def valid(data_processor, max_batches=None, batch_size=1024):
preds_kvpair = []
golds_kvpair = []
batches_sample = 0
while True:
(inputs_seq_batch, inputs_mask_batch, inputs_segment_batch,
outputs_seq_bio_batch,
outputs_seq_attr_batch) = data_processor.get_batch(batch_size)
feed_dict = {
model.inputs_seq: inputs_seq_batch,
model.inputs_mask: inputs_mask_batch,
model.inputs_segment: inputs_segment_batch
}
preds_seq_bio_batch, preds_seq_attr_batch = sess.run(
model.outputs, feed_dict)
for pred_seq_bio, gold_seq_bio, pred_seq_attr, gold_seq_attr, input_seq, mask in zip(
preds_seq_bio_batch, outputs_seq_bio_batch,
preds_seq_attr_batch, outputs_seq_attr_batch,
inputs_seq_batch, inputs_mask_batch):
l = sum(mask) - 2
pred_seq_bio = [i2w_bio[i] for i in pred_seq_bio[1:-1][:l]]
gold_seq_bio = [i2w_bio[i] for i in gold_seq_bio[1:-1][:l]]
char_seq = [i2w_char[i] for i in input_seq[1:-1][:l]]
pred_seq_attr = [
i2w_attr[i] for i in pred_seq_attr[1:-1][:l]
]
gold_seq_attr = [
i2w_attr[i] for i in gold_seq_attr[1:-1][:l]
]
pred_kvpair = extract_kvpairs_in_bioes(
pred_seq_bio, char_seq, pred_seq_attr)
gold_kvpair = extract_kvpairs_in_bioes(
gold_seq_bio, char_seq, gold_seq_attr)
preds_kvpair.append(pred_kvpair)
golds_kvpair.append(gold_kvpair)
if data_processor.end_flag:
data_processor.refresh()
break
batches_sample += 1
if (max_batches
is not None) and (batches_sample >= max_batches):
break
p, r, f1 = cal_f1_score(preds_kvpair, golds_kvpair)
logger.info("Valid Samples: {}".format(len(preds_kvpair)))
logger.info("Valid P/R/F1: {} / {} / {}".format(
round(p * 100, 2), round(r * 100, 2), round(f1 * 100, 2)))
return (p, r, f1)
def valid(data_processor, max_batches=None, batch_size=1024):
preds_kvpair = []
golds_kvpair = []
batches_sample = 0
while True:
(inputs_seq_batch,
inputs_seq_len_batch,
outputs_seq_batch) = data_processor.get_batch(batch_size)
feed_dict = {
model.inputs_seq: inputs_seq_batch,
model.inputs_seq_len: inputs_seq_len_batch,
model.outputs_seq: outputs_seq_batch
}
preds_seq_batch = sess.run(model.outputs, feed_dict)
for pred_seq, gold_seq, input_seq, l in zip(preds_seq_batch,
outputs_seq_batch,
inputs_seq_batch,
inputs_seq_len_batch):
pred_seq = [i2w_bio[i] for i in pred_seq[:l]]
gold_seq = [i2w_bio[i] for i in gold_seq[:l]]
char_seq = [i2w_char[i] for i in input_seq[:l]]
pred_kvpair = extract_kvpairs_in_bio(pred_seq, char_seq)
gold_kvpair = extract_kvpairs_in_bio(gold_seq, char_seq)
print('预测结果:')
print('preds_kvpair', preds_kvpair)
print('gold_kvpair', gold_kvpair)
preds_kvpair.append(pred_kvpair)
golds_kvpair.append(gold_kvpair)
if data_processor.end_flag:
data_processor.refresh()
break
batches_sample += 1
if (max_batches is not None) and (batches_sample >= max_batches):
break
p, r, f1 = cal_f1_score(preds_kvpair, golds_kvpair)
logger.info("Valid Samples: {}".format(len(preds_kvpair)))
logger.info("Valid P/R/F1: {} / {} / {}".format(round(p*100, 2), round(r*100, 2), round(f1*100, 2)))
return (p, r, f1)
assert(params_dict['batch_size'] > 0)
# Start Itearting through
epoch_no = 0
for idx, data in enumerate(train_set):
train_output = train_computation(data)
predictions = train_output['logits']
label_batch = train_output['labels']
niter = idx + 1
# Print training loss and F-1 and EM scores after every 20 iterations
if (idx % 20 == 0):
print('iteration = {}, train loss = {}'.format(
niter, train_output['batch_cost']))
f1_score_int, em_score_int = cal_f1_score(
params_dict, label_batch, predictions)
print("F1_Score and EM_score are", f1_score_int, em_score_int)
divide_val = math.ceil(
len(dev['para']['data']) / params_dict['batch_size'])
if niter % val_frequency == 0:
print('Epoch done:', epoch_no)
epoch_no += 1
f1_score_req = 0
em_score_req = 0
# Compute validation scores
for idx_val, data_val in enumerate(valid_set):
eval_output = valid_computation(data_val)
predictions_val = eval_output['logits']
assert (params_dict['batch_size'] > 0)
# Start Itearting through
epoch_no = 0
for idx, data in enumerate(train_set):
train_output = train_computation(data)
predictions = train_output['logits']
label_batch = train_output['labels']
niter = idx + 1
# Print training loss and F-1 and EM scores after every 20 iterations
if (idx % 20 == 0):
print('iteration = {}, train loss = {}'.format(
niter, train_output['batch_cost']))
f1_score_int, em_score_int = cal_f1_score(params_dict, label_batch,
predictions)
print("F1_Score and EM_score are", f1_score_int, em_score_int)
divide_val = math.ceil(
len(dev['para']['data']) / params_dict['batch_size'])
if niter % val_frequency == 0:
print('Epoch done:', epoch_no)
epoch_no += 1
f1_score_req = 0
em_score_req = 0
# Compute validation scores
for idx_val, data_val in enumerate(valid_set):
eval_output = valid_computation(data_val)
predictions_val = eval_output['logits']