def main():
parser = argparse.ArgumentParser(
description=
"""Computes rationale and final class classification scores""",
formatter_class=argparse.RawTextHelpFormatter)
parser.add_argument(
'--data_dir',
dest='data_dir',
required=True,
help='Which directory contains a {train,val,test}.jsonl file?')
parser.add_argument('--split',
dest='split',
required=True,
help='Which of {train,val,test} are we scoring on?')
parser.add_argument('--strict',
dest='strict',
required=False,
action='store_true',
default=False,
help='Do we perform strict scoring?')
parser.add_argument('--results',
dest='results',
required=True,
help="""Results File
Contents are expected to be jsonl of:
{
"annotation_id": str, required
# these classifications *must not* overlap
# these classifications *must not* overlap
"rationales": List[
{
"docid": str, required
"hard_rationale_predictions": List[{
"start_token": int, inclusive, required
"end_token": int, exclusive, required
}], optional,
# token level classifications, a value must be provided per-token
# in an ideal world, these correspond to the hard-decoding above.
"soft_rationale_predictions": List[float], optional.
# sentence level classifications, a value must be provided for every
# sentence in each document, or not at all
"soft_sentence_predictions": List[float], optional.
}
],
# the classification the model made for the overall classification task
"classification": str, optional
# A probability distribution output by the model. We require this to be normalized.
"classification_scores": Dict[str, float], optional
# The next two fields are measures for how faithful your model is (the
# rationales it predicts are in some sense causal of the prediction), and
# how sufficient they are. We approximate a measure for comprehensiveness by
# asking that you remove the top k%% of tokens from your documents,
# running your models again, and reporting the score distribution in the
# "comprehensiveness_classification_scores" field.
# We approximate a measure of sufficiency by asking exactly the converse
# - that you provide model distributions on the removed k%% tokens.
# 'k' is determined by human rationales, and is documented in our paper.
# You should determine which of these tokens to remove based on some kind
# of information about your model: gradient based, attention based, other
# interpretability measures, etc.
# scores per class having removed k%% of the data, where k is determined by human comprehensive rationales
"comprehensiveness_classification_scores": Dict[str, float], optional
# scores per class having access to only k%% of the data, where k is determined by human comprehensive rationales
"sufficiency_classification_scores": Dict[str, float], optional
# the number of tokens required to flip the prediction - see "Is Attention Interpretable" by Serrano and Smith.
"tokens_to_flip": int, optional
}
When providing one of the optional fields, it must be provided for *every* instance.
The classification, classification_score, and comprehensiveness_classification_scores
must together be present for every instance or absent for every instance.
""")
parser.add_argument('--iou_thresholds',
dest='iou_thresholds',
required=False,
nargs='+',
type=float,
default=[0.5],
help='''Thresholds for IOU scoring.
These are used for "soft" or partial match scoring of rationale spans.
A span is considered a match if the size of the intersection of the prediction
and the annotation, divided by the union of the two spans, is larger than
the IOU threshold. This score can be computed for arbitrary thresholds.
''')
parser.add_argument('--score_file',
dest='score_file',
required=False,
default=None,
help='Where to write results?')
args = parser.parse_args()
results = load_jsonl(args.results)
docids = set(
chain.from_iterable([rat['docid'] for rat in res['rationales']]
for res in results))
docs = load_flattened_documents(args.data_dir, docids)
verify_instances(results, docs)
# load truth
annotations = annotations_from_jsonl(
os.path.join(args.data_dir, args.split + '.jsonl'))
docids |= set(
chain.from_iterable((ev.docid
for ev in chain.from_iterable(ann.evidences))
for ann in annotations))
has_final_predictions = _has_classifications(results)
scores = dict()
if args.strict:
if not args.iou_thresholds:
raise ValueError(
"--iou_thresholds must be provided when running strict scoring"
)
if not has_final_predictions:
raise ValueError(
"We must have a 'classification', 'classification_score', and 'comprehensiveness_classification_score' field in order to perform scoring!"
)
# TODO think about offering a sentence level version of these scores.
if _has_hard_predictions(results):
truth = list(
chain.from_iterable(
Rationale.from_annotation(ann) for ann in annotations))
pred = list(
chain.from_iterable(
Rationale.from_instance(inst) for inst in results))
if args.iou_thresholds is not None:
iou_scores = partial_match_score(truth, pred, args.iou_thresholds)
scores['iou_scores'] = iou_scores
# NER style scoring
rationale_level_prf = score_hard_rationale_predictions(truth, pred)
scores['rationale_prf'] = rationale_level_prf
token_level_truth = list(
chain.from_iterable(rat.to_token_level() for rat in truth))
token_level_pred = list(
chain.from_iterable(rat.to_token_level() for rat in pred))
token_level_prf = score_hard_rationale_predictions(
token_level_truth, token_level_pred)
scores['token_prf'] = token_level_prf
else:
logging.info(
"No hard predictions detected, skipping rationale scoring")
if _has_soft_predictions(results):
flattened_documents = load_flattened_documents(args.data_dir, docids)
paired_scoring = PositionScoredDocument.from_results(
results, annotations, flattened_documents, use_tokens=True)
token_scores = score_soft_tokens(paired_scoring)
scores['token_soft_metrics'] = token_scores
else:
logging.info(
"No soft predictions detected, skipping rationale scoring")
if _has_soft_sentence_predictions(results):
documents = load_documents(args.data_dir, docids)
paired_scoring = PositionScoredDocument.from_results(results,
annotations,
documents,
use_tokens=False)
sentence_scores = score_soft_tokens(paired_scoring)
scores['sentence_soft_metrics'] = sentence_scores
else:
logging.info(
"No sentence level predictions detected, skipping sentence-level diagnostic"
)
if has_final_predictions:
flattened_documents = load_flattened_documents(args.data_dir, docids)
class_results = score_classifications(results, annotations,
flattened_documents)
scores['classification_scores'] = class_results
else:
logging.info(
"No classification scores detected, skipping classification")
pprint.pprint(scores)
if args.score_file:
with open(args.score_file, 'w') as of:
json.dump(scores, of, indent=4, sort_keys=True)
def load_and_cache_examples(args,
model_params,
tokenizer,
evaluate=False,
split="train",
output_examples=False):
if args.local_rank not in [-1, 0] and not evaluate:
torch.distributed.barrier(
) # Make sure only the first process in distributed training process the dataset, and the others will use the cache
# only load one split
input_file = os.path.join(args.data_dir, split)
cached_features_file = os.path.join(
os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
split,
list(filter(None,
model_params["tokenizer_name"].split('/'))).pop(),
str(args.max_seq_length)))
if args.gold_evidence:
cached_features_file += "_goldevidence"
if os.path.exists(cached_features_file
) and not args.overwrite_cache and not output_examples:
logger.info("Loading features from cached file %s",
cached_features_file)
features = torch.load(cached_features_file)
else:
logger.info("Creating features from dataset file at %s", input_file)
dataset = annotations_from_jsonl(
os.path.join(args.data_dir, split + ".jsonl"))
docids = set(e.docid
for e in chain.from_iterable(
chain.from_iterable(
map(lambda ann: ann.evidences, chain(dataset)))))
documents = load_documents(args.data_dir, docids)
if args.out_domain:
examples = read_json(args)
else:
examples = read_examples(args, model_params, dataset, documents,
split)
features = convert_examples_to_features(
args,
model_params,
examples=examples,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
max_query_length=args.max_query_length,
is_training=not evaluate)
if args.local_rank in [-1, 0]:
logger.info("Saving features into cached file %s",
cached_features_file)
torch.save(features, cached_features_file)
if args.local_rank == 0 and not evaluate:
torch.distributed.barrier()
# Tensorize all features
all_input_ids = torch.tensor([f.input_ids for f in features],
dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in features],
dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in features],
dtype=torch.long)
all_cls_index = torch.tensor([f.cls_index for f in features],
dtype=torch.long)
all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
all_unique_ids = torch.tensor([f.unique_id for f in features],
dtype=torch.float)
if evaluate:
all_example_index = torch.arange(all_input_ids.size(0),
dtype=torch.long)
tensorized_dataset = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_example_index,
all_cls_index, all_p_mask,
all_unique_ids)
else:
all_class_labels = torch.tensor([f.class_label for f in features],
dtype=torch.long)
all_evidence_labels = torch.tensor(
[f.evidence_label for f in features], dtype=torch.long)
tensorized_dataset = TensorDataset(all_input_ids, all_input_mask,
all_segment_ids, all_class_labels,
all_cls_index, all_p_mask,
all_unique_ids, all_evidence_labels)
if output_examples:
return tensorized_dataset, examples, features
return tensorized_dataset, features
"annotation_id": annotation_id,
'classification': classification,
"docids": docids,
"query": query,
"query_type": None,
'evidences': [converted_evidences],
}
test_eraser_dataset.append(eraser_ex)
print("Avg evidence sentences %f " % np.average(evs))
train_eraser_file = open(os.path.join(args.data_path, "train.jsonl"), "w+")
val_eraser_file = open(os.path.join(args.data_path, "val.jsonl"), "w+")
test_eraser_file = open(os.path.join(args.data_path, "test.jsonl"), "w+")
for ex in train_eraser_dataset:
train_eraser_file.write(json.dumps(ex) + "\n")
for ex in dev_eraser_dataset:
val_eraser_file.write(json.dumps(ex) + "\n")
for ex in test_eraser_dataset:
test_eraser_file.write(json.dumps(ex) + "\n")
train_eraser_file.close()
val_eraser_file.close()
test_eraser_file.close()
# Verify whether BEER can be opened using
test_dataset = annotations_from_jsonl(
os.path.join(args.data_path, "test.jsonl"))
train_dataset = annotations_from_jsonl(
os.path.join(args.data_path, "train.jsonl"))
val_dataset = annotations_from_jsonl(
os.path.join(args.data_path, "val.jsonl"))
def evaluate(args,
model_params,
model,
tokenizer,
prefix="",
output_examples=False,
split="val"):
# Keep all eraser annotations at the ready to go
annotations = annotations_from_jsonl(
os.path.join(args.data_dir, split + '.jsonl'))
true_rationales = list(
chain.from_iterable(
metrics.Rationale.from_annotation(ann) for ann in annotations))
if output_examples:
dataset, examples, features = load_and_cache_examples(
args,
model_params,
tokenizer,
evaluate=True,
split=split,
output_examples=output_examples)
else:
dataset, features = load_and_cache_examples(
args,
model_params,
tokenizer,
evaluate=True,
split=split,
output_examples=output_examples)
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(
dataset) if args.local_rank == -1 else DistributedSampler(dataset)
eval_dataloader = DataLoader(dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
all_results = []
all_targets = []
# all_rationale_targets = []
all_rationale_results = []
all_results_dictionary = {}
results = []
class_interner = dict(
(y, x) for (x, y) in enumerate(model_params['classes']))
class_labels = [
k for k, v in sorted(class_interner.items(), key=lambda x: x[1])
]
task_performances = []
evidence_performances = []
for run_no in range(10):
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
# 'token_type_ids': None if args.model_type == 'xlm' or args.model_type == 'roberta' else batch[2] # XLM don't use segment_ids
'p_mask': batch[5]
}
example_indices = batch[3]
if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
inputs.update({"evaluate": True})
outputs = model(**inputs)
logits = outputs[0]
hard_preds = to_list(torch.argmax(logits.float(), dim=-1))
hard_rationale_pred = outputs[1]
all_results.extend(hard_preds)
for i, example_index in enumerate(example_indices):
eval_feature = features[example_index.item()]
unique_id = int(eval_feature.unique_id)
all_targets.extend([eval_feature.class_label])
# hard_rationale_binary = np.zeros(eval_feature.evidence_label.shape)
# hard_rationale_binary[np.array(to_list(hard_rationale_pred[i]))] = 1
# all_rationale_results.extend(hard_rationale_binary[:sum(eval_feature.input_mask)])
# all_rationale_targets.extend(eval_feature.evidence_label[:sum(eval_feature.input_mask)])
# all_rationale_dictionary[unique_id] = hard_rationale_binary[:sum(eval_feature.input_mask)]
key = (eval_feature.annotation_id, eval_feature.doc_id)
for chunk in hard_rationale_pred[i]:
chunk = to_list(chunk)
start_token = eval_feature.token_to_orig_map[chunk[0]]
chunk_end = chunk[-1]
if eval_feature.tokens[chunk[-1]] in ["[SEP]", "</s>"]:
# Last chunk which also has SEP and not part of map to original tokens
chunk_end = chunk[-1] - 1
end_token = eval_feature.token_to_orig_map[
chunk_end] + 1 # The final token is still exclusive
# token_to_orig_map hard to implement if Roberta is being used (token level doesn't do as well)
all_rationale_results.append(
metrics.Rationale(ann_id=key[0],
docid=key[1],
start_token=start_token,
end_token=end_token))
all_results_dictionary[key] = hard_preds[i]
output_ratioanale_prediction_file = os.path.join(
args.output_dir, "rationale_predictions.json")
file_pointer = open(output_ratioanale_prediction_file, "w+")
ann_to_rat = metrics._keyed_rationale_from_list(all_rationale_results)
ann_to_gold = metrics._keyed_rationale_from_list(true_rationales)
example_dict = dict()
for ex in examples:
example_dict[(ex.id, ex.doc_id)] = ex
evidence_iou = metrics.partial_match_score(true_rationales,
all_rationale_results,
[0.5])
hard_rationale_metrics = metrics.score_hard_rationale_predictions(
true_rationales, all_rationale_results)
# logger.info("Rationale Partial score metrics:")
# logger.info(evidence_iou)
# logger.info("Rationale hard metrics:")
# logger.info(hard_rationale_metrics)
results = (classification_report(
all_targets,
all_results,
target_names=class_labels,
output_dict=True)['macro avg']['f1-score'],
accuracy_score(all_targets, all_results))
# logger.info('Classification Report: {}'.format(
# classification_report(all_targets, all_results, target_names=class_labels, output_dict=True)))
task_performances.append(results[0])
evidence_performances.append(evidence_iou[0]['macro']['f1'])
# currently considers the last evaluation run sample
for e, feature in enumerate(features):
unique_id = (feature.annotation_id, feature.doc_id)
tokens = feature.tokens
token_level_rationales = ann_to_rat[unique_id]
doc_tokens = example_dict[unique_id].doc_toks
# This technique is aslightly flawed
# rationale = [tokens[i] for i in np.where(token_level_rationale == 1)[0]]
# # Every chunk is separable
# rationale_text = ""
# for i in range(model_params["K"]):
# rationale_text += " ".join(rationale[i*model_params["chunk_size"]:(i+1)*model_params["chunk_size"]]) + "\n"
class_label = class_labels[feature.class_label]
predicted_class_label = all_results_dictionary[unique_id]
# human_rationale = [tokens[i] for i in np.where(np.asarray(feature.evidence_label)== 1)[0]] # This is stale too since there is extra padding into what both models consider
# query = " ".join([tok for tok in feature.tokens[:64] if tok not in ["[CLS]", "[PAD]", "[SEP]"]])
rationales = []
# {
# "docid": str, required
# "hard_rationale_predictions": List[{
# "start_token": int, inclusive, required
# "end_token": int, exclusive, required
# }]
for rationale in token_level_rationales:
start_token = rationale.start_token
end_token = rationale.end_token
rationale_text = doc_tokens[start_token:end_token]
rationales.append({
"start_token": start_token,
"end_token": end_token
})
output_dict = {
"annotation_id":
feature.annotation_id,
"classification":
predicted_class_label,
"rationales": [{
"docid": feature.doc_id,
"hard_rationale_predictions": rationales
}],
"gold_classification":
class_label,
# "gold_rationales" : ann_to_gold[unique_id],
# "predicted_rationales" : ann_to_rat[unique_id],
"query":
example_dict[unique_id].query,
"doc_tokens":
doc_tokens
}
file_pointer.write(json.dumps(output_dict) + "\n")
# Print the rationale chunks predicted to file
file_pointer.close()
logger.info("Averaged classification results: %.4f" %
np.average(task_performances))
logger.info("Averaged Evidence results: %.4f" %
np.average(evidence_performances))
return results
def train(args, model_params, train_dataset, model, tokenizer):
if args.local_rank in [-1, 0] and args.tf_summary:
#if os.path.isdir(os.path.join("runs", os.path.basename(args.output_dir))):
# shutil.rmtree(os.path.join("runs", os.path.basename(args.output_dir)))
tb_writer = SummaryWriter("runs/" + os.path.basename(args.output_dir))
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
train_sampler = RandomSampler(
train_dataset) if args.local_rank == -1 else DistributedSampler(
train_dataset)
train_dataloader = DataLoader(train_dataset,
sampler=train_sampler,
batch_size=args.train_batch_size)
if args.max_steps > 0:
t_total = args.max_steps
args.num_train_epochs = args.max_steps // (
len(train_dataloader) // args.gradient_accumulation_steps) + 1
else:
t_total = len(
train_dataloader
) // args.gradient_accumulation_steps * args.num_train_epochs
no_decay = ["bias", "LayerNorm.weight"]
optimizer_grouped_parameters = [
{
"params": [
p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay)
],
"weight_decay":
args.weight_decay,
},
{
"params": [
p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay)
],
"weight_decay":
0.0
},
]
# TODO: Addiitonal parameters require more learning rate compared to BERT model
optimizer = AdamW(optimizer_grouped_parameters,
lr=args.learning_rate,
eps=args.adam_epsilon)
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=args.warmup_steps,
num_training_steps=t_total)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Distributed training (should be after apex fp16 initialization)
if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(
model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
if args.evaluate_during_training:
dataset, eval_features = load_and_cache_examples(args,
model_params,
tokenizer,
evaluate=True,
split="val",
output_examples=False)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(
1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(
dataset) if args.local_rank == -1 else DistributedSampler(dataset)
eval_dataloader = DataLoader(dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# Keep all eraser annotations at the ready to go
annotations = annotations_from_jsonl(
os.path.join(args.data_dir, args.eval_split + '.jsonl'))
true_rationales = list(
chain.from_iterable(
metrics.Rationale.from_annotation(ann) for ann in annotations))
# Train!
logger.info("***** Running training *****")
logger.info(" Num examples = %d", len(train_dataset))
logger.info(" Num Epochs = %d", args.num_train_epochs)
logger.info(" Instantaneous batch size per GPU = %d",
args.per_gpu_train_batch_size)
logger.info(
" Total train batch size (w. parallel, distributed & accumulation) = %d",
args.train_batch_size * args.gradient_accumulation_steps *
(torch.distributed.get_world_size() if args.local_rank != -1 else 1))
logger.info(" Gradient Accumulation steps = %d",
args.gradient_accumulation_steps)
logger.info(" Total optimization steps = %d", t_total)
global_step = 0
tr_loss, logging_loss = 0.0, 0.0
class_tr_loss, class_logging_loss = 0.0, 0.0
info_tr_loss, info_logging_loss = 0.0, 0.0
model.zero_grad()
train_iterator = trange(int(args.num_train_epochs),
desc="Epoch",
disable=args.local_rank not in [-1, 0])
set_seed(
args) # Added here for reproductibility (even between python 2 and 3)
best_f1 = (-1, -1) # Classfication F1 and accuracy
wait_step = 0
stop_training = False
metric_name = "F1"
epoch = 0
for _ in train_iterator:
epoch_iterator = tqdm(train_dataloader,
desc="Iteration",
disable=args.local_rank not in [-1, 0])
for step, batch in enumerate(epoch_iterator):
model.train()
batch = tuple(t.to(args.device) for t in batch)
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
# 'token_type_ids': None if args.model_type == 'xlm' or args.model_type == 'roberta' else batch[2],
'labels': batch[3],
'p_mask': batch[5]
}
if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
outputs = model(**inputs)
loss = outputs[
0] # model outputs are always tuple in pytorch-transformers (see doc)
class_loss = outputs[1]
info_loss = outputs[2]
if args.n_gpu > 1:
loss = loss.mean(
) # mean() to average on multi-gpu parallel (not distributed) training
class_loss = class_loss.mean()
info_loss = info_loss.mean()
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
class_loss = class_loss / args.gradient_accumulation_steps
info_loss = info_loss / args.gradient_accumulation_steps
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(),
args.max_grad_norm)
tr_loss += loss.item()
class_tr_loss += class_loss.item()
info_tr_loss += info_loss.item()
if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step()
scheduler.step()
model.zero_grad()
global_step += 1
if args.local_rank in [
-1, 0
] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
# Log metrics
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
results = predict(args, model_params, model, tokenizer,
eval_features, eval_dataloader,
true_rationales, tb_writer,
global_step)
#for key, value in results.items():
# tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
if best_f1 < results:
logger.info("Saving model with best %s: %.2f (Acc %.2f) -> %.2f (Acc %.2f) on epoch=%d" % \
(metric_name, best_f1[0] * 100, best_f1[1] * 100, results[0] * 100, results[1] * 100,
epoch))
output_dir = os.path.join(args.output_dir,
'best_model')
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = model.module if hasattr(
model, 'module'
) else model # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
torch.save(
args,
os.path.join(output_dir, 'training_args.bin'))
best_f1 = results
wait_step = 0
stop_training = False
else:
wait_step += 1
if wait_step == args.wait_step:
logger.info("Loosing Patience")
stop_training = True
if args.tf_summary and global_step % 20 == 0:
tb_writer.add_scalar('lr',
scheduler.get_lr()[0], global_step)
tb_writer.add_scalar('loss', (tr_loss - logging_loss) /
args.logging_steps, global_step)
tb_writer.add_scalar('class_loss',
(class_tr_loss - class_logging_loss) /
20, global_step)
tb_writer.add_scalar(
'info_loss', (info_tr_loss - info_logging_loss) / 20,
global_step)
logging_loss = tr_loss
class_logging_loss = class_tr_loss
info_logging_loss = info_tr_loss
if args.local_rank in [
-1, 0
] and args.save_steps > 0 and global_step % args.save_steps == 0:
# Save model checkpoint
output_dir = os.path.join(
args.output_dir, 'checkpoint-{}'.format(global_step))
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = model.module if hasattr(
model, 'module'
) else model # Take care of distributed/parallel training
model_to_save.save_pretrained(output_dir)
torch.save(args,
os.path.join(output_dir, 'training_args.bin'))
logger.info("Saving model checkpoint to %s", output_dir)
if stop_training or args.max_steps > 0 and global_step > args.max_steps:
epoch_iterator.close()
break
if stop_training or args.max_steps > 0 and global_step > args.max_steps:
train_iterator.close()
break
epoch += 1
if args.local_rank in [-1, 0] and args.tf_summary:
tb_writer.close()
return global_step, tr_loss / global_step
from ib_utils import read_examples
import argparse
import json
from itertools import chain
from eraser.rationale_benchmark.utils import load_documents, annotations_from_jsonl
import logging, os
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir', dest='data_dir', required=True)
parser.add_argument('--model_params', dest='model_params', required=True)
parser.add_argument("--split", type=str, default="val")
parser.add_argument("--truncate", default=False, action="store_true")
parser.add_argument("--max_seq_length", default=512, type=int)
parser.add_argument("--max_query_length", default=24, type=int)
parser.add_argument("--max_num_sentences", default=20, type=int)
parser.add_argument("--debug", action="store_true", default=False)
parser.add_argument('--low_resource', action="store_true", default=False)
args = parser.parse_args()
# Parse model args json
with open(args.model_params, 'r') as fp:
logging.debug(f'Loading model parameters from {args.model_params}')
model_params = json.load(fp)
dataset = annotations_from_jsonl(
os.path.join(args.data_dir, args.split + ".jsonl"))
docids = set(e.docid for e in chain.from_iterable(
chain.from_iterable(map(lambda ann: ann.evidences, chain(dataset)))))
documents = load_documents(args.data_dir, docids)
examples = read_examples(args, model_params, dataset, documents, args.split)
def evaluate(args,
model_params,
model,
tokenizer,
prefix="",
output_examples=False,
split="val"):
# Keep all eraser annotations at the ready to go
annotations = annotations_from_jsonl(
os.path.join(args.data_dir, split + '.jsonl'))
true_rationales = list(
chain.from_iterable(
metrics.Rationale.from_annotation(ann) for ann in annotations))
if output_examples:
dataset, examples, features = load_and_cache_examples(
args,
model_params,
tokenizer,
evaluate=True,
split=split,
output_examples=output_examples)
else:
dataset, features = load_and_cache_examples(
args,
model_params,
tokenizer,
evaluate=True,
split=split,
output_examples=output_examples)
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
os.makedirs(args.output_dir)
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
# Note that DistributedSampler samples randomly
eval_sampler = SequentialSampler(
dataset) if args.local_rank == -1 else DistributedSampler(dataset)
eval_dataloader = DataLoader(dataset,
sampler=eval_sampler,
batch_size=args.eval_batch_size)
# Eval!
logger.info("***** Running evaluation {} *****".format(prefix))
logger.info(" Num examples = %d", len(dataset))
logger.info(" Batch size = %d", args.eval_batch_size)
# all_rationale_targets = []
all_rationale_results = []
class_interner = dict(
(y, x) for (x, y) in enumerate(model_params['classes']))
class_labels = [
k for k, v in sorted(class_interner.items(), key=lambda x: x[1])
]
for batch in tqdm(eval_dataloader, desc="Evaluating"):
model.eval()
batch = tuple(t.to(args.device) for t in batch)
with torch.no_grad():
inputs = {
'input_ids': batch[0],
'attention_mask': batch[1],
# 'token_type_ids': None if args.model_type == 'xlm' or args.model_type == 'roberta' else batch[2] # XLM don't use segment_ids
'p_mask': batch[5]
}
example_indices = batch[3]
if args.model_type in ['xlnet', 'xlm']:
inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
inputs.update({
"sentence_starts": batch[7],
"sentence_ends": batch[8],
"sentence_mask": batch[9],
"evidence_labels": batch[10],
})
inputs.update({"evaluate": True})
outputs = model(**inputs)
logits = outputs[0]
hard_rationale_pred = to_list(torch.sigmoid(logits))
hard_rationale_pred = [
np.where(np.asarray(l) > 0.5)[0] for l in hard_rationale_pred
]
for i, example_index in enumerate(example_indices):
eval_feature = features[example_index.item()]
unique_id = int(eval_feature.unique_id)
# based on indices selected; get corresponding start and end ; add those tokens to the Rationale list
key = (eval_feature.annotation_id, eval_feature.doc_id)
last_sentence = sum(eval_feature.sentence_mask)
for sentence in hard_rationale_pred[i]:
if sentence >= last_sentence:
continue
start_token = eval_feature.sentence_starts[sentence]
end_token = eval_feature.sentence_ends[sentence]
orig_start_token = eval_feature.token_to_orig_map[start_token]
if eval_feature.tokens[end_token] in ["[SEP]", "</s>"]:
# last sentence, then end_token points to SEP, we want to get it to point to last word
end_token -= 1
orig_end_token = eval_feature.token_to_orig_map[
end_token] + 1 # +1 since it needs to be exclusive
all_rationale_results.append(
metrics.Rationale(ann_id=key[0],
docid=key[1],
start_token=orig_start_token,
end_token=orig_end_token))
output_ratioanale_prediction_file = os.path.join(
args.output_dir, "rationale_predictions.json")
file_pointer = open(output_ratioanale_prediction_file, "w+")
ann_to_rat = metrics._keyed_rationale_from_list(all_rationale_results)
ann_to_gold = metrics._keyed_rationale_from_list(true_rationales)
example_dict = dict()
for ex in examples:
example_dict[(ex.id, ex.doc_id)] = ex
evidence_iou = metrics.partial_match_score(true_rationales,
all_rationale_results,
[0.1, 0.5, 0.9])
for e, feature in enumerate(features):
unique_id = (feature.annotation_id, feature.doc_id)
tokens = feature.tokens
token_level_rationales = ann_to_rat[unique_id]
doc_tokens = example_dict[unique_id].doc_toks
class_label = class_labels[feature.class_label]
# human_rationale = [tokens[i] for i in np.where(np.asarray(feature.evidence_label)== 1)[0]] # This is stale too since there is extra padding into what both models consider
# query = " ".join([tok for tok in feature.tokens[:64] if tok not in ["[CLS]", "[PAD]", "[SEP]"]])
rationales = []
for rationale in token_level_rationales:
start_token = rationale.start_token
end_token = rationale.end_token
rationale_text = doc_tokens[start_token:end_token]
rationales.append({
"start_token": start_token,
"end_token": end_token
})
output_dict = {
"annotation_id":
feature.annotation_id,
"rationales": [{
"docid": feature.doc_id,
"hard_rationale_predictions": rationales
}],
"gold_classification":
class_label,
# "gold_rationales" : ann_to_gold[unique_id],
# "predicted_rationales" : ann_to_rat[unique_id],
"query":
example_dict[unique_id].query,
"doc_tokens":
doc_tokens
}
file_pointer.write(json.dumps(output_dict) + "\n")
# Print the rationale chunks predicted to file
result = evidence_iou
file_pointer.close()
return result
