def predict(test_exe,
test_program,
test_pyreader,
fetch_list,
processor,
test_path=None):
if not os.path.exists(args.checkpoints):
os.makedirs(args.checkpoints)
output_prediction_file = os.path.join(args.checkpoints, "predictions.json")
output_nbest_file = os.path.join(args.checkpoints,
"nbest_predictions.json")
output_null_log_odds_file = os.path.join(args.checkpoints,
"null_odds.json")
test_pyreader.start()
all_results = []
time_begin = time.time()
while True:
try:
np_unique_ids, np_start_logits, np_end_logits, np_num_seqs = test_exe.run(
fetch_list=fetch_list, program=test_program)
for idx in range(np_unique_ids.shape[0]):
if len(all_results) % 1000 == 0:
print("Processing example: %d" % len(all_results))
unique_id = int(np_unique_ids[idx])
start_logits = [float(x) for x in np_start_logits[idx].flat]
end_logits = [float(x) for x in np_end_logits[idx].flat]
all_results.append(
RawResult(unique_id=unique_id,
start_logits=start_logits,
end_logits=end_logits))
except fluid.core.EOFException:
test_pyreader.reset()
break
time_end = time.time()
features = processor.get_features(processor.predict_examples,
is_training=False)
if test_path is None:
adv_f1 = write_predictions(
processor.predict_examples, features, all_results,
args.n_best_size, args.max_answer_length, args.do_lower_case,
output_prediction_file, output_nbest_file,
output_null_log_odds_file, args.version_2_with_negative,
args.null_score_diff_threshold, args.verbose, args.predict_file)
else:
adv_f1 = write_predictions(
processor.predict_examples, features, all_results,
args.n_best_size, args.max_answer_length, args.do_lower_case,
output_prediction_file, output_nbest_file,
output_null_log_odds_file, args.version_2_with_negative,
args.null_score_diff_threshold, args.verbose, test_path)
print(adv_f1)
return adv_f1
def predict(test_exe, test_program, test_data_loader, fetch_list, processor,
name):
if not os.path.exists(args.checkpoints):
os.makedirs(args.checkpoints)
output_prediction_file = os.path.join(args.checkpoints,
name + "predictions.json")
output_nbest_file = os.path.join(args.checkpoints,
name + "nbest_predictions.json")
output_null_log_odds_file = os.path.join(args.checkpoints,
name + "null_odds.json")
test_data_loader.start()
all_results = []
time_begin = time.time()
while True:
try:
outputs = test_exe.run(fetch_list=fetch_list, program=test_program)
np_unique_ids, np_start_top_log_probs, np_start_top_index, np_end_top_log_probs, np_end_top_index, np_cls_logits, \
= outputs[0:6]
for idx in range(np_unique_ids.shape[0]):
if len(all_results) % 1000 == 0:
print("Processing example: %d" % len(all_results))
unique_id = int(np_unique_ids[idx])
start_top_log_probs = [
float(x) for x in np_start_top_log_probs[idx].flat
]
start_top_index = [
int(x) for x in np_start_top_index[idx].flat
]
end_top_log_probs = [
float(x) for x in np_end_top_log_probs[idx].flat
]
end_top_index = [int(x) for x in np_end_top_index[idx].flat]
cls_logits = float(np_cls_logits[idx].flat[0])
all_results.append(
RawResult(unique_id=unique_id,
start_top_log_probs=start_top_log_probs,
start_top_index=start_top_index,
end_top_log_probs=end_top_log_probs,
end_top_index=end_top_index,
cls_logits=cls_logits))
except fluid.core.EOFException:
test_data_loader.reset()
break
time_end = time.time()
with io.open(args.predict_file, "r", encoding="utf8") as f:
orig_data = json.load(f)["data"]
features = processor.get_features(processor.predict_examples,
is_training=False)
ret = write_predictions(processor.predict_examples, features, all_results,
args.n_best_size, args.max_answer_length,
output_prediction_file, output_nbest_file,
output_null_log_odds_file, orig_data, args)
# Log current result
print("=" * 80)
log_str = "Result | "
for key, val in ret.items():
log_str += "{} {} | ".format(key, val)
print(log_str)
print("=" * 80)
