Skip to content
This repository was archived by the owner on Jan 15, 2024. It is now read-only.

Add sorting of chunks to evaluation #1588

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
Open

Add sorting of chunks to evaluation #1588

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3d27013
Add sorting of chunks to evaluation
bartekkuncer Apr 8, 2022
fa89f9e
Fix sorting chunks with quantization
bartekkuncer Apr 20, 2022
d1c078d
Add sort flag
bartekkuncer Jul 12, 2022
cf93fef
Fix bad commit
bartekkuncer Jul 12, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions scripts/question_answering/run_squad.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -126,6 +126,8 @@ def parse_args():
'may increase for mixed precision training on GPUs with TensorCores.')
parser.add_argument('--overwrite_cache', action='store_true',
help='Whether to overwrite the feature cache.')
parser.add_argument('--sort_input_data', action='store_true',
help='Whether to sort input data before evaluation.')
# Evaluation hyperparameters
parser.add_argument('--start_top_n', type=int, default=5,
help='Number of start-position candidates')
Expand Down Expand Up @@ -837,6 +839,8 @@ def evaluate(args, last=True):

logging.info('Prepare dev data')
dev_features = get_squad_features(args, tokenizer, segment='dev')
if args.sort_input_data:
dev_features.sort(key=lambda x: x.qas_id)
dev_data_path = os.path.join(args.data_dir, 'dev-v{}.json'.format(args.version))
dataset_processor = SquadDatasetProcessor(tokenizer=tokenizer,
doc_stride=args.doc_stride,
Expand All @@ -848,6 +852,8 @@ def evaluate(args, last=True):
chunk_features = dataset_processor.process_sample(feature)
dev_all_chunk_features.extend(chunk_features)
dev_chunk_feature_ptr.append(dev_chunk_feature_ptr[-1] + len(chunk_features))
if args.sort_input_data:
dev_all_chunk_features.sort(key=lambda x: x.valid_length, reverse=True)

def eval_validation(ckpt_name, best_eval):
"""
Expand Down Expand Up @@ -912,6 +918,9 @@ def eval_validation(ckpt_name, best_eval):
all_predictions = collections.OrderedDict()
all_nbest_json = collections.OrderedDict()
no_answer_score_json = collections.OrderedDict()
if args.sort_input_data:
all_results.sort(key=lambda x: x.qas_id)
dev_all_chunk_features.sort(key=lambda x: x.qas_id)
for index, (left_index, right_index) in enumerate(zip(dev_chunk_feature_ptr[:-1],
dev_chunk_feature_ptr[1:])):
chunked_features = dev_all_chunk_features[left_index:right_index]
Expand Down