Spaces:
No application file
No application file
# coding=utf-8 | |
# Copyright 2018 The Google AI Team 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. | |
# Lint as: python2, python3 | |
"""Run masked LM/next sentence masked_lm pre-training for ALBERT.""" | |
from __future__ import absolute_import | |
from __future__ import division | |
from __future__ import print_function | |
import os | |
import time | |
from albert import modeling | |
from albert import optimization | |
from six.moves import range | |
import tensorflow.compat.v1 as tf | |
from tensorflow.contrib import cluster_resolver as contrib_cluster_resolver | |
from tensorflow.contrib import data as contrib_data | |
from tensorflow.contrib import tpu as contrib_tpu | |
flags = tf.flags | |
FLAGS = flags.FLAGS | |
## Required parameters | |
flags.DEFINE_string( | |
"albert_config_file", None, | |
"The config json file corresponding to the pre-trained ALBERT model. " | |
"This specifies the model architecture.") | |
flags.DEFINE_string( | |
"input_file", None, | |
"Input TF example files (can be a glob or comma separated).") | |
flags.DEFINE_string( | |
"output_dir", None, | |
"The output directory where the model checkpoints will be written.") | |
## Other parameters | |
flags.DEFINE_string( | |
"init_checkpoint", None, | |
"Initial checkpoint (usually from a pre-trained ALBERT model).") | |
flags.DEFINE_integer( | |
"max_seq_length", 512, | |
"The maximum total input sequence length after WordPiece tokenization. " | |
"Sequences longer than this will be truncated, and sequences shorter " | |
"than this will be padded. Must match data generation.") | |
flags.DEFINE_integer( | |
"max_predictions_per_seq", 20, | |
"Maximum number of masked LM predictions per sequence. " | |
"Must match data generation.") | |
flags.DEFINE_bool("do_train", True, "Whether to run training.") | |
flags.DEFINE_bool("do_eval", False, "Whether to run eval on the dev set.") | |
flags.DEFINE_integer("train_batch_size", 4096, "Total batch size for training.") | |
flags.DEFINE_integer("eval_batch_size", 64, "Total batch size for eval.") | |
flags.DEFINE_enum("optimizer", "lamb", ["adamw", "lamb"], | |
"The optimizer for training.") | |
flags.DEFINE_float("learning_rate", 0.00176, "The initial learning rate.") | |
flags.DEFINE_float("poly_power", 1.0, "The power of poly decay.") | |
flags.DEFINE_integer("num_train_steps", 125000, "Number of training steps.") | |
flags.DEFINE_integer("num_warmup_steps", 3125, "Number of warmup steps.") | |
flags.DEFINE_integer("start_warmup_step", 0, "The starting step of warmup.") | |
flags.DEFINE_integer("save_checkpoints_steps", 5000, | |
"How often to save the model checkpoint.") | |
flags.DEFINE_integer("keep_checkpoint_max", 5, | |
"How many checkpoints to keep.") | |
flags.DEFINE_integer("iterations_per_loop", 1000, | |
"How many steps to make in each estimator call.") | |
flags.DEFINE_integer("max_eval_steps", 100, "Maximum number of eval steps.") | |
flags.DEFINE_bool("use_tpu", False, "Whether to use TPU or GPU/CPU.") | |
flags.DEFINE_bool("init_from_group0", False, "Whether to initialize" | |
"parameters of other groups from group 0") | |
tf.flags.DEFINE_string( | |
"tpu_name", None, | |
"The Cloud TPU to use for training. This should be either the name " | |
"used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 " | |
"url.") | |
tf.flags.DEFINE_string( | |
"tpu_zone", None, | |
"[Optional] GCE zone where the Cloud TPU is located in. If not " | |
"specified, we will attempt to automatically detect the GCE project from " | |
"metadata.") | |
tf.flags.DEFINE_string( | |
"gcp_project", None, | |
"[Optional] Project name for the Cloud TPU-enabled project. If not " | |
"specified, we will attempt to automatically detect the GCE project from " | |
"metadata.") | |
tf.flags.DEFINE_string("master", None, "[Optional] TensorFlow master URL.") | |
flags.DEFINE_integer( | |
"num_tpu_cores", 8, | |
"Only used if `use_tpu` is True. Total number of TPU cores to use.") | |
flags.DEFINE_float( | |
"masked_lm_budget", 0, | |
"If >0, the ratio of masked ngrams to unmasked ngrams. Default 0," | |
"for offline masking") | |
def model_fn_builder(albert_config, init_checkpoint, learning_rate, | |
num_train_steps, num_warmup_steps, use_tpu, | |
use_one_hot_embeddings, optimizer, poly_power, | |
start_warmup_step): | |
"""Returns `model_fn` closure for TPUEstimator.""" | |
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument | |
"""The `model_fn` for TPUEstimator.""" | |
tf.logging.info("*** Features ***") | |
for name in sorted(features.keys()): | |
tf.logging.info(" name = %s, shape = %s" % (name, features[name].shape)) | |
input_ids = features["input_ids"] | |
input_mask = features["input_mask"] | |
segment_ids = features["segment_ids"] | |
masked_lm_positions = features["masked_lm_positions"] | |
masked_lm_ids = features["masked_lm_ids"] | |
masked_lm_weights = features["masked_lm_weights"] | |
# Note: We keep this feature name `next_sentence_labels` to be compatible | |
# with the original data created by lanzhzh@. However, in the ALBERT case | |
# it does represent sentence_order_labels. | |
sentence_order_labels = features["next_sentence_labels"] | |
is_training = (mode == tf.estimator.ModeKeys.TRAIN) | |
model = modeling.AlbertModel( | |
config=albert_config, | |
is_training=is_training, | |
input_ids=input_ids, | |
input_mask=input_mask, | |
token_type_ids=segment_ids, | |
use_one_hot_embeddings=use_one_hot_embeddings) | |
(masked_lm_loss, masked_lm_example_loss, | |
masked_lm_log_probs) = get_masked_lm_output(albert_config, | |
model.get_sequence_output(), | |
model.get_embedding_table(), | |
masked_lm_positions, | |
masked_lm_ids, | |
masked_lm_weights) | |
# (sentence_order_loss, sentence_order_example_loss, | |
# sentence_order_log_probs) = get_sentence_order_output( | |
# albert_config, model.get_pooled_output(), sentence_order_labels) | |
total_loss = masked_lm_loss # + sentence_order_loss | |
tvars = tf.trainable_variables() | |
initialized_variable_names = {} | |
scaffold_fn = None | |
if init_checkpoint: | |
tf.logging.info("number of hidden group %d to initialize", | |
albert_config.num_hidden_groups) | |
num_of_initialize_group = 1 | |
if FLAGS.init_from_group0: | |
num_of_initialize_group = albert_config.num_hidden_groups | |
if albert_config.net_structure_type > 0: | |
num_of_initialize_group = albert_config.num_hidden_layers | |
(assignment_map, initialized_variable_names | |
) = modeling.get_assignment_map_from_checkpoint( | |
tvars, init_checkpoint, num_of_initialize_group) | |
if use_tpu: | |
def tpu_scaffold(): | |
for gid in range(num_of_initialize_group): | |
tf.logging.info("initialize the %dth layer", gid) | |
tf.logging.info(assignment_map[gid]) | |
tf.train.init_from_checkpoint(init_checkpoint, assignment_map[gid]) | |
return tf.train.Scaffold() | |
scaffold_fn = tpu_scaffold | |
else: | |
for gid in range(num_of_initialize_group): | |
tf.logging.info("initialize the %dth layer", gid) | |
tf.logging.info(assignment_map[gid]) | |
tf.train.init_from_checkpoint(init_checkpoint, assignment_map[gid]) | |
tf.logging.info("**** Trainable Variables ****") | |
for var in tvars: | |
init_string = "" | |
if var.name in initialized_variable_names: | |
init_string = ", *INIT_FROM_CKPT*" | |
tf.logging.info(" name = %s, shape = %s%s", var.name, var.shape, | |
init_string) | |
output_spec = None | |
if mode == tf.estimator.ModeKeys.TRAIN: | |
train_op = optimization.create_optimizer( | |
total_loss, learning_rate, num_train_steps, num_warmup_steps, | |
use_tpu, optimizer, poly_power, start_warmup_step) | |
output_spec = contrib_tpu.TPUEstimatorSpec( | |
mode=mode, | |
loss=total_loss, | |
train_op=train_op, | |
scaffold_fn=scaffold_fn) | |
elif mode == tf.estimator.ModeKeys.EVAL: | |
def metric_fn(*args): | |
"""Computes the loss and accuracy of the model.""" | |
(masked_lm_example_loss, masked_lm_log_probs, masked_lm_ids, | |
masked_lm_weights, sentence_order_example_loss, | |
sentence_order_log_probs, sentence_order_labels) = args[:7] | |
masked_lm_log_probs = tf.reshape(masked_lm_log_probs, | |
[-1, masked_lm_log_probs.shape[-1]]) | |
masked_lm_predictions = tf.argmax( | |
masked_lm_log_probs, axis=-1, output_type=tf.int32) | |
masked_lm_example_loss = tf.reshape(masked_lm_example_loss, [-1]) | |
masked_lm_ids = tf.reshape(masked_lm_ids, [-1]) | |
masked_lm_weights = tf.reshape(masked_lm_weights, [-1]) | |
masked_lm_accuracy = tf.metrics.accuracy( | |
labels=masked_lm_ids, | |
predictions=masked_lm_predictions, | |
weights=masked_lm_weights) | |
masked_lm_mean_loss = tf.metrics.mean( | |
values=masked_lm_example_loss, weights=masked_lm_weights) | |
metrics = { | |
"masked_lm_accuracy": masked_lm_accuracy, | |
"masked_lm_loss": masked_lm_mean_loss, | |
} | |
sentence_order_log_probs = tf.reshape( | |
sentence_order_log_probs, [-1, sentence_order_log_probs.shape[-1]]) | |
sentence_order_predictions = tf.argmax( | |
sentence_order_log_probs, axis=-1, output_type=tf.int32) | |
sentence_order_labels = tf.reshape(sentence_order_labels, [-1]) | |
sentence_order_accuracy = tf.metrics.accuracy( | |
labels=sentence_order_labels, | |
predictions=sentence_order_predictions) | |
sentence_order_mean_loss = tf.metrics.mean( | |
values=sentence_order_example_loss) | |
metrics.update({ | |
"sentence_order_accuracy": sentence_order_accuracy, | |
"sentence_order_loss": sentence_order_mean_loss | |
}) | |
return metrics | |
metric_values = [ | |
masked_lm_example_loss, masked_lm_log_probs, masked_lm_ids, | |
masked_lm_weights, sentence_order_example_loss, | |
sentence_order_log_probs, sentence_order_labels | |
] | |
eval_metrics = (metric_fn, metric_values) | |
output_spec = contrib_tpu.TPUEstimatorSpec( | |
mode=mode, | |
loss=total_loss, | |
eval_metrics=eval_metrics, | |
scaffold_fn=scaffold_fn) | |
else: | |
raise ValueError("Only TRAIN and EVAL modes are supported: %s" % (mode)) | |
return output_spec | |
return model_fn | |
def get_masked_lm_output(albert_config, input_tensor, output_weights, positions, | |
label_ids, label_weights): | |
"""Get loss and log probs for the masked LM.""" | |
input_tensor = gather_indexes(input_tensor, positions) | |
with tf.variable_scope("cls/predictions"): | |
# We apply one more non-linear transformation before the output layer. | |
# This matrix is not used after pre-training. | |
with tf.variable_scope("transform"): | |
input_tensor = tf.layers.dense( | |
input_tensor, | |
units=albert_config.embedding_size, | |
activation=modeling.get_activation(albert_config.hidden_act), | |
kernel_initializer=modeling.create_initializer( | |
albert_config.initializer_range)) | |
input_tensor = modeling.layer_norm(input_tensor) | |
# The output weights are the same as the input embeddings, but there is | |
# an output-only bias for each token. | |
output_bias = tf.get_variable( | |
"output_bias", | |
shape=[albert_config.vocab_size], | |
initializer=tf.zeros_initializer()) | |
logits = tf.matmul(input_tensor, output_weights, transpose_b=True) | |
logits = tf.nn.bias_add(logits, output_bias) | |
log_probs = tf.nn.log_softmax(logits, axis=-1) | |
label_ids = tf.reshape(label_ids, [-1]) | |
label_weights = tf.reshape(label_weights, [-1]) | |
one_hot_labels = tf.one_hot( | |
label_ids, depth=albert_config.vocab_size, dtype=tf.float32) | |
# The `positions` tensor might be zero-padded (if the sequence is too | |
# short to have the maximum number of predictions). The `label_weights` | |
# tensor has a value of 1.0 for every real prediction and 0.0 for the | |
# padding predictions. | |
per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels, axis=[-1]) | |
numerator = tf.reduce_sum(label_weights * per_example_loss) | |
denominator = tf.reduce_sum(label_weights) + 1e-5 | |
loss = numerator / denominator | |
return (loss, per_example_loss, log_probs) | |
def get_sentence_order_output(albert_config, input_tensor, labels): | |
"""Get loss and log probs for the next sentence prediction.""" | |
# Simple binary classification. Note that 0 is "next sentence" and 1 is | |
# "random sentence". This weight matrix is not used after pre-training. | |
with tf.variable_scope("cls/seq_relationship"): | |
output_weights = tf.get_variable( | |
"output_weights", | |
shape=[2, albert_config.hidden_size], | |
initializer=modeling.create_initializer( | |
albert_config.initializer_range)) | |
output_bias = tf.get_variable( | |
"output_bias", shape=[2], initializer=tf.zeros_initializer()) | |
logits = tf.matmul(input_tensor, output_weights, transpose_b=True) | |
logits = tf.nn.bias_add(logits, output_bias) | |
log_probs = tf.nn.log_softmax(logits, axis=-1) | |
labels = tf.reshape(labels, [-1]) | |
one_hot_labels = tf.one_hot(labels, depth=2, dtype=tf.float32) | |
per_example_loss = -tf.reduce_sum(one_hot_labels * log_probs, axis=-1) | |
loss = tf.reduce_mean(per_example_loss) | |
return (loss, per_example_loss, log_probs) | |
def gather_indexes(sequence_tensor, positions): | |
"""Gathers the vectors at the specific positions over a minibatch.""" | |
sequence_shape = modeling.get_shape_list(sequence_tensor, expected_rank=3) | |
batch_size = sequence_shape[0] | |
seq_length = sequence_shape[1] | |
width = sequence_shape[2] | |
flat_offsets = tf.reshape( | |
tf.range(0, batch_size, dtype=tf.int32) * seq_length, [-1, 1]) | |
flat_positions = tf.reshape(positions + flat_offsets, [-1]) | |
flat_sequence_tensor = tf.reshape(sequence_tensor, | |
[batch_size * seq_length, width]) | |
output_tensor = tf.gather(flat_sequence_tensor, flat_positions) | |
return output_tensor | |
def input_fn_builder(input_files, | |
max_seq_length, | |
max_predictions_per_seq, | |
is_training, | |
num_cpu_threads=4): | |
"""Creates an `input_fn` closure to be passed to TPUEstimator.""" | |
def input_fn(params): | |
"""The actual input function.""" | |
batch_size = params["batch_size"] | |
name_to_features = { | |
"input_ids": tf.FixedLenFeature([max_seq_length], tf.int64), | |
"input_mask": tf.FixedLenFeature([max_seq_length], tf.int64), | |
"segment_ids": tf.FixedLenFeature([max_seq_length], tf.int64), | |
# Note: We keep this feature name `next_sentence_labels` to be | |
# compatible with the original data created by lanzhzh@. However, in | |
# the ALBERT case it does represent sentence_order_labels. | |
"next_sentence_labels": tf.FixedLenFeature([1], tf.int64), | |
} | |
if FLAGS.masked_lm_budget: | |
name_to_features.update({ | |
"token_boundary": | |
tf.FixedLenFeature([max_seq_length], tf.int64)}) | |
else: | |
name_to_features.update({ | |
"masked_lm_positions": | |
tf.FixedLenFeature([max_predictions_per_seq], tf.int64), | |
"masked_lm_ids": | |
tf.FixedLenFeature([max_predictions_per_seq], tf.int64), | |
"masked_lm_weights": | |
tf.FixedLenFeature([max_predictions_per_seq], tf.float32)}) | |
# For training, we want a lot of parallel reading and shuffling. | |
# For eval, we want no shuffling and parallel reading doesn't matter. | |
if is_training: | |
d = tf.data.Dataset.from_tensor_slices(tf.constant(input_files)) | |
d = d.repeat() | |
d = d.shuffle(buffer_size=len(input_files)) | |
# `cycle_length` is the number of parallel files that get read. | |
cycle_length = min(num_cpu_threads, len(input_files)) | |
# `sloppy` mode means that the interleaving is not exact. This adds | |
# even more randomness to the training pipeline. | |
d = d.apply( | |
contrib_data.parallel_interleave( | |
tf.data.TFRecordDataset, | |
sloppy=is_training, | |
cycle_length=cycle_length)) | |
d = d.shuffle(buffer_size=100) | |
else: | |
d = tf.data.TFRecordDataset(input_files) | |
# Since we evaluate for a fixed number of steps we don't want to encounter | |
# out-of-range exceptions. | |
d = d.repeat() | |
# We must `drop_remainder` on training because the TPU requires fixed | |
# size dimensions. For eval, we assume we are evaluating on the CPU or GPU | |
# and we *don't* want to drop the remainder, otherwise we wont cover | |
# every sample. | |
d = d.apply( | |
tf.data.experimental.map_and_batch_with_legacy_function( | |
lambda record: _decode_record(record, name_to_features), | |
batch_size=batch_size, | |
num_parallel_batches=num_cpu_threads, | |
drop_remainder=True)) | |
tf.logging.info(d) | |
return d | |
return input_fn | |
def _decode_record(record, name_to_features): | |
"""Decodes a record to a TensorFlow example.""" | |
example = tf.parse_single_example(record, name_to_features) | |
# tf.Example only supports tf.int64, but the TPU only supports tf.int32. | |
# So cast all int64 to int32. | |
for name in list(example.keys()): | |
t = example[name] | |
if t.dtype == tf.int64: | |
t = tf.to_int32(t) | |
example[name] = t | |
return example | |
def main(_): | |
tf.logging.set_verbosity(tf.logging.INFO) | |
if not FLAGS.do_train and not FLAGS.do_eval: | |
raise ValueError("At least one of `do_train` or `do_eval` must be True.") | |
albert_config = modeling.AlbertConfig.from_json_file(FLAGS.albert_config_file) | |
tf.gfile.MakeDirs(FLAGS.output_dir) | |
input_files = [] | |
for input_pattern in FLAGS.input_file.split(","): | |
input_files.extend(tf.gfile.Glob(input_pattern)) | |
tf.logging.info("*** Input Files ***") | |
for input_file in input_files: | |
tf.logging.info(" %s" % input_file) | |
tpu_cluster_resolver = None | |
if FLAGS.use_tpu and FLAGS.tpu_name: | |
tpu_cluster_resolver = contrib_cluster_resolver.TPUClusterResolver( | |
FLAGS.tpu_name, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project) | |
is_per_host = contrib_tpu.InputPipelineConfig.PER_HOST_V2 | |
run_config = contrib_tpu.RunConfig( | |
cluster=tpu_cluster_resolver, | |
master=FLAGS.master, | |
model_dir=FLAGS.output_dir, | |
save_checkpoints_steps=FLAGS.save_checkpoints_steps, | |
keep_checkpoint_max=FLAGS.keep_checkpoint_max, | |
tpu_config=contrib_tpu.TPUConfig( | |
iterations_per_loop=FLAGS.iterations_per_loop, | |
num_shards=FLAGS.num_tpu_cores, | |
per_host_input_for_training=is_per_host)) | |
model_fn = model_fn_builder( | |
albert_config=albert_config, | |
init_checkpoint=FLAGS.init_checkpoint, | |
learning_rate=FLAGS.learning_rate, | |
num_train_steps=FLAGS.num_train_steps, | |
num_warmup_steps=FLAGS.num_warmup_steps, | |
use_tpu=FLAGS.use_tpu, | |
use_one_hot_embeddings=FLAGS.use_tpu, | |
optimizer=FLAGS.optimizer, | |
poly_power=FLAGS.poly_power, | |
start_warmup_step=FLAGS.start_warmup_step) | |
# If TPU is not available, this will fall back to normal Estimator on CPU | |
# or GPU. | |
estimator = contrib_tpu.TPUEstimator( | |
use_tpu=FLAGS.use_tpu, | |
model_fn=model_fn, | |
config=run_config, | |
train_batch_size=FLAGS.train_batch_size, | |
eval_batch_size=FLAGS.eval_batch_size) | |
if FLAGS.do_train: | |
tf.logging.info("***** Running training *****") | |
tf.logging.info(" Batch size = %d", FLAGS.train_batch_size) | |
train_input_fn = input_fn_builder( | |
input_files=input_files, | |
max_seq_length=FLAGS.max_seq_length, | |
max_predictions_per_seq=FLAGS.max_predictions_per_seq, | |
is_training=True) | |
estimator.train(input_fn=train_input_fn, max_steps=FLAGS.num_train_steps) | |
if FLAGS.do_eval: | |
tf.logging.info("***** Running evaluation *****") | |
tf.logging.info(" Batch size = %d", FLAGS.eval_batch_size) | |
global_step = -1 | |
output_eval_file = os.path.join(FLAGS.output_dir, "eval_results.txt") | |
writer = tf.gfile.GFile(output_eval_file, "w") | |
eval_input_fn = input_fn_builder( | |
input_files=input_files, | |
max_seq_length=FLAGS.max_seq_length, | |
max_predictions_per_seq=FLAGS.max_predictions_per_seq, | |
is_training=False) | |
best_perf = 0 | |
key_name = "masked_lm_accuracy" | |
while global_step < FLAGS.num_train_steps: | |
if estimator.latest_checkpoint() is None: | |
tf.logging.info("No checkpoint found yet. Sleeping.") | |
time.sleep(1) | |
else: | |
result = estimator.evaluate( | |
input_fn=eval_input_fn, steps=FLAGS.max_eval_steps) | |
global_step = result["global_step"] | |
tf.logging.info("***** Eval results *****") | |
checkpoint_path = estimator.latest_checkpoint() | |
for key in sorted(result.keys()): | |
tf.logging.info(" %s = %s", key, str(result[key])) | |
writer.write("%s = %s\n" % (key, str(result[key]))) | |
if result[key_name] > best_perf: | |
best_perf = result[key_name] | |
for ext in ["meta", "data-00000-of-00001", "index"]: | |
src_ckpt = checkpoint_path + ".{}".format(ext) | |
tgt_ckpt = checkpoint_path.rsplit( | |
"-", 1)[0] + "-best.{}".format(ext) | |
tf.logging.info("saving {} to {}".format(src_ckpt, tgt_ckpt)) | |
tf.gfile.Copy(src_ckpt, tgt_ckpt, overwrite=True) | |
writer.write("saved {} to {}\n".format(src_ckpt, tgt_ckpt)) | |
if __name__ == "__main__": | |
flags.mark_flag_as_required("input_file") | |
flags.mark_flag_as_required("albert_config_file") | |
flags.mark_flag_as_required("output_dir") | |
tf.app.run() | |