Spaces:
Sleeping
Sleeping
File size: 6,856 Bytes
c130734 |
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 |
# Copyright 2024 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""A converter from a tf1 ALBERT encoder checkpoint to a tf2 encoder checkpoint.
The conversion will yield an object-oriented checkpoint that can be used
to restore an AlbertEncoder object.
"""
import os
from absl import app
from absl import flags
import tensorflow as tf, tf_keras
from official.legacy.albert import configs
from official.modeling import tf_utils
from official.nlp.modeling import models
from official.nlp.modeling import networks
from official.nlp.tools import tf1_bert_checkpoint_converter_lib
FLAGS = flags.FLAGS
flags.DEFINE_string("albert_config_file", None,
"Albert configuration file to define core bert layers.")
flags.DEFINE_string(
"checkpoint_to_convert", None,
"Initial checkpoint from a pretrained BERT model core (that is, only the "
"BertModel, with no task heads.)")
flags.DEFINE_string("converted_checkpoint_path", None,
"Name for the created object-based V2 checkpoint.")
flags.DEFINE_string("checkpoint_model_name", "encoder",
"The name of the model when saving the checkpoint, i.e., "
"the checkpoint will be saved using: "
"tf.train.Checkpoint(FLAGS.checkpoint_model_name=model).")
flags.DEFINE_enum(
"converted_model", "encoder", ["encoder", "pretrainer"],
"Whether to convert the checkpoint to a `AlbertEncoder` model or a "
"`BertPretrainerV2` model (with mlm but without classification heads).")
ALBERT_NAME_REPLACEMENTS = (
("bert/encoder/", ""),
("bert/", ""),
("embeddings/word_embeddings", "word_embeddings/embeddings"),
("embeddings/position_embeddings", "position_embedding/embeddings"),
("embeddings/token_type_embeddings", "type_embeddings/embeddings"),
("embeddings/LayerNorm", "embeddings/layer_norm"),
("embedding_hidden_mapping_in", "embedding_projection"),
("group_0/inner_group_0/", ""),
("attention_1/self", "self_attention"),
("attention_1/output/dense", "self_attention/attention_output"),
("transformer/LayerNorm/", "transformer/self_attention_layer_norm/"),
("ffn_1/intermediate/dense", "intermediate"),
("ffn_1/intermediate/output/dense", "output"),
("transformer/LayerNorm_1/", "transformer/output_layer_norm/"),
("pooler/dense", "pooler_transform"),
("cls/predictions", "bert/cls/predictions"),
("cls/predictions/output_bias", "cls/predictions/output_bias/bias"),
("cls/seq_relationship/output_bias", "predictions/transform/logits/bias"),
("cls/seq_relationship/output_weights",
"predictions/transform/logits/kernel"),
)
def _create_albert_model(cfg):
"""Creates an ALBERT keras core model from BERT configuration.
Args:
cfg: A `AlbertConfig` to create the core model.
Returns:
A keras model.
"""
albert_encoder = networks.AlbertEncoder(
vocab_size=cfg.vocab_size,
hidden_size=cfg.hidden_size,
embedding_width=cfg.embedding_size,
num_layers=cfg.num_hidden_layers,
num_attention_heads=cfg.num_attention_heads,
intermediate_size=cfg.intermediate_size,
activation=tf_utils.get_activation(cfg.hidden_act),
dropout_rate=cfg.hidden_dropout_prob,
attention_dropout_rate=cfg.attention_probs_dropout_prob,
max_sequence_length=cfg.max_position_embeddings,
type_vocab_size=cfg.type_vocab_size,
initializer=tf_keras.initializers.TruncatedNormal(
stddev=cfg.initializer_range))
return albert_encoder
def _create_pretrainer_model(cfg):
"""Creates a pretrainer with AlbertEncoder from ALBERT configuration.
Args:
cfg: A `BertConfig` to create the core model.
Returns:
A BertPretrainerV2 model.
"""
albert_encoder = _create_albert_model(cfg)
pretrainer = models.BertPretrainerV2(
encoder_network=albert_encoder,
mlm_activation=tf_utils.get_activation(cfg.hidden_act),
mlm_initializer=tf_keras.initializers.TruncatedNormal(
stddev=cfg.initializer_range))
# Makes sure masked_lm layer's variables in pretrainer are created.
_ = pretrainer(pretrainer.inputs)
return pretrainer
def convert_checkpoint(bert_config, output_path, v1_checkpoint,
checkpoint_model_name,
converted_model="encoder"):
"""Converts a V1 checkpoint into an OO V2 checkpoint."""
output_dir, _ = os.path.split(output_path)
# Create a temporary V1 name-converted checkpoint in the output directory.
temporary_checkpoint_dir = os.path.join(output_dir, "temp_v1")
temporary_checkpoint = os.path.join(temporary_checkpoint_dir, "ckpt")
tf1_bert_checkpoint_converter_lib.convert(
checkpoint_from_path=v1_checkpoint,
checkpoint_to_path=temporary_checkpoint,
num_heads=bert_config.num_attention_heads,
name_replacements=ALBERT_NAME_REPLACEMENTS,
permutations=tf1_bert_checkpoint_converter_lib.BERT_V2_PERMUTATIONS,
exclude_patterns=["adam", "Adam"])
# Create a V2 checkpoint from the temporary checkpoint.
if converted_model == "encoder":
model = _create_albert_model(bert_config)
elif converted_model == "pretrainer":
model = _create_pretrainer_model(bert_config)
else:
raise ValueError("Unsupported converted_model: %s" % converted_model)
tf1_bert_checkpoint_converter_lib.create_v2_checkpoint(
model, temporary_checkpoint, output_path, checkpoint_model_name)
# Clean up the temporary checkpoint, if it exists.
try:
tf.io.gfile.rmtree(temporary_checkpoint_dir)
except tf.errors.OpError:
# If it doesn't exist, we don't need to clean it up; continue.
pass
def main(_):
output_path = FLAGS.converted_checkpoint_path
v1_checkpoint = FLAGS.checkpoint_to_convert
checkpoint_model_name = FLAGS.checkpoint_model_name
converted_model = FLAGS.converted_model
albert_config = configs.AlbertConfig.from_json_file(FLAGS.albert_config_file)
convert_checkpoint(albert_config, output_path, v1_checkpoint,
checkpoint_model_name,
converted_model=converted_model)
if __name__ == "__main__":
app.run(main)
|