Spaces:
Sleeping
Sleeping
Pradeep Kumar
commited on
Delete tf1_bert_checkpoint_converter_lib.py
Browse files
tf1_bert_checkpoint_converter_lib.py
DELETED
@@ -1,201 +0,0 @@
|
|
1 |
-
# Copyright 2024 The TensorFlow Authors. All Rights Reserved.
|
2 |
-
#
|
3 |
-
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
-
# you may not use this file except in compliance with the License.
|
5 |
-
# You may obtain a copy of the License at
|
6 |
-
#
|
7 |
-
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
-
#
|
9 |
-
# Unless required by applicable law or agreed to in writing, software
|
10 |
-
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
-
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
-
# See the License for the specific language governing permissions and
|
13 |
-
# limitations under the License.
|
14 |
-
|
15 |
-
r"""Convert checkpoints created by Estimator (tf1) to be Keras compatible."""
|
16 |
-
|
17 |
-
import numpy as np
|
18 |
-
import tensorflow.compat.v1 as tf # TF 1.x
|
19 |
-
|
20 |
-
# Mapping between old <=> new names. The source pattern in original variable
|
21 |
-
# name will be replaced by destination pattern.
|
22 |
-
BERT_NAME_REPLACEMENTS = (
|
23 |
-
("bert", "bert_model"),
|
24 |
-
("embeddings/word_embeddings", "word_embeddings/embeddings"),
|
25 |
-
("embeddings/token_type_embeddings",
|
26 |
-
"embedding_postprocessor/type_embeddings"),
|
27 |
-
("embeddings/position_embeddings",
|
28 |
-
"embedding_postprocessor/position_embeddings"),
|
29 |
-
("embeddings/LayerNorm", "embedding_postprocessor/layer_norm"),
|
30 |
-
("attention/self", "self_attention"),
|
31 |
-
("attention/output/dense", "self_attention_output"),
|
32 |
-
("attention/output/LayerNorm", "self_attention_layer_norm"),
|
33 |
-
("intermediate/dense", "intermediate"),
|
34 |
-
("output/dense", "output"),
|
35 |
-
("output/LayerNorm", "output_layer_norm"),
|
36 |
-
("pooler/dense", "pooler_transform"),
|
37 |
-
)
|
38 |
-
|
39 |
-
BERT_V2_NAME_REPLACEMENTS = (
|
40 |
-
("bert/", ""),
|
41 |
-
("encoder", "transformer"),
|
42 |
-
("embeddings/word_embeddings", "word_embeddings/embeddings"),
|
43 |
-
("embeddings/token_type_embeddings", "type_embeddings/embeddings"),
|
44 |
-
("embeddings/position_embeddings", "position_embedding/embeddings"),
|
45 |
-
("embeddings/LayerNorm", "embeddings/layer_norm"),
|
46 |
-
("attention/self", "self_attention"),
|
47 |
-
("attention/output/dense", "self_attention/attention_output"),
|
48 |
-
("attention/output/LayerNorm", "self_attention_layer_norm"),
|
49 |
-
("intermediate/dense", "intermediate"),
|
50 |
-
("output/dense", "output"),
|
51 |
-
("output/LayerNorm", "output_layer_norm"),
|
52 |
-
("pooler/dense", "pooler_transform"),
|
53 |
-
("cls/predictions", "bert/cls/predictions"),
|
54 |
-
("cls/predictions/output_bias", "cls/predictions/output_bias/bias"),
|
55 |
-
("cls/seq_relationship/output_bias", "predictions/transform/logits/bias"),
|
56 |
-
("cls/seq_relationship/output_weights",
|
57 |
-
"predictions/transform/logits/kernel"),
|
58 |
-
)
|
59 |
-
|
60 |
-
BERT_PERMUTATIONS = ()
|
61 |
-
|
62 |
-
BERT_V2_PERMUTATIONS = (("cls/seq_relationship/output_weights", (1, 0)),)
|
63 |
-
|
64 |
-
|
65 |
-
def _bert_name_replacement(var_name, name_replacements):
|
66 |
-
"""Gets the variable name replacement."""
|
67 |
-
for src_pattern, tgt_pattern in name_replacements:
|
68 |
-
if src_pattern in var_name:
|
69 |
-
old_var_name = var_name
|
70 |
-
var_name = var_name.replace(src_pattern, tgt_pattern)
|
71 |
-
tf.logging.info("Converted: %s --> %s", old_var_name, var_name)
|
72 |
-
return var_name
|
73 |
-
|
74 |
-
|
75 |
-
def _has_exclude_patterns(name, exclude_patterns):
|
76 |
-
"""Checks if a string contains substrings that match patterns to exclude."""
|
77 |
-
for p in exclude_patterns:
|
78 |
-
if p in name:
|
79 |
-
return True
|
80 |
-
return False
|
81 |
-
|
82 |
-
|
83 |
-
def _get_permutation(name, permutations):
|
84 |
-
"""Checks whether a variable requires transposition by pattern matching."""
|
85 |
-
for src_pattern, permutation in permutations:
|
86 |
-
if src_pattern in name:
|
87 |
-
tf.logging.info("Permuted: %s --> %s", name, permutation)
|
88 |
-
return permutation
|
89 |
-
|
90 |
-
return None
|
91 |
-
|
92 |
-
|
93 |
-
def _get_new_shape(name, shape, num_heads):
|
94 |
-
"""Checks whether a variable requires reshape by pattern matching."""
|
95 |
-
if "self_attention/attention_output/kernel" in name:
|
96 |
-
return tuple([num_heads, shape[0] // num_heads, shape[1]])
|
97 |
-
if "self_attention/attention_output/bias" in name:
|
98 |
-
return shape
|
99 |
-
|
100 |
-
patterns = [
|
101 |
-
"self_attention/query", "self_attention/value", "self_attention/key"
|
102 |
-
]
|
103 |
-
for pattern in patterns:
|
104 |
-
if pattern in name:
|
105 |
-
if "kernel" in name:
|
106 |
-
return tuple([shape[0], num_heads, shape[1] // num_heads])
|
107 |
-
if "bias" in name:
|
108 |
-
return tuple([num_heads, shape[0] // num_heads])
|
109 |
-
return None
|
110 |
-
|
111 |
-
|
112 |
-
def create_v2_checkpoint(model,
|
113 |
-
src_checkpoint,
|
114 |
-
output_path,
|
115 |
-
checkpoint_model_name="model"):
|
116 |
-
"""Converts a name-based matched TF V1 checkpoint to TF V2 checkpoint."""
|
117 |
-
# Uses streaming-restore in eager model to read V1 name-based checkpoints.
|
118 |
-
model.load_weights(src_checkpoint).assert_existing_objects_matched()
|
119 |
-
if hasattr(model, "checkpoint_items"):
|
120 |
-
checkpoint_items = model.checkpoint_items
|
121 |
-
else:
|
122 |
-
checkpoint_items = {}
|
123 |
-
|
124 |
-
checkpoint_items[checkpoint_model_name] = model
|
125 |
-
checkpoint = tf.train.Checkpoint(**checkpoint_items)
|
126 |
-
checkpoint.save(output_path)
|
127 |
-
|
128 |
-
|
129 |
-
def convert(checkpoint_from_path,
|
130 |
-
checkpoint_to_path,
|
131 |
-
num_heads,
|
132 |
-
name_replacements,
|
133 |
-
permutations,
|
134 |
-
exclude_patterns=None):
|
135 |
-
"""Migrates the names of variables within a checkpoint.
|
136 |
-
|
137 |
-
Args:
|
138 |
-
checkpoint_from_path: Path to source checkpoint to be read in.
|
139 |
-
checkpoint_to_path: Path to checkpoint to be written out.
|
140 |
-
num_heads: The number of heads of the model.
|
141 |
-
name_replacements: A list of tuples of the form (match_str, replace_str)
|
142 |
-
describing variable names to adjust.
|
143 |
-
permutations: A list of tuples of the form (match_str, permutation)
|
144 |
-
describing permutations to apply to given variables. Note that match_str
|
145 |
-
should match the original variable name, not the replaced one.
|
146 |
-
exclude_patterns: A list of string patterns to exclude variables from
|
147 |
-
checkpoint conversion.
|
148 |
-
|
149 |
-
Returns:
|
150 |
-
A dictionary that maps the new variable names to the Variable objects.
|
151 |
-
A dictionary that maps the old variable names to the new variable names.
|
152 |
-
"""
|
153 |
-
with tf.Graph().as_default():
|
154 |
-
tf.logging.info("Reading checkpoint_from_path %s", checkpoint_from_path)
|
155 |
-
reader = tf.train.NewCheckpointReader(checkpoint_from_path)
|
156 |
-
name_shape_map = reader.get_variable_to_shape_map()
|
157 |
-
new_variable_map = {}
|
158 |
-
conversion_map = {}
|
159 |
-
for var_name in name_shape_map:
|
160 |
-
if exclude_patterns and _has_exclude_patterns(var_name, exclude_patterns):
|
161 |
-
continue
|
162 |
-
# Get the original tensor data.
|
163 |
-
tensor = reader.get_tensor(var_name)
|
164 |
-
|
165 |
-
# Look up the new variable name, if any.
|
166 |
-
new_var_name = _bert_name_replacement(var_name, name_replacements)
|
167 |
-
|
168 |
-
# See if we need to reshape the underlying tensor.
|
169 |
-
new_shape = None
|
170 |
-
if num_heads > 0:
|
171 |
-
new_shape = _get_new_shape(new_var_name, tensor.shape, num_heads)
|
172 |
-
if new_shape:
|
173 |
-
tf.logging.info("Veriable %s has a shape change from %s to %s",
|
174 |
-
var_name, tensor.shape, new_shape)
|
175 |
-
tensor = np.reshape(tensor, new_shape)
|
176 |
-
|
177 |
-
# See if we need to permute the underlying tensor.
|
178 |
-
permutation = _get_permutation(var_name, permutations)
|
179 |
-
if permutation:
|
180 |
-
tensor = np.transpose(tensor, permutation)
|
181 |
-
|
182 |
-
# Create a new variable with the possibly-reshaped or transposed tensor.
|
183 |
-
var = tf.Variable(tensor, name=var_name)
|
184 |
-
|
185 |
-
# Save the variable into the new variable map.
|
186 |
-
new_variable_map[new_var_name] = var
|
187 |
-
|
188 |
-
# Keep a list of converter variables for sanity checking.
|
189 |
-
if new_var_name != var_name:
|
190 |
-
conversion_map[var_name] = new_var_name
|
191 |
-
|
192 |
-
saver = tf.train.Saver(new_variable_map)
|
193 |
-
|
194 |
-
with tf.Session() as sess:
|
195 |
-
sess.run(tf.global_variables_initializer())
|
196 |
-
tf.logging.info("Writing checkpoint_to_path %s", checkpoint_to_path)
|
197 |
-
saver.save(sess, checkpoint_to_path, write_meta_graph=False)
|
198 |
-
|
199 |
-
tf.logging.info("Summary:")
|
200 |
-
tf.logging.info(" Converted %d variable name(s).", len(new_variable_map))
|
201 |
-
tf.logging.info(" Converted: %s", str(conversion_map))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|