forked from tensorflow/ranking
-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtfrbert_example.py
271 lines (217 loc) · 9.88 KB
/
tfrbert_example.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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
# Copyright 2020 The TensorFlow Ranking Authors.
#
# 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.
r"""An example of using BERT output for finetuning a TF-Ranking model.
Please download `bert_config_file` and `bert_init_ckpt` from tensorflow models
website: https://github.com/tensorflow/models/tree/master/official/nlp/bert.
Note that those checkpoints are TF 2.x compatible, which are different from the
checkpoints downloaded here: https://github.com/google-research/bert. You may
convert a TF 1.x checkpoint to TF 2.x using `tf2_encoder_checkpoint_converter`
under https://github.com/tensorflow/models/tree/master/official/nlp/bert.
The following command downloads an uncased BERT-base model checkpoint for you:
mkdir /tmp/bert && \
wget https://storage.googleapis.com/cloud-tpu-checkpoints/bert/keras_bert/\
uncased_L-12_H-768_A-12.tar.gz -P /tmp/bert && \
tar -xvf /tmp/bert/uncased_L-12_H-768_A-12.tar.gz -C /tmp/bert/ && \
Then, use the following command to run training and evaluation locally with cpu
or gpu. For GPU, please add `CUDA_VISIBLE_DEVICES=0` and `--config=cuda`. The
example toy data contains 3 lists in train and test respectively. Due to the
large number of BERT parameters, if running into the `out-of-memory` issue,
plese see: https://github.com/google-research/bert#out-of-memory-issues.
BERT_DIR="/tmp/bert/uncased_L-12_H-768_A-12" && \
OUTPUT_DIR="/tmp/tfr/model/" && \
DATA_DIR="tensorflow_ranking/extension/testdata" && \
rm -rf "${OUTPUT_DIR}" && \
bazel build -c opt \
tensorflow_ranking/extension/examples:tfrbert_example_py_binary && \
./bazel-bin/tensorflow_ranking/extension/examples/tfrbert_example_py_binary \
--train_input_pattern=${DATA_DIR}/tfrbert_elwc_train.tfrecord \
--eval_input_pattern=${DATA_DIR}/tfrbert_elwc_test.tfrecord \
--bert_config_file=${BERT_DIR}/bert_config.json \
--bert_init_ckpt=${BERT_DIR}/bert_model.ckpt \
--bert_max_seq_length=128 \
--model_dir="${OUTPUT_DIR}" \
--list_size=3 \
--loss=softmax_loss \
--train_batch_size=8 \
--eval_batch_size=8 \
--learning_rate=1e-5 \
--num_train_steps=50 \
--num_eval_steps=10 \
--checkpoint_secs=120 \
--num_checkpoints=20
You can use TensorBoard to display the training results stored in $OUTPUT_DIR.
Notes:
* Use --alsologtostderr if the output is not printed into screen.
* The training and evaluation data should be stored in TFRecord format.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from absl import flags
import tensorflow as tf
import tensorflow_ranking as tfr
from tensorflow_ranking.extension import tfrbert
flags.DEFINE_bool("local_training", True, "If true, run training locally.")
flags.DEFINE_string("train_input_pattern", "",
"Input file path pattern used for training.")
flags.DEFINE_string("eval_input_pattern", "",
"Input file path pattern used for eval.")
flags.DEFINE_float("learning_rate", 0.005, "Learning rate for the optimizer.")
flags.DEFINE_integer("train_batch_size", 8,
"Number of input records used per batch for training.")
flags.DEFINE_integer("eval_batch_size", 8,
"Number of input records used per batch for eval.")
flags.DEFINE_integer("checkpoint_secs", 120,
"Saves a model checkpoint every checkpoint_secs seconds.")
flags.DEFINE_integer("num_checkpoints", 1000,
"Saves at most num_checkpoints checkpoints in workspace.")
flags.DEFINE_integer(
"num_train_steps", 200000,
"Number of training iterations. Default means continuous training.")
flags.DEFINE_integer("num_eval_steps", 100, "Number of evaluation iterations.")
flags.DEFINE_string(
"loss", "softmax_loss",
"The RankingLossKey deciding the loss function used in training.")
flags.DEFINE_integer("list_size", None, "List size used for training.")
flags.DEFINE_bool("convert_labels_to_binary", False,
"If true, relevance labels are set to either 0 or 1.")
flags.DEFINE_string("model_dir", None, "Output directory for models.")
flags.DEFINE_float("dropout_rate", 0.1, "The dropout rate.")
# The followings are BERT related flags.
flags.DEFINE_string(
"bert_config_file", None,
"The config json file corresponding to the pre-trained BERT model. This "
"specifies the model architecture. Please download the model from "
"the link: https://github.com/google-research/bert")
flags.DEFINE_string(
"bert_init_ckpt", None,
"Initial checkpoint from a pre-trained BERT model. Please download from "
"the link: https://github.com/google-research/bert")
flags.DEFINE_integer(
"bert_max_seq_length", 512,
"The maximum input sequence length (#words) after WordPiece tokenization. "
"Sequences longer than this will be truncated, and sequences shorter "
"than this will be padded.")
flags.DEFINE_integer(
"bert_num_warmup_steps", 10000,
"This is used for adjust learning rate. If global_step < num_warmup_steps, "
"the learning rate will be `global_step/num_warmup_steps * init_lr`. This "
"is implemented in the bert/optimization.py file.")
FLAGS = flags.FLAGS
_SIZE = "example_list_size"
_NETWORK_NAME = "tfrbert"
def context_feature_columns():
"""Returns context feature names to column definitions."""
return {}
def example_feature_columns():
"""Returns example feature names to column definitions.
`input_ids`, `input_mask` and `segment_ids` are derived from query-document
pair sequence: [CLS] all query tokens [SEP] all document tokens [SEP]. The
original tokens are mapped to ids (based on BERT vocabulary) in `input_ids`.
"""
feature_columns = {}
feature_columns.update({
"input_ids":
tf.feature_column.numeric_column(
"input_ids",
shape=(FLAGS.bert_max_seq_length,),
default_value=0,
dtype=tf.int64),
"input_mask":
tf.feature_column.numeric_column(
"input_mask",
shape=(FLAGS.bert_max_seq_length,),
default_value=0,
dtype=tf.int64),
"segment_ids":
tf.feature_column.numeric_column(
"segment_ids",
shape=(FLAGS.bert_max_seq_length,),
default_value=0,
dtype=tf.int64),
})
return feature_columns
def get_estimator(hparams):
"""Create Keras ranking estimator."""
util = tfrbert.TFRBertUtil(
bert_config_file=hparams.get("bert_config_file"),
bert_init_ckpt=hparams.get("bert_init_ckpt"),
bert_max_seq_length=hparams.get("bert_max_seq_length"))
network = tfrbert.TFRBertRankingNetwork(
context_feature_columns=context_feature_columns(),
example_feature_columns=example_feature_columns(),
bert_config_file=hparams.get("bert_config_file"),
bert_max_seq_length=hparams.get("bert_max_seq_length"),
bert_output_dropout=hparams.get("dropout_rate"),
name=_NETWORK_NAME)
loss = tfr.keras.losses.get(
hparams.get("loss"),
reduction=tf.compat.v2.losses.Reduction.SUM_OVER_BATCH_SIZE)
metrics = tfr.keras.metrics.default_keras_metrics()
config = tf.estimator.RunConfig(
model_dir=hparams.get("model_dir"),
keep_checkpoint_max=hparams.get("num_checkpoints"),
save_checkpoints_secs=hparams.get("checkpoint_secs"))
optimizer = util.create_optimizer(
init_lr=hparams.get("learning_rate"),
train_steps=hparams.get("num_train_steps"),
warmup_steps=hparams.get("bert_num_warmup_steps"))
ranker = tfr.keras.model.create_keras_model(
network=network,
loss=loss,
metrics=metrics,
optimizer=optimizer,
size_feature_name=_SIZE)
return tfr.keras.estimator.model_to_estimator(
model=ranker,
model_dir=hparams.get("model_dir"),
config=config,
warm_start_from=util.get_warm_start_settings(exclude=_NETWORK_NAME))
def train_and_eval():
"""Runs the training and evaluation jobs for a BERT ranking model."""
# The below contains a set of common flags for a TF-Ranking model. You need to
# include all of them for adopting the `RankingPipeline`.
hparams = dict(
train_input_pattern=FLAGS.train_input_pattern,
eval_input_pattern=FLAGS.eval_input_pattern,
learning_rate=FLAGS.learning_rate,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
checkpoint_secs=FLAGS.checkpoint_secs,
num_checkpoints=FLAGS.num_checkpoints,
num_train_steps=FLAGS.num_train_steps,
num_eval_steps=FLAGS.num_eval_steps,
loss=FLAGS.loss,
dropout_rate=FLAGS.dropout_rate,
list_size=FLAGS.list_size,
listwise_inference=True, # Only supports `True` in keras Ranking Network.
convert_labels_to_binary=FLAGS.convert_labels_to_binary,
model_dir=FLAGS.model_dir,
bert_config_file=FLAGS.bert_config_file,
bert_init_ckpt=FLAGS.bert_init_ckpt,
bert_max_seq_length=FLAGS.bert_max_seq_length,
bert_num_warmup_steps=FLAGS.bert_num_warmup_steps)
bert_ranking_pipeline = tfr.ext.pipeline.RankingPipeline(
context_feature_columns=context_feature_columns(),
example_feature_columns=example_feature_columns(),
hparams=hparams,
estimator=get_estimator(hparams),
label_feature_name="relevance",
label_feature_type=tf.int64,
size_feature_name=_SIZE)
bert_ranking_pipeline.train_and_eval(local_training=FLAGS.local_training)
def main(_):
train_and_eval()
if __name__ == "__main__":
tf.compat.v1.app.run()