-
Notifications
You must be signed in to change notification settings - Fork 4
/
run_languagemodel.py
175 lines (156 loc) · 7.75 KB
/
run_languagemodel.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
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
import json
import glob
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import argparse
import numpy as np
import tensorflow
import tensorflow.compat.v1 as tf
import modeling
import time
from encode_swe import SWEEncoder_ja as Encoder
import modeling
# Mesh-tensorflow用の設定
tf.logging.set_verbosity(tf.logging.ERROR)
tf.disable_v2_behavior()
tf.enable_eager_execution()
# プログラム引数
tf.flags.DEFINE_string("model", help="model folder", default="gptsan-backbone-2.8B" )
tf.flags.DEFINE_string('context', help="input context", default='')
tf.flags.DEFINE_string('mask', help="mask sentence to fill context", default='[MASK]')
tf.flags.DEFINE_integer('pos_vector', help="token position to pull internal vector", default=-1)
tf.flags.DEFINE_string('output', help="output json file", default='')
tf.flags.DEFINE_string('vocabulary', help="vocabulary file", default='ja-swe36k.txt')
tf.flags.DEFINE_string('tpu_nodes', help="tpu nodes", default='')
args = tf.flags.FLAGS
def main():
# 引数チェック
assert os.path.isdir(args.model), f'model not found; {args.model}'
assert os.path.isfile(os.path.join(args.model,'parameters.json')), f'parameter file not found in {args.model}'
assert os.path.isfile(os.path.join(args.model,'checkpoint')), f'checkpoint not found in {args.model}'
assert os.path.isfile(args.vocabulary), f'vocabulary file not found; {args.vocabulary}'
assert args.output=='' or not os.path.isfile(args.output), f'file is exists in {args.output}'
if args.tpu_nodes=='':
print('TPU node not foud. Using GPU device.')
# テキストエンコーダー作成
with open(os.path.join(args.model,'parameters.json'), encoding='utf-8') as f:
saved_params = json.loads(f.read())
with open(args.vocabulary, encoding='utf-8') as f:
bpe = f.read().split('\n')
with open('emoji.json', encoding='utf-8') as f:
emoji = json.loads(f.read())
enc = Encoder(bpe, emoji)
# モデル設定
NUM_CTX = saved_params['model_params']['num_contexts']
MODE = saved_params['model_params']['train_mode']
NUM_TOKENS = len(bpe)
SOT_TOKEN = NUM_TOKENS-7
MSK_TOKEN = NUM_TOKENS-6
SEP_TOKEN = NUM_TOKENS-5
NOT_TOKEN = NUM_TOKENS-4
BAG_TOKEN = NUM_TOKENS-3
SEG_TOKEN = NUM_TOKENS-2
EOT_TOKEN = NUM_TOKENS-1
# マルチモーダル（画像など→テキスト）用のベクトル入力パラメーター
TOTAL_LAYERS = saved_params['model_params']['num_switch_layers']+saved_params['model_params']['num_ext_layers']
NUM_HEADERS = saved_params['model_params']['num_header']
NUM_CHANNELS = saved_params['model_params']['num_hidden'] // NUM_HEADERS
EXT_INPUTS = TOTAL_LAYERS*NUM_HEADERS*NUM_CHANNELS*2
# モデル実行モード
assert MODE in ["lm","hybrid"], "invalid mode"
# SEP_TOKENが文章の区切りを意味するので、出力時に変換するための変数
DOT_TOKENS = enc.encode("。｡．.？！?!:：;；")
NL_TOKEN = enc.encode("\n")[0]
LAST_TOKEN = enc.encode("<|byte0|>")[0]-1
TOKEN_IS_DOT_NL = [(t in DOT_TOKENS or t==NL_TOKEN) for t in range(NUM_TOKENS)]
# 続きを生成する場合の、直前の文章
if MODE=="lm":
pre_input = [SOT_TOKEN] + enc.encode(args.context)
connected_inputs = 0 # hybridで入力するトークン列数
else:
if args.mask in args.context:
pre_inps = args.context.split(args.mask)
inp_token = [enc.encode(inp)+[MSK_TOKEN] for inp in pre_inps]
inp_token = sum(inp_token,[])[:-1]
pre_input = [SOT_TOKEN] + inp_token
connected_inputs = len(inp_token) # hybridで入力するトークン列数
else:
inp_token = enc.encode(args.context)
pre_input = [SOT_TOKEN] + inp_token
connected_inputs = len(inp_token) # hybridで入力するトークン列数
# 実行環境設定
if args.tpu_nodes != "":
tpu = tf.distribute.cluster_resolver.TPUClusterResolver(args.tpu_nodes)
tf.config.experimental_connect_to_cluster(tpu)
topology = tf.tpu.experimental.initialize_tpu_system(tpu)
else:
tpu = None
topology = None
saved_params['model_params']['num_pallarelizm'] = min(saved_params['model_params']['num_pallarelizm'],
len(tf.config.experimental.list_physical_devices('GPU')))
# Mesh-tensorflowを使うので、Estimatorでモデルを読み込むので、その関数
def model_fn(features, labels, mode, params):
x = features["x"]
pos_vector = features["pos_vector"]
num_precontext = features["num_precontext"]
model, run = modeling.model(tpu, params, saved_params, False, False, False)
return run(model(x=x, num_precontext=num_precontext, pos_vector=pos_vector))
# Mesh-tensorflowを使うので、Estimatorでデータを読み込むので、その関数
def input_fn(params):
input_size = min(len(pre_input), NUM_CTX) # Transformerへ入力する長さ
def input_gen(): # モデルへの入力を一つずつ返す
input_tokens = pre_input[:input_size] # モデルの最大入力数まで
pos_vector = args.pos_vector if args.pos_vector>=0 else len(input_tokens)-1 # 内部のデータを取り出す位置
yield {"x":[input_tokens+[EOT_TOKEN]*(input_size-len(input_tokens))],
"pos_vector":[[pos_vector]],
"num_precontext":[[connected_inputs]]}, [0]
output_type = {"x":tf.int32,
"pos_vector":tf.int32,
"num_precontext":tf.int32}
output_shape = {"x":[1,input_size],
"pos_vector":[1,1],
"num_precontext":[1,1]}
dataset = tf.data.Dataset.from_generator(input_gen,
output_types=(output_type,tf.int32),
output_shapes=(output_shape,1))
return dataset
# モデルの実行を定義
if tpu is not None:
run_config = tf.estimator.tpu.RunConfig(
cluster=tpu,
master=None,
model_dir=args.model,
tpu_config=tf.estimator.tpu.TPUConfig(
iterations_per_loop=1,
num_cores_per_replica=1,
per_host_input_for_training=tf.estimator.tpu.InputPipelineConfig.BROADCAST))
estimator = tf.estimator.tpu.TPUEstimator(use_tpu=True, model_fn=model_fn, config=run_config, train_batch_size=1, predict_batch_size=1)
else:
estimator = tf.estimator.Estimator(model_fn=model_fn, model_dir=args.model)
# 文章生成を実行
result = list(estimator.predict(input_fn=input_fn))
pred_token, pred_score = [], []
for i in range(len(pre_input)-1):
p = int(result[0]['logits'][i].argmax()) if pre_input[i+1]==MSK_TOKEN else pre_input[i+1]
s = float(result[0]['logits'][i][pre_input[i+1]])
pred_token.append(p)
pred_score.append(s)
print("{OUTPUT TEXTS}")
print(enc.decode(pred_token))
print("{OUTPUT TOKENS}")
print(pred_token)
print("{OUTPUT SCORES}")
print(pred_score)
print("{OUTPUT VECTOR SHAPE}")
print(result[0]['vector'].shape)
if args.output!='':
with open(args.output, "w") as wf:
wf.write(json.dumps({"output_text":enc.decode(pred_token),
"output_tokens":pred_token,
"output_scores":pred_score,
"output_vector":result[0]['vector'].astype(float).tolist(),
"input_text":args.context,
"input_tokens":pre_input,
}))
if __name__ == "__main__":
main()