def validate_unofficial(args, data_loader, model, global_stats, mode):
"""Run one full unofficial validation.
Unofficial = doesn't use SQuAD script.
"""
eval_time = utils.Timer()
start_acc = utils.AverageMeter()
end_acc = utils.AverageMeter()
exact_match = utils.AverageMeter()
loss = utils.AverageMeter()
# Make predictions
examples = 0
for ex in data_loader:
batch_size = ex[0].size(0)
(pred_s, pred_e, _), batch_loss = model.predict(ex)
target_s, target_e = ex[-3:-1]
# We get metrics for independent start/end and joint start/end
accuracies = eval_accuracies(pred_s, target_s, pred_e, target_e)
start_acc.update(accuracies[0], batch_size)
end_acc.update(accuracies[1], batch_size)
exact_match.update(accuracies[2], batch_size)
loss.update(batch_loss, batch_size)
# If getting train accuracies, sample max 10k
examples += batch_size
if mode == 'train' and examples >= 1e4:
break
logger.info('%s valid unofficial: Epoch = %d | start = %.2f | ' %
(mode, global_stats['epoch'], start_acc.avg) +
'end = %.2f | exact = %.2f | examples = %d | ' %
(end_acc.avg, exact_match.avg, examples) +
'valid time = %.2f (s)' % eval_time.time())
return {'loss': loss.avg, 'exact_match': exact_match.avg}
def validate_official(args, data_loader, model, global_stats, offsets, texts,
answers):
"""Run one full official validation. Uses exact spans and same
exact match/F1 score computation as in the SQuAD script.
Extra arguments:
offsets: The character start/end indices for the tokens in each context.
texts: Map of qid --> raw text of examples context (matches offsets).
answers: Map of qid --> list of accepted answers.
"""
eval_time = utils.Timer()
f1 = utils.AverageMeter()
exact_match = utils.AverageMeter()
# Run through examples
examples = 0
for ex in data_loader:
ex_id, batch_size = ex[-1], ex[0].size(0)
pred_s, pred_e, _ = model.predict(ex)
for i in range(batch_size):
s_offset = offsets[ex_id[i]][pred_s[i][0]][0]
e_offset = offsets[ex_id[i]][pred_e[i][0]][1]
prediction = texts[ex_id[i]][s_offset:e_offset]
# Compute metrics
ground_truths = answers[ex_id[i]]
exact_match.update(
utils.metric_max_over_ground_truths(utils.exact_match_score,
prediction, ground_truths))
f1.update(
utils.metric_max_over_ground_truths(utils.f1_score, prediction,
ground_truths))
examples += batch_size
logger.info('dev valid official: Epoch = %d | EM = %.2f | ' %
(global_stats['epoch'], exact_match.avg * 100) +
'F1 = %.2f | examples = %d | valid time = %.2f (s)' %
(f1.avg * 100, examples, eval_time.time()))
return {'exact_match': exact_match.avg * 100, 'f1': f1.avg * 100}
def train(args, data_loader, model, global_stats):
"""Run through one epoch of model training with the provided data loader."""
# Initialize meters + timers
train_loss = utils.AverageMeter()
epoch_time = utils.Timer()
# Run one epoch
for idx, ex in enumerate(data_loader):
train_loss.update(*model.update(ex))
if idx % args.display_iter == 0:
logger.info('train: Epoch = %d | iter = %d/%d | ' %
(global_stats['epoch'], idx, len(data_loader)) +
'loss = %.2f | elapsed time = %.2f (s)' %
(train_loss.avg, global_stats['timer'].time()))
train_loss.reset()
logger.info('train: Epoch %d done. Time for epoch = %.2f (s)' %
(global_stats['epoch'], epoch_time.time()))
# Checkpoint
if args.checkpoint:
model.checkpoint(args.model_file + '.checkpoint',
global_stats['epoch'] + 1)
def main(args):
# --------------------------------------------------------------------------
# DATA
logger.info('-' * 100)
logger.info('Load data files')
train_exs = utils.load_data(args, args.train_file, skip_no_answer=True)
logger.info('Num train examples = %d' % len(train_exs))
dev_exs = utils.load_data(args, args.dev_file)
logger.info('Num dev examples = %d' % len(dev_exs))
# If we are doing offician evals then we need to:
# 1) Load the original text to retrieve spans from offsets.
# 2) Load the (multiple) text answers for each question.
if args.official_eval:
dev_texts = utils.load_text(args.dev_json)
dev_offsets = {ex['id']: ex['offsets'] for ex in dev_exs}
dev_answers = utils.load_answers(args.dev_json)
# --------------------------------------------------------------------------
# MODEL
logger.info('-' * 100)
start_epoch = 0
if args.checkpoint and os.path.isfile(args.model_file + '.checkpoint'):
# Just resume training, no modifications.
logger.info('Found a checkpoint...')
checkpoint_file = args.model_file + '.checkpoint'
model, start_epoch = DocReader.load_checkpoint(checkpoint_file, args)
else:
# Training starts fresh. But the model state is either pretrained or
# newly (randomly) initialized.
if args.pretrained:
logger.info('Using pretrained model...')
model = DocReader.load(args.pretrained, args)
if args.expand_dictionary:
logger.info('Expanding dictionary for new data...')
# Add words in training + dev examples
words = utils.load_words(args, train_exs + dev_exs)
added = model.expand_dictionary(words)
# Load pretrained embeddings for added words
if args.embedding_file:
model.load_embeddings(added, args.embedding_file)
else:
logger.info('Training model from scratch...')
model = init_from_scratch(args, train_exs, dev_exs)
# Set up partial tuning of embeddings
if args.tune_partial > 0:
logger.info('-' * 100)
logger.info('Counting %d most frequent question words' %
args.tune_partial)
top_words = utils.top_question_words(
args, train_exs, model.word_dict
)
for word in top_words[:5]:
logger.info(word)
logger.info('...')
for word in top_words[-6:-1]:
logger.info(word)
model.tune_embeddings([w[0] for w in top_words])
# Set up optimizer
model.init_optimizer()
# Use the GPU?
if args.cuda:
model.cuda()
# Use multiple GPUs?
if args.parallel:
model.parallelize()
# --------------------------------------------------------------------------
# DATA ITERATORS
# Two datasets: train and dev. If we sort by length it's faster.
logger.info('-' * 100)
logger.info('Make data loaders')
train_dataset = data.ReaderDataset(train_exs, model, single_answer=True)
if args.sort_by_len:
train_sampler = data.SortedBatchSampler(train_dataset.lengths(),
args.batch_size,
shuffle=True)
else:
train_sampler = torch.utils.data.sampler.RandomSampler(train_dataset)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
sampler=train_sampler,
num_workers=args.data_workers,
collate_fn=vector.batchify,
pin_memory=args.cuda,
)
dev_dataset = data.ReaderDataset(dev_exs, model, single_answer=False)
if args.sort_by_len:
dev_sampler = data.SortedBatchSampler(dev_dataset.lengths(),
args.test_batch_size,
shuffle=False)
else:
dev_sampler = torch.utils.data.sampler.SequentialSampler(dev_dataset)
dev_loader = torch.utils.data.DataLoader(
dev_dataset,
batch_size=args.test_batch_size,
sampler=dev_sampler,
num_workers=args.data_workers,
collate_fn=vector.batchify,
pin_memory=args.cuda,
)
# -------------------------------------------------------------------------
# PRINT CONFIG
logger.info('-' * 100)
logger.info('CONFIG:\n%s' %
json.dumps(vars(args), indent=4, sort_keys=True))
# --------------------------------------------------------------------------
# TRAIN/VALID LOOP
logger.info('-' * 100)
logger.info('Starting training...')
stats = {'timer': utils.Timer(), 'epoch': 0, 'best_valid': 0}
writer = SummaryWriter('data/runs')
for epoch in range(start_epoch, args.num_epochs):
stats['epoch'] = epoch
# Train
train(args, train_loader, model, stats)
# Validate unofficial (train)
result = validate_unofficial(args, train_loader, model, stats, mode='train')
tr_loss = result['loss']
# Validate unofficial (dev)
result = validate_unofficial(args, dev_loader, model, stats, mode='dev')
dev_loss = result['loss']
dev_em, dev_f1 = 0, 0
# Validate official
if args.official_eval:
result = validate_official(args, dev_loader, model, stats,
dev_offsets, dev_texts, dev_answers)
dev_em, dev_f1 = result['exact_match'], result['f1']
writer.add_scalars('bidaf/%s_loss' % args.exp_id, {'tr_loss': torch.Tensor(1).fill_(tr_loss), 'dev_loss': torch.Tensor(1).fill_(dev_loss)}, epoch)
writer.add_scalars('bidaf/%s_metric' % args.exp_id, {'em': torch.Tensor(1).fill_(dev_em), 'f1': torch.Tensor(1).fill_(dev_f1)}, epoch)
# Save best valid
if result[args.valid_metric] > stats['best_valid']:
logger.info('Best valid: %s = %.2f (epoch %d, %d updates)' %
(args.valid_metric, result[args.valid_metric],
stats['epoch'], model.updates))
model.save(args.model_file)
stats['best_valid'] = result[args.valid_metric]