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()
support_acc = utils.AverageMeter()
# Make predictions
examples = 0
for ex in data_loader:
batch_size = ex[0].size(0)
pred_s, pred_e, _, pred_sp = model.predict(ex)
target_s, target_e, target_sp = ex[-5:-2]
# We get metrics for independent start/end and joint start/end
accuracies = eval_accuracies_spans(pred_s, target_s, pred_e, target_e)
support_accuracy = eval_accuracies_support(pred_sp, target_sp, mode)
start_acc.update(accuracies[0], batch_size)
end_acc.update(accuracies[1], batch_size)
exact_match.update(accuracies[2], batch_size)
support_acc.update(support_accuracy, 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) +
'sentence_selection = %.2f | ' % (support_acc.avg) +
'valid time = %.2f (s)' % eval_time.time())
return {'exact_match': exact_match.avg}
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 = []
for t_file in args.train_file:
train_exs += utils.load_data(args, t_file, skip_no_answer=True)
np.random.shuffle(train_exs)
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)
## OFFSET comes from the gold sentence; the predicted sentence value shoule be maintained and sent to official validation set
# --------------------------------------------------------------------------
# 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.
# Sentence selection objective : run the sentence selector as a submodule
logger.info('-' * 100)
logger.info('Make data loaders')
train_dataset = reader_data.ReaderDataset(train_exs, model, single_answer=True)
# Filter out None examples in training dataset (where sentence selection fails)
#train_dataset.examples = [t for t in train_dataset.examples if t is not None]
if args.sort_by_len:
train_sampler = reader_data.SortedBatchSampler(train_dataset.lengths(),
args.batch_size,
shuffle=True)
else:
train_sampler = torch.utils.data.sampler.RandomSampler(train_dataset)
if args.use_sentence_selector:
train_batcher = reader_vector.sentence_batchifier(model, single_answer=True)
# batching_function = train_batcher.batchify
batching_function = reader_vector.batchify
else:
batching_function = reader_vector.batchify
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
sampler=train_sampler,
num_workers=args.data_workers,
collate_fn=batching_function,
pin_memory=args.cuda,
)
dev_dataset = reader_data.ReaderDataset(dev_exs, model, single_answer=False)
#dev_dataset.examples = [t for t in dev_dataset.examples if t is not None]
if args.sort_by_len:
dev_sampler = reader_data.SortedBatchSampler(dev_dataset.lengths(),
args.test_batch_size,
shuffle=False)
else:
dev_sampler = torch.utils.data.sampler.SequentialSampler(dev_dataset)
if args.use_sentence_selector:
dev_batcher = reader_vector.sentence_batchifier(model, single_answer=False)
# batching_function = dev_batcher.batchify
batching_function = reader_vector.batchify
else:
batching_function = reader_vector.batchify
dev_loader = torch.utils.data.DataLoader(
dev_dataset,
batch_size=args.test_batch_size,
sampler=dev_sampler,
num_workers=args.data_workers,
collate_fn=batching_function,
pin_memory=args.cuda,
)
## Dev dataset for measuring performance of the trained sentence selector
if args.use_sentence_selector:
dev_dataset1 = selector_data.SentenceSelectorDataset(dev_exs, model.sentence_selector, single_answer=False)
#dev_dataset1.examples = [t for t in dev_dataset.examples if t is not None]
if args.sort_by_len:
dev_sampler1 = selector_data.SortedBatchSampler(dev_dataset1.lengths(),
args.test_batch_size,
shuffle=False)
else:
dev_sampler1 = torch.utils.data.sampler.SequentialSampler(dev_dataset1)
dev_loader1 = torch.utils.data.DataLoader(
dev_dataset1,
#batch_size=args.test_batch_size,
#sampler=dev_sampler1,
batch_sampler = dev_sampler1,
num_workers=args.data_workers,
collate_fn=selector_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}
# --------------------------------------------------------------------------
# QUICKLY VALIDATE ON PRETRAINED MODEL
if args.global_mode == "test":
result1 = validate_unofficial(args, dev_loader, model, stats, mode='dev')
result2 = validate_official(args, dev_loader, model, stats,
dev_offsets, dev_texts, dev_answers)
print(result2[args.valid_metric])
print(result1["exact_match"])
if args.use_sentence_selector:
sent_stats = {'timer': utils.Timer(), 'epoch': 0, 'best_valid': 0}
#sent_selector_results = validate_selector(model.sentence_selector.args, dev_loader1, model.sentence_selector, sent_stats, mode="dev")
#print("Sentence Selector model acheives:")
#print(sent_selector_results["accuracy"])
if len(args.adv_dev_json) > 0:
validate_adversarial(args, model, stats, mode="dev")
exit(0)
valid_history = []
bad_counter = 0
for epoch in range(start_epoch, args.num_epochs):
stats['epoch'] = epoch
# Train
train(args, train_loader, model, stats)
# Validate unofficial (train)
validate_unofficial(args, train_loader, model, stats, mode='train')
# Validate unofficial (dev)
result = validate_unofficial(args, dev_loader, model, stats, mode='dev')
# Validate official
if args.official_eval:
result = validate_official(args, dev_loader, model, stats,
dev_offsets, dev_texts, dev_answers)
# 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]
bad_counter = 0
else:
bad_counter += 1
if bad_counter > args.patience:
logger.info("Early Stopping at epoch: %d" % epoch)
exit(0)
def validate_adversarial(args, model, global_stats, mode="dev"):
# create dataloader for dev sets, load thier jsons, integrate the function
for idx, dataset_file in enumerate(args.adv_dev_json):
predictions = {}
logger.info("Validating Adversarial Dataset %s" % dataset_file)
exs = utils.load_data(args, args.adv_dev_file[idx])
logger.info('Num dev examples = %d' % len(exs))
## Create dataloader
dev_dataset = reader_data.ReaderDataset(exs, model, single_answer=False)
if args.sort_by_len:
dev_sampler = reader_data.SortedBatchSampler(dev_dataset.lengths(),
args.test_batch_size,
shuffle=False)
else:
dev_sampler = torch.utils.data.sampler.SequentialSampler(dev_dataset)
if args.use_sentence_selector:
dev_batcher = reader_vector.sentence_batchifier(model, single_answer=False)
#batching_function = dev_batcher.batchify
batching_function = reader_vector.batchify
else:
batching_function = reader_vector.batchify
dev_loader = torch.utils.data.DataLoader(
dev_dataset,
batch_size=args.test_batch_size,
sampler=dev_sampler,
num_workers=args.data_workers,
collate_fn=batching_function,
pin_memory=args.cuda,
)
texts = utils.load_text(dataset_file)
offsets = {ex['id']: ex['offsets'] for ex in exs}
answers = utils.load_answers(dataset_file)
eval_time = utils.Timer()
f1 = utils.AverageMeter()
exact_match = utils.AverageMeter()
examples = 0
bad_examples = 0
for ex in dev_loader:
ex_id, batch_size = ex[-1], ex[0].size(0)
chosen_offset = ex[-2]
pred_s, pred_e, _, pred_sp = model.predict(ex)
for i in range(batch_size):
if pred_s[i][0] >= len(offsets[ex_id[i]]) or pred_e[i][0] >= len(offsets[ex_id[i]]):
bad_examples += 1
continue
if args.use_sentence_selector:
s_offset = chosen_offset[i][pred_s[i][0]][0]
e_offset = chosen_offset[i][pred_e[i][0]][1]
else:
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]
if args.select_k > 1:
prediction = ""
offset_subset = chosen_offset[i][pred_s[i][0]: pred_e[i][0]]
for enum_, o in enumerate(offset_subset):
prediction += texts[ex_id[i]][o[0]:o[1]] + " "
prediction = prediction.strip()
predictions[ex_id[i]] = prediction
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 for dev file %s : Epoch = %d | EM = %.2f | ' %
(dataset_file, global_stats['epoch'], exact_match.avg * 100) +
'F1 = %.2f | examples = %d | valid time = %.2f (s)' %
(f1.avg * 100, examples, eval_time.time()))
orig_f1_score = 0.0
orig_exact_match_score = 0.0
adv_f1_scores = {} # Map from original ID to F1 score
adv_exact_match_scores = {} # Map from original ID to exact match score
adv_ids = {}
all_ids = set() # Set of all original IDs
f1 = exact_match = 0
dataset = json.load(open(dataset_file))['data']
for article in dataset:
for paragraph in article['paragraphs']:
for qa in paragraph['qas']:
orig_id = qa['id'].split('-')[0]
all_ids.add(orig_id)
if qa['id'] not in predictions:
message = 'Unanswered question ' + qa['id'] + ' will receive score 0.'
# logger.info(message)
continue
ground_truths = list(map(lambda x: x['text'], qa['answers']))
prediction = predictions[qa['id']]
cur_exact_match = utils.metric_max_over_ground_truths(utils.exact_match_score,
prediction, ground_truths)
cur_f1 = utils.metric_max_over_ground_truths(utils.f1_score, prediction, ground_truths)
if orig_id == qa['id']:
# This is an original example
orig_f1_score += cur_f1
orig_exact_match_score += cur_exact_match
if orig_id not in adv_f1_scores:
# Haven't seen adversarial example yet, so use original for adversary
adv_ids[orig_id] = orig_id
adv_f1_scores[orig_id] = cur_f1
adv_exact_match_scores[orig_id] = cur_exact_match
else:
# This is an adversarial example
if (orig_id not in adv_f1_scores or adv_ids[orig_id] == orig_id
or adv_f1_scores[orig_id] > cur_f1):
# Always override if currently adversary currently using orig_id
adv_ids[orig_id] = qa['id']
adv_f1_scores[orig_id] = cur_f1
adv_exact_match_scores[orig_id] = cur_exact_match
orig_f1 = 100.0 * orig_f1_score / len(all_ids)
orig_exact_match = 100.0 * orig_exact_match_score / len(all_ids)
adv_exact_match = 100.0 * sum(adv_exact_match_scores.values()) / len(all_ids)
adv_f1 = 100.0 * sum(adv_f1_scores.values()) / len(all_ids)
logger.info("For the file %s Original Exact Match : %.4f ; Original F1 : : %.4f | "
% (dataset_file, orig_exact_match, orig_f1)
+ "Adversarial Exact Match : %.4f ; Adversarial F1 : : %.4f " % (adv_exact_match, adv_f1))
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.
"""
clean_id_file = open(os.path.join(DATA_DIR, "clean_qids.txt"), "w+")
eval_time = utils.Timer()
f1 = utils.AverageMeter()
exact_match = utils.AverageMeter()
# Run through examples
examples = 0
bad_examples = 0
for ex in data_loader:
ex_id, batch_size = ex[-1], ex[0].size(0)
chosen_offset = ex[-2]
pred_s, pred_e, _ , pred_sp = model.predict(ex)
for i in range(batch_size):
if pred_s[i][0] >= len(offsets[ex_id[i]]) or pred_e[i][0] >= len(offsets[ex_id[i]]):
bad_examples += 1
continue
if args.use_sentence_selector:
s_offset = chosen_offset[i][pred_s[i][0]][0]
e_offset = chosen_offset[i][pred_e[i][0]][1]
else:
s_offset = offsets[ex_id[i]][pred_s[i][0]][0]
e_offset = offsets[ex_id[i]][pred_e[i][0]][1]
# If sentence selector is not turned on
if not args.use_sentence_selector or args.select_k == 1:
prediction = texts[ex_id[i]][s_offset:e_offset]
if args.select_k > 1:
prediction = ""
offset_subset = chosen_offset[i][pred_s[i][0]: pred_e[i][0] + 1]
for enum_, o in enumerate(offset_subset):
prediction += texts[ex_id[i]][o[0]:o[1]] + " "
prediction = prediction.strip()
# 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))
f1_example = utils.metric_max_over_ground_truths(
utils.f1_score, prediction, ground_truths)
if f1_example != 0:
clean_id_file.write(ex_id[i] + "\n")
examples += batch_size
clean_id_file.close()
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()))
logger.info('Bad Offset Examples during official eval: %d' % bad_examples)
return {'exact_match': exact_match.avg * 100, 'f1': f1.avg * 100}