-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy patheval_roberta_align.py
139 lines (120 loc) · 5.76 KB
/
eval_roberta_align.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
import tensorflow as tf
import xlrd
import json
import numpy as np
import modeling_align
import tokenization_roberta
all_count=0
correct=0
results=[]
paths=["./data/data1.xlsx","./data/data2.xlsx"]
bert_config = modeling_align.BertConfig.from_json_file("./pretrained/config_tf.json")
tokenizer = tokenization_roberta.RobertaTokenizer.from_pretrained(r"./pretrained")
graph=tf.Graph()
with graph.as_default():
input_ids_placehold=tf.placeholder(shape=[None,64],dtype=tf.int32)
input_mask_placehold=tf.placeholder(shape=[None,64],dtype=tf.int32)
align_mask_placehold = tf.placeholder(shape=[None,64,64], dtype=tf.int32)
model = modeling_align.BertModel(
config=bert_config,
is_training=False,
input_ids=input_ids_placehold,
input_mask=input_mask_placehold,
align_mask=align_mask_placehold,
token_type_ids=None,
use_one_hot_embeddings=False)
output_layer = model.get_pooled_output()
hidden_size = output_layer.shape[-1].value
output_weights = tf.get_variable(
"output_weights", [2, hidden_size],
initializer=tf.truncated_normal_initializer(stddev=0.02))
output_bias = tf.get_variable(
"output_bias", [2], initializer=tf.zeros_initializer())
logits = tf.matmul(output_layer, output_weights, transpose_b=True)
logits = tf.nn.bias_add(logits, output_bias)
probabilities = tf.nn.softmax(logits, axis=-1)
saver = tf.train.Saver()
with tf.Session(graph=graph) as sess:
model_file=tf.train.latest_checkpoint("./output/roberta_align/")
saver.restore(sess,model_file)
for path in paths:
data = xlrd.open_workbook(path)
table = data.sheets()[0]
nrows = table.nrows
for j in range(1, nrows):
all_input_ids = None
all_input_masks=None
all_align_masks=None
all_labels=None
all_count += 1
samples = []
sentences = table.row_values(j)[6].split("|")
for sentence in sentences:
sample = []
sample.append(sentence)
sample.append(table.row_values(j)[4])
samples.append(sample)
for i in range(len(samples)):
guid = "%s" % (i)
text_a = samples[i][0].lower()
text_b = samples[i][1].lower()
label = str(int(table.row_values(j)[3] >= 0.5))
pad_token = tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0]
pad_token_segment_id = 0
label_map = {}
for (i, label) in enumerate(["0","1"]):
label_map[label] = i
inputs = tokenizer.encode_plus(
text_a,
text_b,
add_special_tokens=True,
max_length=64,
)
input_ids, token_type_ids = inputs["input_ids"], inputs["token_type_ids"]
text_a_len = token_type_ids.count(0)
text_b_len = len(token_type_ids) - text_a_len
attention_mask = [1] * len(input_ids)
# Zero-pad up to the sequence length.
padding_length = 64 - len(input_ids)
input_ids = input_ids + ([pad_token] * padding_length)
attention_mask = attention_mask + ([0] * padding_length)
token_type_ids = token_type_ids + ([pad_token_segment_id] * padding_length)
align_mask = [[0] * text_a_len + [1] * text_b_len + [0] * padding_length] * text_a_len + [
[1] * text_a_len + [0] * (text_b_len + padding_length)] * text_b_len + [
[0] * len(input_ids)] * padding_length
label_id = label_map[label]
#print(input_ids)
if all_input_ids is None:
all_input_ids=np.array(input_ids,dtype=np.int).reshape([1,64])
else:
all_input_ids=np.append(all_input_ids,np.array(input_ids,dtype=np.int).reshape([1,64]),axis=0)
if all_input_masks is None:
all_input_masks=np.array(attention_mask,dtype=np.int).reshape(1,64)
else:
all_input_masks=np.append(all_input_masks,np.array(attention_mask,dtype=np.int).reshape(1,64),axis=0)
if all_align_masks is None:
all_align_masks = np.array(align_mask, dtype=np.int).reshape([1,64, 64])
else:
all_align_masks = np.append(all_align_masks, np.array(align_mask, dtype=np.int).reshape([1,64, 64]), axis=0)
if all_labels is None:
all_labels=np.array([label_id],dtype=np.int)
else:
all_labels=np.append(all_labels,np.array([label_id],dtype=np.int),axis=0)
#print(all_input_ids.shape)
#print(all_input_masks.shape)
#print(all_labels.shape)
prob=sess.run(probabilities,feed_dict={input_ids_placehold:all_input_ids,input_mask_placehold:all_input_masks,align_mask_placehold:all_align_masks})
prob=np.array(prob)
#print(prob.shape)
output = np.argmax(prob, axis=1)
#print(output)
correct += int(np.any(output == all_labels))
results.append({"answer": table.row_values(j)[4], "ref": table.row_values(j)[6],
"logits": output.tolist(), "label": int(all_labels[0]),
"iscorrect": int(np.any(output == all_labels))})
print(correct)
print(all_count)
print(correct / all_count)
with open("./results_roberta_align.json", "w", encoding="utf8") as fout:
for result in results:
fout.write(json.dumps(result, ensure_ascii=False) + "\n")