forked from OpenNMT/CTranslate2
-
Notifications
You must be signed in to change notification settings - Fork 0
/
opennmt_tf.py
257 lines (226 loc) · 11 KB
/
opennmt_tf.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
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
import shutil
import os
import six
from ctranslate2.converters import utils
from ctranslate2.converters.converter import Converter
from ctranslate2.specs import common_spec
from ctranslate2.specs import transformer_spec
def load_model(model_path, src_vocab=None, tgt_vocab=None):
"""Loads variables and vocabularies from a TensorFlow checkpoint or SavedModel."""
import tensorflow as tf
def _extract_variables(structure, scope=""):
from tensorflow.python.training.tracking import tracking
variables = {}
if isinstance(structure, tf.Variable):
variables[scope] = structure
elif isinstance(structure, list):
for i, value in enumerate(structure):
variables.update(_extract_variables(value, scope="%s/%d" % (scope, i)))
elif isinstance(structure, tracking.AutoTrackable):
for key, value in six.iteritems(structure.__dict__):
if key.startswith("_") or key == "keras_api":
continue
variables.update(_extract_variables(
value, scope="%s/%s" % (scope, key) if scope else key))
return variables
model_version = 1
tf_version = int(tf.version.VERSION[0])
# Force beam search kernel loading.
if tf_version == 1:
from tensorflow.contrib.seq2seq.python.ops import beam_search_ops
elif tf_version == 2:
from tensorflow_addons.seq2seq import beam_search_decoder
else:
raise ValueError("Unsupported TensorFlow version %d" % tf_version)
if tf.saved_model.contains_saved_model(model_path):
if tf_version == 2:
model_version = 2
imported = tf.saved_model.load(model_path)
variables = {
"model/%s" % scope:variable.numpy()
for scope, variable in six.iteritems(_extract_variables(imported))}
elif tf_version == 1:
config = tf.compat.v1.ConfigProto(device_count={'GPU': 0})
with tf.compat.v1.Graph().as_default():
with tf.compat.v1.Session(config=config) as sess:
meta_graph = tf.compat.v1.saved_model.loader.load(sess, ["serve"], model_path)
variables = sess.run(
{variable.op.name:variable for variable in tf.compat.v1.global_variables()})
assets = sess.run(tf.compat.v1.get_collection(tf.GraphKeys.ASSET_FILEPATHS))
src_vocab = os.path.join(six.b(model_path), b"assets", os.path.basename(assets[0]))
tgt_vocab = os.path.join(six.b(model_path), b"assets", os.path.basename(assets[1]))
else:
if src_vocab is None or tgt_vocab is None:
raise ValueError("vocabularies must be passed as argument when converting checkpoint")
if os.path.isdir(model_path):
checkpoint = tf.train.latest_checkpoint(model_path)
else:
checkpoint = model_path
reader = tf.train.load_checkpoint(checkpoint)
variables = {
name:reader.get_tensor(name)
for name in six.iterkeys(reader.get_variable_to_shape_map())}
if os.path.basename(checkpoint).startswith("ckpt"):
model_version = 2
variables = {
name.replace("/.ATTRIBUTES/VARIABLE_VALUE", ""):value
for name, value in six.iteritems(variables)}
return model_version, variables, src_vocab, tgt_vocab
class OpenNMTTFConverter(Converter):
"""Converts models generated by OpenNMT-tf."""
def __init__(self, model_path, src_vocab=None, tgt_vocab=None):
self._model_path = model_path
self._src_vocab = src_vocab
self._tgt_vocab = tgt_vocab
def _load(self, model_spec):
version, variables, src_vocab, tgt_vocab = load_model(
self._model_path,
src_vocab=self._src_vocab,
tgt_vocab=self._tgt_vocab)
if isinstance(model_spec, transformer_spec.TransformerSpec):
if version == 2:
set_transformer_spec_v2(model_spec, variables)
else:
set_transformer_spec(model_spec, variables)
else:
raise NotImplementedError()
return src_vocab, tgt_vocab
def _save_vocabulary(self, vocab, destination):
shutil.copy(vocab, destination)
def _vocabulary_size(self, vocab):
with open(vocab, "rb") as vocab_file:
num_tokens = 0
for _ in vocab_file:
num_tokens += 1
return num_tokens + 1 # Add OOV token.
def set_transformer_spec_v2(spec, variables):
set_embeddings(
spec.encoder.embeddings, variables, "model/examples_inputter/features_inputter", version=2)
try:
target_embedding_name = set_embeddings(
spec.decoder.embeddings,
variables,
"model/examples_inputter/labels_inputter",
version=2)
except KeyError:
target_embedding_name = set_embeddings(
spec.decoder.embeddings,
variables,
"model/examples_inputter/features_inputter",
version=2)
set_transformer_encoder_v2(spec.encoder, variables, "model/encoder")
set_transformer_decoder_v2(spec.decoder, variables, "model/decoder", target_embedding_name)
def set_transformer_encoder_v2(spec, variables, scope):
set_layer_norm(spec.layer_norm, variables, "%s/layer_norm" % scope)
for i, layer in enumerate(spec.layer):
set_transformer_encoder_layer_v2(layer, variables, "%s/layers/%d" % (scope, i))
def set_transformer_decoder_v2(spec, variables, scope, target_embedding_name):
try:
set_linear(spec.projection, variables, "%s/output_layer" % scope)
except KeyError:
set_linear(
spec.projection,
variables,
"%s/output_layer" % scope,
weight_name=target_embedding_name,
transpose=False)
set_layer_norm(spec.layer_norm, variables, "%s/layer_norm" % scope)
for i, layer in enumerate(spec.layer):
set_transformer_decoder_layer_v2(layer, variables, "%s/layers/%d" % (scope, i))
def set_transformer_encoder_layer_v2(spec, variables, scope):
set_ffn_v2(spec.ffn, variables, "%s/ffn" % scope)
set_multi_head_attention_v2(
spec.self_attention, variables, "%s/self_attention" % scope, self_attention=True)
def set_transformer_decoder_layer_v2(spec, variables, scope):
set_ffn_v2(spec.ffn, variables, "%s/ffn" % scope)
set_multi_head_attention_v2(
spec.self_attention, variables, "%s/self_attention" % scope, self_attention=True)
set_multi_head_attention_v2(
spec.attention, variables, "%s/attention/0" % scope)
def set_ffn_v2(spec, variables, scope):
set_layer_norm(spec.layer_norm, variables, "%s/input_layer_norm" % scope)
set_linear(spec.linear_0, variables, "%s/layer/inner" % scope)
set_linear(spec.linear_1, variables, "%s/layer/outer" % scope)
def set_multi_head_attention_v2(spec, variables, scope, self_attention=False):
set_layer_norm(spec.layer_norm, variables, "%s/input_layer_norm" % scope)
if self_attention:
split_layers = [common_spec.LinearSpec() for _ in range(3)]
set_linear(split_layers[0], variables, "%s/layer/linear_queries" % scope)
set_linear(split_layers[1], variables, "%s/layer/linear_keys" % scope)
set_linear(split_layers[2], variables, "%s/layer/linear_values" % scope)
utils.fuse_linear(spec.linear[0], split_layers)
else:
set_linear(spec.linear[0], variables, "%s/layer/linear_queries" % scope)
split_layers = [common_spec.LinearSpec() for _ in range(2)]
set_linear(split_layers[0], variables, "%s/layer/linear_keys" % scope)
set_linear(split_layers[1], variables, "%s/layer/linear_values" % scope)
utils.fuse_linear(spec.linear[1], split_layers)
set_linear(spec.linear[-1], variables, "%s/layer/linear_output" % scope)
def set_transformer_spec(spec, variables):
set_transformer_encoder(spec.encoder, variables)
set_transformer_decoder(spec.decoder, variables)
def set_transformer_encoder(spec, variables):
set_layer_norm(spec.layer_norm, variables, "transformer/encoder/LayerNorm")
try:
set_embeddings(spec.embeddings, variables, "transformer/encoder")
except KeyError:
# Try shared embeddings scope instead.
set_embeddings(spec.embeddings, variables, "transformer/shared_embeddings")
for i, layer in enumerate(spec.layer):
set_transformer_encoder_layer(layer, variables, "transformer/encoder/layer_%d" % i)
def set_transformer_decoder(spec, variables):
try:
embeddings_name = set_embeddings(spec.embeddings, variables, "transformer/decoder")
except KeyError:
# Try shared embeddings scope instead.
embeddings_name = set_embeddings(spec.embeddings, variables, "transformer/shared_embeddings")
try:
set_linear(spec.projection, variables, "transformer/decoder/dense")
except KeyError:
# Try reusing the target embeddings.
set_linear(
spec.projection,
variables,
"transformer",
weight_name=embeddings_name,
transpose=False)
set_layer_norm(spec.layer_norm, variables, "transformer/decoder/LayerNorm")
for i, layer in enumerate(spec.layer):
set_transformer_decoder_layer(layer, variables, "transformer/decoder/layer_%d" % i)
def set_transformer_encoder_layer(spec, variables, scope):
set_ffn(spec.ffn, variables, "%s/ffn" % scope)
set_multi_head_attention(
spec.self_attention, variables, "%s/multi_head" % scope, self_attention=True)
def set_transformer_decoder_layer(spec, variables, scope):
set_ffn(spec.ffn, variables, "%s/ffn" % scope)
set_multi_head_attention(
spec.self_attention, variables, "%s/masked_multi_head" % scope, self_attention=True)
set_multi_head_attention(spec.attention, variables, "%s/multi_head" % scope)
def set_ffn(spec, variables, scope):
set_layer_norm(spec.layer_norm, variables, "%s/LayerNorm" % scope)
set_linear(spec.linear_0, variables, "%s/conv1d" % scope)
set_linear(spec.linear_1, variables, "%s/conv1d_1" % scope)
def set_multi_head_attention(spec, variables, scope, self_attention=False):
set_layer_norm(spec.layer_norm, variables, "%s/LayerNorm" % scope)
set_linear(spec.linear[0], variables, "%s/conv1d" % scope)
set_linear(spec.linear[1], variables, "%s/conv1d_1" % scope)
if not self_attention:
set_linear(spec.linear[2], variables, "%s/conv1d_2" % scope)
def set_layer_norm(spec, variables, scope):
spec.gamma = variables["%s/gamma" % scope]
spec.beta = variables["%s/beta" % scope]
def set_linear(spec, variables, scope, weight_name=None, transpose=True):
if weight_name is None:
weight_name = "%s/kernel" % scope
spec.weight = variables[weight_name].squeeze()
if transpose:
spec.weight = spec.weight.transpose()
spec.bias = variables["%s/bias" % scope]
def set_embeddings(spec, variables, scope, version=1):
if version == 2:
name = "embedding"
else:
name = "w_embs"
variable_name = "%s/%s" % (scope, name)
spec.weight = variables[variable_name]
return variable_name