1_qanet_s_e, log_p2_qanet_s_e = model_qanet_s_e(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_qanet_s_e = F.nll_loss(log_p1_qanet_s_e, y1) + F.nll_loss(
log_p2_qanet_s_e, y2)
nll_meter_qanet_s_e.update(loss_qanet_s_e.item(), batch_size)
# Get F1 and EM scores
l_p1 += [log_p1_qanet_s_e.exp()]
l_p2 += [log_p2_qanet_s_e.exp()]
if (args.save_probabilities):
save_prob_qanet_s_e_start += [
log_p1_qanet_s_e.exp().detach().cpu().numpy()
]
save_prob_qanet_s_e_end += [
log_p2_qanet_s_e.exp().detach().cpu().numpy()
]
if (args.load_path_bidaf):
log_p1_bidaf, log_p2_bidaf = model_bidaf(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_bidaf = F.nll_loss(log_p1_bidaf, y1) + F.nll_loss(
log_p2_bidaf, y2)
nll_meter_bidaf.update(loss_bidaf.item(), batch_size)
l_p1 += [log_p1_bidaf.exp()]
l_p2 += [log_p2_bidaf.exp()]
if (args.save_probabilities):
save_prob_bidaf_start += [
log_p1_bidaf.exp().detach().cpu().numpy()
]
save_prob_bidaf_end += [
log_p2_bidaf.exp().detach().cpu().numpy()
]
if (args.load_path_bidaf_fusion):
log_p1_bidaf_fu, log_p2_bidaf_fu = model_bidaf_fu(
cw_idxs, cc_idxs, qw_idxs, qc_idxs)
loss_bidaf_fu = F.nll_loss(log_p1_bidaf_fu, y1) + F.nll_loss(
log_p2_bidaf_fu, y2)
nll_meter_bidaf_fu.update(loss_bidaf_fu.item(), batch_size)
l_p1 += [log_p1_bidaf_fu.exp()]
l_p2 += [log_p2_bidaf_fu.exp()]
if (args.save_probabilities):
save_prob_bidaf_fu_start += [
log_p1_bidaf_fu.exp().detach().cpu().numpy()
]
save_prob_bidaf_fu_end += [
log_p2_bidaf_fu.exp().detach().cpu().numpy()
]
p1, p2 = l_p1[0], l_p2[0]
for i in range(1, nbr_model):
p1 += l_p1[i]
p2 += l_p2[i]
p1 /= nbr_model
p2 /= nbr_model
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
if (args.load_path_qanet):
progress_bar.set_postfix(NLL=nll_meter_qanet.avg)
elif (args.load_path_bidaf):
progress_bar.set_postfix(NLL=nll_meter_bidaf.avg)
elif (args.load_path_bidaf_fusion):
progress_bar.set_postfix(NLL=nll_meter_bidaf_fu.avg)
elif (args.load_path_qanet_old):
progress_bar.set_postfix(NLL=nll_meter_qanet_old.avg)
elif (args.load_path_qanet_inde):
progress_bar.set_postfix(NLL=nll_meter_qanet_inde.avg)
elif (args.load_path_qanet_s_e):
progress_bar.set_postfix(NLL=nll_meter_qanet_s_e.avg)
else:
progress_bar.set_postfix(NLL=nll_meter_baseline.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
if (args.save_probabilities):
if (args.load_path_baseline):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_baseline_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_baseline_end, fp)
if (args.load_path_bidaf):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_bidaf_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_bidaf_end, fp)
if (args.load_path_bidaf_fusion):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_bidaf_fu_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_bidaf_fu_end, fp)
if (args.load_path_qanet):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_end, fp)
if (args.load_path_qanet_old):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_old_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_old_end, fp)
if (args.load_path_qanet_inde):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_inde_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_inde_end, fp)
if (args.load_path_qanet_s_e):
with open(args.save_dir + "/probs_start", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_s_e_start, fp)
with open(args.save_dir + "/probs_end", "wb") as fp: #Pickling
pickle.dump(save_prob_qanet_s_e_end, fp)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
if (args.load_path_qanet):
meter_avg = nll_meter_qanet.avg
elif (args.load_path_bidaf):
meter_avg = nll_meter_bidaf.avg
elif (args.load_path_bidaf_fusion):
meter_avg = nll_meter_bidaf_fu.avg
elif (args.load_path_qanet_inde):
meter_avg = nll_meter_qanet_inde.avg
elif (args.load_path_qanet_s_e):
meter_avg = nll_meter_qanet_s_e.avg
elif (args.load_path_qanet_old):
meter_avg = nll_meter_qanet_old.avg
else:
meter_avg = nll_meter_baseline.avg
results_list = [('NLL', meter_avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
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):
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(args):
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
#log = util.get_logger(args.save_dir, args.name)
#log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
print('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
print('Building model...')
model = BiDAF(word_vectors=word_vectors, hidden_size=args.hidden_size)
model = nn.DataParallel(model, gpu_ids)
print(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
print('Building dataset...')
#record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD("./data/my_test.npz", args.use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
print(f'Evaluating on {args.split} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
#eval_file = vars(args)[f'{args.split}_eval_file']
with open("./data/my_test_eval.json", 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_loader:
print("viewing the dataset")
print(cw_idxs, cc_idxs, qw_idxs, qc_idxs)
# 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, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
#if args.split != 'test':
# No labels for the test set, so NLL would be invalid
#progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
print("my evaluation ....")
for el in pred_dict:
print(el, pred_dict[el])
for el in sub_dict:
print(el, sub_dict[el])
def main(args):
# Set up logging
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
# Get model
log.info('Building model...')
if args.model == 'bidaf':
model = BiDAF(word_vectors=word_vectors, hidden_size=args.hidden_size)
elif args.model == 'bidafextra':
model = BiDAFExtra(word_vectors=word_vectors, args=args)
elif args.model == 'fusionnet':
model = FusionNet(word_vectors=word_vectors, args=args)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
log.info('Building dataset...')
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# print("*"*80)
# print(len(dataset.question_idxs))
# for question_idx in dataset.question_idxs:
# print(question_idx)
# print("*" * 80)
# print(self.question_idxs[question_idx])
# self.question_idxs[idx]
# print("data_loader: ",data_loader)
# Evaluate
log.info(f'Evaluating on {args.split} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
# create statistics
# print("*"*80)
# print(len(gold_dict))
# print(gold_dict['1']['question'])
count_questions_type = defaultdict(lambda: 0)
audit_trail_from_question_type = defaultdict(lambda: [])
list_of_interrogative_pronouns = [
"what", "whose", "why", "which", "where", "when", "how", "who", "whom"
]
for index in range(1, len(gold_dict)):
# transform the question in lower case to simplify the analysis, thus losing the benefit of the capital letters
# possibly indicating the position of the interrogative pronoun in the sentence.
question_lower_case = gold_dict[str(index)]['question'].lower()
list_question_lower_case_with_punctuation = question_lower_case.translate(
{ord(i): " "
for i in "'"}).split()
#
question_lower_case = []
for item in list_question_lower_case_with_punctuation:
question_lower_case.append(
item.translate({ord(i): ""
for i in ",.<>!@£$%^&*()_-+=?"}))
# defining a variable for the first word
first_word_question_lower_case = question_lower_case[0]
# defining variable for the second word
second_word_question_lower_case = question_lower_case[1]
# defining variable for the first and second word
combined_first_and_second_words = first_word_question_lower_case + " " + second_word_question_lower_case
#printing on the screen test for debugging purpose
# Analyzing the sentence
if first_word_question_lower_case in list_of_interrogative_pronouns:
count_questions_type[first_word_question_lower_case] += 1
audit_trail_from_question_type[
first_word_question_lower_case].append(str(index))
# composed question starting by in
elif first_word_question_lower_case == "in":
if second_word_question_lower_case in list_of_interrogative_pronouns and second_word_question_lower_case != "whose":
count_questions_type[combined_first_and_second_words] += 1
audit_trail_from_question_type[
combined_first_and_second_words].append(str(index))
else:
pronoun = find_first_interrogative_pronoun(
list_of_interrogative_pronouns, question_lower_case)
count_questions_type[pronoun] += 1
audit_trail_from_question_type[pronoun].append(str(index))
# composed question starting by by
elif first_word_question_lower_case == "by":
if second_word_question_lower_case in list_of_interrogative_pronouns \
and second_word_question_lower_case !="whom"\
and second_word_question_lower_case !="which"\
and second_word_question_lower_case !="when"\
and second_word_question_lower_case !="how":
count_questions_type[combined_first_and_second_words] += 1
audit_trail_from_question_type[
combined_first_and_second_words].append(str(index))
else:
pronoun = find_first_interrogative_pronoun(
list_of_interrogative_pronouns, question_lower_case)
count_questions_type[pronoun] += 1
audit_trail_from_question_type[pronoun].append(str(index))
else:
pronoun = find_first_interrogative_pronoun(
list_of_interrogative_pronouns, question_lower_case)
#if pronoun =="":
# print(">>", question_lower_case)
# print("@@@", gold_dict[str(index)]['question'])
count_questions_type[pronoun] += 1
audit_trail_from_question_type[pronoun].append(str(index))
# if pronoun =="":
# print(">>", question_lower_case.split())
# print()
#if first_word_question_lower_case == "if":
# print(">>", question_lower_case.split())
# print(count_questions_type)
# if gold_dict[str(index)]['question'].lower().split()[0] == "in":
# print(gold_dict[str(index)]['question'])
reverse_dict_by_value = OrderedDict(
sorted(count_questions_type.items(), key=lambda x: x[1]))
# print(count_questions_type)
total_questions = sum(count_questions_type.values())
# print(reverse_dict)
#for k, v in reverse_dict_by_value.items():
# print( "%s: %s and in percentage: %s" % (k, v, 100*v/total_questions))
#print(audit_trail_from_question_type)
# exit()
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, cw_pos, cw_ner, cw_freq, cqw_extra, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
# Forward
if args.model == 'bidaf':
log_p1, log_p2 = model(cw_idxs, qw_idxs)
else:
log_p1, log_p2 = model(cw_idxs, qw_idxs, cw_pos, cw_ner,
cw_freq, cqw_extra)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
# Printing information for questions without interrogative pronouns
""""
print("len(gold_dict): ", len(gold_dict))
print("len(pred_dict): ", len(pred_dict))
print("Is gold_dict.keys() identical to pred_dict.keys(): ", gold_dict.keys()==pred_dict.keys())
if gold_dict.keys()!=pred_dict.keys():
for key in gold_dict.keys():
if key not in pred_dict.keys():
print("key ", key, " missing in pred_dict.keys(")
"""
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Computing the F1 score for each type of question
#
# audit_trail_from_question_type[pronoun].append(str(index))
# create a list of the types of questions by extracting the keys from the dict audit_trail_from_question_type
types_of_questions = list(audit_trail_from_question_type.keys())
gold_dict_per_type_of_questions = defaultdict(lambda: [])
pred_dict_per_type_of_questions = {}
gold_dict_per_type_of_questions_start = {}
pred_dict_per_type_of_questions_start = {}
gold_dict_per_type_of_questions_middle = {}
pred_dict_per_type_of_questions_middle = {}
gold_dict_per_type_of_questions_end = {}
pred_dict_per_type_of_questions_end = {}
for type_of_questions in types_of_questions:
#gold_pred = {key: value for key, value in gold_dict.items() if key in audit_trail_from_question_type[type_of_questions]}
#lst_pred = {key: value for key, value in pred_dict.items() if key in audit_trail_from_question_type[type_of_questions]}
# Create two dictionnaries for each type of sentence for gold_dict_per_type_of_questions and pred_dict_per_type_of_questions
gold_dict_per_type_of_questions[type_of_questions] = {
key: value
for key, value in gold_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
pred_dict_per_type_of_questions[type_of_questions] = {
key: value
for key, value in pred_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
# print(type_of_questions," F1 score: ", util.eval_dicts(gold_dict_per_type_of_questions[type_of_questions], pred_dict_per_type_of_questions[type_of_questions], args.use_squad_v2)['F1'])
gold_dict_per_type_of_questions_start[type_of_questions] = {
key: value
for key, value in gold_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
pred_dict_per_type_of_questions_start[type_of_questions] = {
key: value
for key, value in pred_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
gold_dict_per_type_of_questions_middle[type_of_questions] = {
key: value
for key, value in gold_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
pred_dict_per_type_of_questions_middle[type_of_questions] = {
key: value
for key, value in pred_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
gold_dict_per_type_of_questions_end[type_of_questions] = {
key: value
for key, value in gold_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
pred_dict_per_type_of_questions_end[type_of_questions] = {
key: value
for key, value in pred_dict.items()
if key in audit_trail_from_question_type[type_of_questions]
and key in pred_dict.keys()
}
for key, value in gold_dict.items():
#if key in audit_trail_from_question_type[type_of_questions] and key in pred_dict.keys():
if key in audit_trail_from_question_type[
type_of_questions] and type_of_questions != "" and key in pred_dict_per_type_of_questions[
type_of_questions]:
"""
print("type_of_questions: ",type_of_questions)
print("key: ", key)
print("question: ", value["question"])
sub_index = value["question"].lower().find(type_of_questions)
print("sub_index: ",sub_index)
test_fc = value["question"].lower().find(type_of_questions)
print("present type of the var: ",type(test_fc))
#print("question: ", value["question"][str(key)])
print("length of the question: ", len(value["question"]))
print('Position of the interrogative pronoun in the question:', )
"""
# Create two dictionnaries for each type of sentence based at the start of the sentence
if value["question"].lower().find(
type_of_questions) == 1 or value["question"].lower(
).find(type_of_questions) == 0:
#print("BEGINNING")
if type_of_questions != "":
try:
del gold_dict_per_type_of_questions_middle[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_middle[
type_of_questions][key]
except KeyError:
pass
try:
del gold_dict_per_type_of_questions_end[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_end[
type_of_questions][key]
except KeyError:
pass
#pred_dict_per_type_of_questions_start[type_of_questions] = {key: pred_dict[key] for key in
# gold_dict_per_type_of_questions_start[
# type_of_questions].keys()}
elif value["question"].lower(
).find(type_of_questions) >= len(
value["question"]) - len(type_of_questions) - 5:
#print("END")
if type_of_questions != "":
try:
del gold_dict_per_type_of_questions_middle[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_middle[
type_of_questions][key]
except KeyError:
pass
try:
del gold_dict_per_type_of_questions_start[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_start[
type_of_questions][key]
except KeyError:
pass
#print("type_of_questions: ",type_of_questions)
#sub_index = value["question"].lower().find(type_of_questions)
#print("sub_index: ", sub_index)
#print("len(value['question']) - len(type_of_questions) - 2: ", len(value["question"])-len(type_of_questions)-2)
#start_string = len(value["question"])-len(type_of_questions)-6
#end_string = len(value["question"])-1
#print("extract at the end: ", value["question"][start_string:end_string])
else:
#print("MIDDLE")
if type_of_questions != "":
try:
del gold_dict_per_type_of_questions_start[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_start[
type_of_questions][key]
except KeyError:
pass
try:
del gold_dict_per_type_of_questions_end[
type_of_questions][key]
except KeyError:
pass
try:
del pred_dict_per_type_of_questions_end[
type_of_questions][key]
except KeyError:
pass
pass
"""
if type_of_questions != "":
gold_dict_per_type_of_questions_start[type_of_questions] = {key: value for key, value in gold_dict.items() if (key in audit_trail_from_question_type[type_of_questions] \
and (value["question"].lower().find(type_of_questions) <= 1) \
and key in pred_dict_per_type_of_questions[type_of_questions]) }
"""
"""
for key in gold_dict_per_type_of_questions_start[type_of_questions].keys():
print("key:: ", key )
print("type(key):: ", type(key) )
print("pred_dict[,key,] : ", pred_dict[key])
print("@@@@@@@@@@@@@@@@@@@@@@@@")
pred_dict_per_type_of_questions_start[type_of_questions] = {key: pred_dict[key] for key in gold_dict_per_type_of_questions_start[type_of_questions].keys()}
#pred_dict_per_type_of_questions_start[type_of_questions] = {key: value for key, value in pred_dict.items() if key in list(gold_dict_per_type_of_questions_start[type_of_questions].keys()) }
# Create two dictionnaries for each type of sentence based at the end of the sentence
gold_dict_per_type_of_questions_end[type_of_questions] = {key: value for key, value in gold_dict.items() if key in audit_trail_from_question_type[type_of_questions] \
and value["question"].lower().find(type_of_questions) >= len(value["question"])-len(type_of_questions)-2 \
and key in pred_dict_per_type_of_questions[type_of_questions]}
pred_dict_per_type_of_questions_end[type_of_questions] = {key: pred_dict[key] for key in list(gold_dict_per_type_of_questions_end[type_of_questions].keys())}
#print("*"*80)
# Create two dictionnaries for each type of sentence based at the middle of the sentencecount_questions_type
gold_dict_per_type_of_questions_middle[type_of_questions] = {key: value for key, value in gold_dict.items() if key not in list(gold_dict_per_type_of_questions_start[type_of_questions].keys()) \
and key not in list(gold_dict_per_type_of_questions_end[type_of_questions].keys())}
pred_dict_per_type_of_questions_middle[type_of_questions] = {key: pred_dict[key] for key in list(gold_dict_per_type_of_questions_end[type_of_questions].keys())}
else:
gold_dict_per_type_of_questions_start[""] = gold_dict_per_type_of_questions[""]
pred_dict_per_type_of_questions_start[""] = pred_dict_per_type_of_questions[""]
gold_dict_per_type_of_questions_end[""] = gold_dict_per_type_of_questions[""]
pred_dict_per_type_of_questions_end[""] = pred_dict_per_type_of_questions[""]
gold_dict_per_type_of_questions_middle[""] = gold_dict_per_type_of_questions[""]
pred_dict_per_type_of_questions_middle[""] = pred_dict_per_type_of_questions[""]
"""
positions_in_question = ["beginning", "middle", "end"]
# print(type_of_questions," F1 score: ", util.eval_dicts(gold_dict_per_type_of_questions[type_of_questions], pred_dict_per_type_of_questions[type_of_questions], args.use_squad_v2)['F1'])
list_beginning = [
util.eval_dicts(
gold_dict_per_type_of_questions_start[type_of_questions],
pred_dict_per_type_of_questions_start[type_of_questions],
args.use_squad_v2)['F1']
for type_of_questions in types_of_questions
]
list_middle = [
util.eval_dicts(
gold_dict_per_type_of_questions_middle[type_of_questions],
pred_dict_per_type_of_questions_middle[type_of_questions],
args.use_squad_v2)['F1']
for type_of_questions in types_of_questions
]
list_end = [
util.eval_dicts(
gold_dict_per_type_of_questions_end[type_of_questions],
pred_dict_per_type_of_questions_end[type_of_questions],
args.use_squad_v2)['F1']
for type_of_questions in types_of_questions
]
#for type_of_questions in types_of_questions:
# print("gold_dict_per_type_of_questions_start[type_of_questions]: ",gold_dict_per_type_of_questions_start[type_of_questions])
# print("pred_dict_per_type_of_questions[type_of_questions]: ",pred_dict_per_type_of_questions[type_of_questions])
F1 = np.array([list_beginning, list_middle, list_end])
m, n = F1.shape
value_to_ignore = []
for i in range(m):
for j in range(n):
if F1[i, j] == "NA" or F1[i, j] == 0:
value_to_ignore.append((i, j))
print("value to ignore: ", value_to_ignore)
#F1 = np.array([[0, 0, 0, 0, 0.1, 0, 0, 0, 0, 0, 0, 0, 0],
# [0, 0, 0, 0, 0.1, 0, 0, 0, 0, 0, 0, 0, 0],
# [0, 0, 0, 0, 0.1, 0, 0, 0, 0, 0, 0, 0, 0]])
data_label = copy.deepcopy(F1)
for row in data_label:
for column_idx in range(len(row)):
if row[column_idx] == "NA":
row[column_idx] = ""
# print question without interrogative pronoun required for the second part of the analysis:
for key, value in gold_dict.items():
if key in audit_trail_from_question_type[
""] and key in pred_dict.keys():
print("question: ", gold_dict_per_type_of_questions[''])
print("golden answers: ", )
print("prediction: ", pred_dict[key])
print()
fig, ax = plt.subplots()
types_of_questions[types_of_questions.index(
"")] = "Implicit question without interrogative pronoun"
im, cbar = heatmap(F1, positions_in_question, types_of_questions, ax=ax, \
cmap="YlGn", cbarlabel="F1 scores")
texts = annotate_heatmap(im,
data=data_label,
valfmt="{x:.1f}",
ignore=value_to_ignore)
fig.tight_layout()
plt.show()
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
def test_model(questions,
context,
use_squad_v2=True,
model_path="../save/training-02/best.pth.tar"):
# Set up logging
#args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
#log = util.get_logger(args.save_dir, args.name)
#log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
#args = get_test_args()
device, gpu_ids = util.get_available_devices()
batch_size = 64 * max(1, len(gpu_ids))
# Get embeddings
#print('Loading embeddings...')
word_vectors = util.torch_from_json('../data/word_emb.json')
# Get model
#print('Building model...')
model = BiDAF(word_vectors=word_vectors, hidden_size=100)
model = nn.DataParallel(model, gpu_ids)
#model_path = "../save/training-02/best.pth.tar"
#print(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model, model_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
#print('Building dataset...')
#record_file = vars(args)[f'{args.split}_record_file']
# my code start here
# this is a simple approch when dealing with the user date
# according to your approch of creating the interface you can change this code
# and also you have to check the function "process_file" in the setup.py file
processed_questions = []
for index, question in enumerate(questions):
processed_question = {
"question": question,
"id": index,
"answers": [{
"answer_start": 0,
"text": "never mind"
}]
}
processed_questions.append(processed_question)
source = {"paragraphs": [{"qas": processed_questions, "context": context}]}
word_counter, char_counter = Counter(), Counter()
with open("../data/word2idx.json", "r") as f1:
word2idx_dict = json.load(f1)
with open("../data/char2idx.json", "r") as f2:
char2idx_dict = json.load(f2)
my_test_examples, my_test_eval = process_file(source, "my_test",
word_counter, char_counter)
npz = build_features(my_test_examples,
"my_test",
word2idx_dict,
char2idx_dict,
is_test=True)
#my code end here
dataset = SQuAD(npz, use_squad_v2)
data_loader = data.DataLoader(dataset,
batch_size=batch_size,
shuffle=False,
num_workers=4,
collate_fn=collate_fn)
# Evaluate
#print(f'Evaluating on {args.split} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
#eval_file = vars(args)[f'{args.split}_eval_file']
gold_dict = my_test_eval
#print("gold_dict", gold_dict)
#print("data_loader", data_loader)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_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, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, 15, use_squad_v2)
print("starts ", starts, " ends ", ends)
# Log info
progress_bar.update(batch_size)
#if args.split != 'test':
# No labels for the test set, so NLL would be invalid
#progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
#print("my evaluation ....")
#for el in pred_dict:
#print(el, pred_dict[el])
#for el in sub_dict:
#print(el, sub_dict[el])
return pred_dict
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 (unchanged from train.py)
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) # train only, not in test
device, args.gpu_ids = util.get_available_devices() # todo(small): should this be args (compare test_para)
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
args.batch_size *= max(1, len(args.gpu_ids)) # args.py: default size is 64
# Set random seed (unchanged) - train only
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)
# Prepare BiDAF model (must already trained)
log.info('Building BiDAF model (should be pretrained)')
bidaf_model = BiDAF(word_vectors=word_vectors, # todo: these word vectors shouldn't matter?
hidden_size=args.hidden_size) # since they will be loaded in during load_model?
#drop_prob=args.drop_prob) # no drop probability since we are not training
bidaf_model = nn.DataParallel(bidaf_model, args.gpu_ids)
if args.short_test:
args.hidden_size = 5
elif not args.load_path:
log.info("Trying to trian paraphraser withou bidaf model. "
"First train BiDAF and then specify the load path. Exiting")
exit(1)
else:
log.info(f'Loading checkpoint from {args.load_path}...')
bidaf_model = util.load_model(bidaf_model, args.load_path, args.gpu_ids, return_step=False) # don't need step since we aren't training
bidaf_model = bidaf_model.to(device)
bidaf_model.eval() # we eval only (vs train)
# todo: Setup the Paraphraser model
paraphaser_model = Paraphraser(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
# Get data loader
log.info('Building dataset...')
# New for paraphrase: squad_paraphrase has extra fields
train_dataset = SQuAD_paraphrase(args.train_record_file, args.use_squad_v2) # train.npz (from setup.py, build_features())
train_loader = data.DataLoader(train_dataset, # this dataloader used for all epoch iteration
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
collate_fn=collate_fn_para)
dev_dataset = SQuAD_paraphrase(args.dev_record_file, args.use_squad_v2) # dev.npz (same as above)
dev_loader = data.DataLoader(dev_dataset, # dev.npz used in evaluate() fcn
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn_para)
# todo: this is just for looking at the paraphrases
idx2word_dict = load(args.idx2word_file)
#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
# ema = util.EMA(paraphaser_model, args.ema_decay)
# optimizer = optim.Adadelta(paraphaser_model.parameters(), args.lr,
# weight_decay=args.l2_wd)
# scheduler = sched.LambdaLR(optimizer, lambda s: 1.) # Constant LR
# Train
step = 0
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, cphr_idxs, qphr_idxs, qphr_types, ids in train_loader:
# Setup for forward
# note that cc_idxs, qc_idxs are not used! (character indices)
cw_idxs = cw_idxs.to(device) # todo what does this actually do
qw_idxs = qw_idxs.to(device)
cphr_idxs = cphr_idxs.to(device)
qphr_idxs = qphr_idxs.to(device)
qphr_types = qphr_types.to(device)
batch_size = cw_idxs.size(0)
# if args.short_test:
# print(f'batch size: {batch_size}')
# for i, type in enumerate(cphr_idxs[0]):
# print(f'type: {i}')
# pp(type)
# for x in (qphr_idxs[0], qphr_types[0]):
# pp(x)
# return
paraphrased = paraphaser_model(qphr_idxs, qphr_types, cphr_idxs)
for idx, p in enumerate(paraphrased): # enumerate over batch_size
non_zeros = p[p.nonzero()].squeeze()
#paraphrased[idx] = non_zeros
sentence_as_list = [idx2word_dict[str(w.item())] for w in non_zeros]
pp(" ".join(sentence_as_list))
#pp([idx2word_dict[w] for w in non_zeros])
if args.short_test:
return
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) # // is floor division
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, # call eval with dev_loader
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)
# ######################################
# 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 test(args):
# Set up logging
log = util.get_logger(args.save_dir, args.name)
log.info(f"Args: {dumps(vars(args), indent=4, sort_keys=True)}")
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
# 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,
use_glove=args.use_glove,
)
model = nn.DataParallel(model, gpu_ids)
log.info(f"Loading checkpoint from {args.load_path}...")
model = util.load_model(model, args.load_path, gpu_ids, return_step=False)
model = model.to(device)
model.eval()
# Get data loader
log.info("Building dataset...")
record_file = vars(args)[f"{args.split}_record_file"]
dataset = SQuAD(record_file, args.use_squad_v2)
data_loader = data.DataLoader(
dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn,
)
# Evaluate
log.info(f"Evaluating on {args.split} split...")
nll_meter = stats.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f"{args.split}_eval_file"]
with open(eval_file, "r") as fh:
gold_dict = json_load(fh)
with torch.no_grad(), tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in data_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, qw_idxs)
y1, y2 = y1.to(device), y2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Get F1 and EM scores
p1, p2 = log_p1.exp(), log_p2.exp()
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != "test":
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(
gold_dict,
ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2,
)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != "test":
results = {"NLL": nll_meter.avg}
results.update(eval.eval_dicts(gold_dict, pred_dict,
args.use_squad_v2))
# Log to console
results_str = ", ".join(f"{k}: {v:05.2f}" for k, v in results.items())
log.info(f"{args.split.title()} {results_str}")
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(
tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals,
)
# Write submission file
if args.split == "dev":
sub_path = join(args.save_dir, "val" + "_" + args.sub_file)
else:
sub_path = join(args.save_dir, args.split + "_" + args.sub_file)
log.info(f"Writing submission file to {sub_path}...")
eval.write_submission(sub_path, sub_dict)
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 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):
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)
# 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
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=False)
log = util.get_logger(args.save_dir, args.name)
device, gpu_ids = util.get_available_devices()
args.batch_size *= max(1, len(gpu_ids))
log.info(f'Args: {dumps(vars(args), indent=4, sort_keys=True)}')
# Get embeddings
log.info('Loading embeddings...')
word_vectors = util.torch_from_json(args.word_emb_file)
models = {}
if args.use_ensemble:
total_models = 0
for model_name in ['bidaf', 'bidafextra', 'fusionnet']:
models_list = []
for model_file in glob.glob(
f'{args.load_path}/{model_name}-*/{args.ensemble_models}'):
# Get model
log.info('Building model...')
if model_name == 'bidaf':
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size)
elif model_name == 'bidafextra':
model = BiDAFExtra(word_vectors=word_vectors, args=args)
elif model_name == 'fusionnet':
model = FusionNet(word_vectors=word_vectors, args=args)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {model_file}...')
model = util.load_model(model,
model_file,
gpu_ids,
return_step=False)
# Load each model on CPU (have plenty of RAM ...)
model = model.cpu()
model.eval()
models_list.append(model)
models[model_name] = models_list
total_models += len(models_list)
log.info(f'Using an ensemble of {total_models} models')
else:
device, gpu_ids = util.get_available_devices()
# Get model
log.info('Building model...')
if args.model == 'bidaf':
model = BiDAF(word_vectors=word_vectors,
hidden_size=args.hidden_size)
elif args.model == 'bidafextra':
model = BiDAFExtra(word_vectors=word_vectors, args=args)
elif args.model == 'fusionnet':
model = FusionNet(word_vectors=word_vectors, args=args)
model = nn.DataParallel(model, gpu_ids)
log.info(f'Loading checkpoint from {args.load_path}...')
model = util.load_model(model,
args.load_path,
gpu_ids,
return_step=False)
model = model.to(device)
model.eval()
models[args.model] = [model]
# Get data loader
log.info('Building dataset...')
record_file = vars(args)[f'{args.split}_record_file']
dataset = SQuAD(record_file, args)
data_loader = data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_fn)
# Evaluate
log.info(f'Evaluating on {args.split} split...')
nll_meter = util.AverageMeter()
pred_dict = {} # Predictions for TensorBoard
sub_dict = {} # Predictions for submission
eval_file = vars(args)[f'{args.split}_eval_file']
with open(eval_file, 'r') as fh:
gold_dict = json_load(fh)
with torch.no_grad(), \
tqdm(total=len(dataset)) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, cw_pos, cw_ner, cw_freq, cqw_extra, y1, y2, ids in data_loader:
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
p1s = []
p2s = []
for model_name in models:
for model in models[model_name]:
# Move model to GPU to evaluate
model = model.to(device)
# Forward
if model_name == 'bidaf':
log_p1, log_p2 = model.to(device)(cw_idxs, qw_idxs)
else:
log_p1, log_p2 = model.to(device)(cw_idxs, qw_idxs,
cw_pos, cw_ner,
cw_freq, cqw_extra)
log_p1, log_p2 = log_p1.cpu(), log_p2.cpu()
if not args.use_ensemble:
y1, y2 = y1.to(device), y2.to(device)
log_p1, log_p2 = log_p1.to(device), log_p2.to(device)
loss = F.nll_loss(log_p1, y1) + F.nll_loss(log_p2, y2)
nll_meter.update(loss.item(), batch_size)
# Move model back to CPU to release GPU memory
model = model.cpu()
# Get F1 and EM scores
p1, p2 = log_p1.exp().unsqueeze(
-1).cpu(), log_p2.exp().unsqueeze(-1).cpu()
p1s.append(p1), p2s.append(p2)
best_ps = torch.max(
torch.cat([
torch.cat(p1s, -1).unsqueeze(-1),
torch.cat(p2s, -1).unsqueeze(-1)
], -1), -2)[0]
p1, p2 = best_ps[:, :, 0], best_ps[:, :, 1]
starts, ends = util.discretize(p1, p2, args.max_ans_len,
args.use_squad_v2)
# Log info
progress_bar.update(batch_size)
if args.split != 'test':
# No labels for the test set, so NLL would be invalid
progress_bar.set_postfix(NLL=nll_meter.avg)
idx2pred, uuid2pred = util.convert_tokens(gold_dict, ids.tolist(),
starts.tolist(),
ends.tolist(),
args.use_squad_v2)
pred_dict.update(idx2pred)
sub_dict.update(uuid2pred)
# Log results (except for test set, since it does not come with labels)
if args.split != 'test':
results = util.eval_dicts(gold_dict, pred_dict, args.use_squad_v2)
results_list = [('NLL', nll_meter.avg), ('F1', results['F1']),
('EM', results['EM'])]
if args.use_squad_v2:
results_list.append(('AvNA', results['AvNA']))
results = OrderedDict(results_list)
# Log to console
results_str = ', '.join(f'{k}: {v:05.2f}' for k, v in results.items())
log.info(f'{args.split.title()} {results_str}')
# Log to TensorBoard
tbx = SummaryWriter(args.save_dir)
util.visualize(tbx,
pred_dict=pred_dict,
eval_path=eval_file,
step=0,
split=args.split,
num_visuals=args.num_visuals)
# Write submission file
sub_path = join(args.save_dir, args.split + '_' + args.sub_file)
log.info(f'Writing submission file to {sub_path}...')
with open(sub_path, 'w', newline='', encoding='utf-8') as csv_fh:
csv_writer = csv.writer(csv_fh, delimiter=',')
csv_writer.writerow(['Id', 'Predicted'])
for uuid in sorted(sub_dict):
csv_writer.writerow([uuid, sub_dict[uuid]])
def main(args):
# Set up logging and devices
startime = datetime.now()
args.save_dir = util.get_save_dir(args.save_dir, args.name, training=True)
log = util.get_logger(args.save_dir, args.name)
time_log = args.log_time
if time_log > 0:
log.info(f'Start training at: {startime.strftime("%H:%M:%S")}')
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))
model_type = args.model
# 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)
# check this
#useCharEmbeddings = args.model == 'BiDAFplus'
# Get embeddings
log.info('Loading embeddings...')
print(f'{args.word_emb_file}')
word_vectors = util.torch_from_json(args.word_emb_file)
char_vectors = util.torch_from_json(args.char_emb_file)
if time_log > 0:
log.info(f'Loaded embeddings: {(datetime.now()-startime).seconds}')
# load_char_vectors
# Get model
log.info('Building model...')
if model_type == 'BiDAFplus': #
model = BiDAFplus(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=args.hidden_size,
params=get_params(model_type, args.params))
elif model_type == 'BiDAFbase':
model = BiDAFbase(word_vectors=word_vectors,
hidden_size=args.hidden_size,
drop_prob=args.drop_prob)
elif model_type == "Transformer":
model = TransformerModel(word_vectors=word_vectors,
char_vectors=char_vectors,
params=get_params(model_type, args.params))
elif model_type == 'BiDAF':
model = BiDAF(word_vectors=word_vectors,
char_vectors=char_vectors,
hidden_size=args.hidden_size,
params=get_params(model_type, args.params))
model = nn.DataParallel(model, args.gpu_ids)
if time_log > 0:
log.info(f'Built model: {(datetime.now()-startime).seconds}')
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...')
if args.mode != 'quick_eval':
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)
else:
loaded_data = quick_eval_data_loader()
train_loader = [loaded_data for _ in range(5)]
dev_loader = [quick_eval_data_loader(dev=True)]
train_dataset = train_loader
dev_dataset = dev_loader
log.info('Built dataset: {}:{}'.format(*divmod((datetime.now() -
startime).seconds, 60)))
# Train
log.info('Training...')
steps_till_eval = args.eval_steps
epoch = step // len(train_dataset)
if time_log > 0:
traintime = datetime.now()
total_iterations = 0
while epoch != args.num_epochs:
epoch += 1
log.info(f'Starting epoch {epoch}...')
if time_log > 0:
epochtime = datetime.now()
if args.mode != 'quick_eval':
progress_len = len(train_loader.dataset)
else:
progress_len = len(train_loader)
with torch.enable_grad(), \
tqdm(total=progress_len) as progress_bar:
for cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids in train_loader:
#quick_eval_data_saver(cw_idxs, cc_idxs, qw_idxs, qc_idxs, y1, y2, ids)
#########
if time_log > 0:
itertime = datetime.now()
# Setup for forward
cw_idxs = cw_idxs.to(device)
qw_idxs = qw_idxs.to(device)
batch_size = cw_idxs.size(0)
optimizer.zero_grad()
if model_type == 'BiDAF' or model_type == "Transformer":
cc_idxs = cc_idxs.to(device)
qc_idxs = qc_idxs.to(device)
log_p1, log_p2 = model(cc_idxs, qc_idxs, cw_idxs, qw_idxs)
# Forward
elif model_type == 'BiDAFbase':
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()
if time_log > 2:
forwardtime = datetime.now()
log.info('Forward time {}:{}'.format(
*divmod((forwardtime - itertime).seconds, 60)))
# 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)
if time_log > 2:
backwardtime = datetime.now()
log.info('Backward time {}:{}'.format(
*divmod((backwardtime - forwardtime).seconds, 60)))
# 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)
if time_log > 0:
enditertime = datetime.now()
#log.info('Iteration {} {}:{}'.format(total_iterations,
# *divmod((enditertime-itertime).seconds, 60)))
steps_till_eval -= batch_size
if steps_till_eval <= 0 or args.mode == 'quick_eval':
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,
model_type,
quick_eval=args.mode == 'quick_eval')
saver.save(step, model, results[args.metric_name], device)
ema.resume(model)
# Log to console
if time_log > 1:
log.info('Eval time {}:{}'.format(
*divmod((datetime.now() -
enditertime).seconds, 60)))
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)
total_iterations += 1
if ((time_log == 2) and (total_iterations % 10 == 0)) or (
(time_log == 1) and (total_iterations % 100 == 0)):
log.info('Mean iteration time {}:{}'.format(
*divmod((enditertime - traintime).seconds /
total_iterations, 60)))
if time_log > 0:
endepochtime = datetime.now()
log.info('Epoch time {}:{}'.format(
*divmod((endepochtime - epochtime).seconds, 60)))