def get_result(checkpoint):
"""Evaluate the checkpoint on SQuAD v2.0."""
# If running eval on the TPU, you will need to specify the number of
# steps.
reader = tf.train.NewCheckpointReader(checkpoint)
global_step = reader.get_tensor(tf.GraphKeys.GLOBAL_STEP)
all_results = []
for result in estimator.predict(predict_input_fn,
yield_single_examples=True,
checkpoint_path=checkpoint):
if len(all_results) % 1000 == 0:
tf.logging.info("Processing example: %d" %
(len(all_results)))
unique_id = int(result["unique_ids"])
start_top_log_probs = ([
float(x) for x in result["start_top_log_probs"].flat
])
start_top_index = [
int(x) for x in result["start_top_index"].flat
]
end_top_log_probs = ([
float(x) for x in result["end_top_log_probs"].flat
])
end_top_index = [int(x) for x in result["end_top_index"].flat]
cls_logits = float(result["cls_logits"].flat[0])
all_results.append(
squad_utils.RawResultV2(
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))
output_prediction_file = os.path.join(FLAGS.output_dir,
"predictions.json")
output_nbest_file = os.path.join(FLAGS.output_dir,
"nbest_predictions.json")
output_null_log_odds_file = os.path.join(FLAGS.output_dir,
"null_odds.json")
result_dict = {}
cls_dict = {}
squad_utils.accumulate_predictions_v2(
result_dict, cls_dict, eval_examples, eval_features,
all_results, FLAGS.n_best_size, FLAGS.max_answer_length,
FLAGS.start_n_top, FLAGS.end_n_top)
return squad_utils.evaluate_v2(
result_dict, cls_dict, prediction_json, eval_examples,
eval_features, all_results, FLAGS.n_best_size,
FLAGS.max_answer_length, output_prediction_file,
output_nbest_file, output_null_log_odds_file), int(global_step)
def get_result(checkpoint):
"""Evaluate the checkpoint on SQuAD v2.0."""
# If running eval on the TPU, you will need to specify the number of
# steps.
reader = tf.train.NewCheckpointReader(checkpoint)
global_step = reader.get_tensor(tf.GraphKeys.GLOBAL_STEP)
all_results = []
for result in estimator.predict(predict_input_fn,
yield_single_examples=True,
checkpoint_path=checkpoint):
if len(all_results) % 1000 == 0:
tf.logging.info("Processing example: %d" %
(len(all_results)))
unique_id = int(result["unique_ids"])
cls_logits = float(result["cls_logits"].flat[0])
all_results.append(
squad_utils.RawResultV2(unique_id=unique_id,
cls_logits=cls_logits))
output_prediction_file = os.path.join(FLAGS.output_dir,
"predictions.json")
output_nbest_file = os.path.join(FLAGS.output_dir,
"nbest_predictions.json")
output_null_log_odds_file = os.path.join(FLAGS.output_dir,
"null_odds.json")
result_dict = {}
cls_dict = {}
squad_utils.accumulate_predictions_v2(
result_dict, cls_dict, eval_examples, eval_features,
all_results, FLAGS.n_best_size, FLAGS.max_answer_length,
FLAGS.start_n_top, FLAGS.end_n_top)
from squad_utils import make_qid_to_has_ans
import numpy as np
qid_to_has_ans = make_qid_to_has_ans(
prediction_json) # maps qid to True/False
has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
print("has_ans", len(has_ans_qids))
print("no_ans", len(no_ans_qids))
def compute_metrics_with_threshold(threshold):
nonlocal result_dict
result_dict = {}
tp = 0
tn = 0
fp = 0
fn = 0
for example_index, example in enumerate(eval_examples):
m = np.min(cls_dict[example_index])
predict_is_impossible = 1 / (1 + np.exp(-m)) > threshold
# predict_is_impossible = m > threshold
result_dict[example.qas_id] = m
if example.is_impossible:
if predict_is_impossible:
tp += 1
else:
fn += 1
else:
if predict_is_impossible:
fp += 1
else:
tn += 1
precision = tp / (tp + fp)
recall = tp / (fn + tp)
f1 = 2 * tp / (2 * tp + fp + fn)
tf.logging.info(f"precision: {precision}"
f"recall: {recall}"
f"f1: {f1}")
return precision, recall, f1
# precision, recall, f1 = compute_metrics_with_threshold(0.4)
precision, recall, f1 = compute_metrics_with_threshold(0.5)
# precision, recall, f1 = compute_metrics_with_threshold(0.6)
with tf.gfile.GFile(output_prediction_file, "w") as writer:
writer.write(json.dumps(result_dict, indent=4) + "\n")
return {
"precision": precision,
"recall": recall,
"f1": f1,
"total": len(eval_examples)
}, int(global_step)