In [ ]:
#@title 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
#
# https://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.
In this example, we will work through fine-tuning a BERT model using the tensorflow-models PIP package.
The pretrained BERT model this tutorial is based on is also available on TensorFlow Hub, to see how to use it refer to the Hub Appendix
In [ ]:
!pip install -q tf-nightly
!pip install -q tf-models-nightly
In [ ]:
import os
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
import tensorflow_hub as hub
import tensorflow_datasets as tfds
tfds.disable_progress_bar()
from official.modeling import tf_utils
from official import nlp
from official.nlp import bert
# Load the required submodules
import official.nlp.optimization
import official.nlp.bert.bert_models
import official.nlp.bert.configs
import official.nlp.bert.run_classifier
import official.nlp.bert.tokenization
import official.nlp.data.classifier_data_lib
import official.nlp.modeling.losses
import official.nlp.modeling.models
import official.nlp.modeling.networks
This directory contains the configuration, vocabulary, and a pre-trained checkpoint used in this tutorial:
In [ ]:
gs_folder_bert = "gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-12_H-768_A-12"
tf.io.gfile.listdir(gs_folder_bert)
You can get a pre-trained BERT encoder from TensorFlow Hub:
In [ ]:
hub_url_bert = "https://tfhub.dev/tensorflow/bert_en_uncased_L-12_H-768_A-12/2"
For this example we used the GLUE MRPC dataset from TFDS.
This dataset is not set up so that it can be directly fed into the BERT model, so this section also handles the necessary preprocessing.
The Microsoft Research Paraphrase Corpus (Dolan & Brockett, 2005) is a corpus of sentence pairs automatically extracted from online news sources, with human annotations for whether the sentences in the pair are semantically equivalent.
In [ ]:
glue, info = tfds.load('glue/mrpc', with_info=True,
# It's small, load the whole dataset
batch_size=-1)
In [ ]:
list(glue.keys())
The info object describes the dataset and it's features:
In [ ]:
info.features
The two classes are:
In [ ]:
info.features['label'].names
Here is one example from the training set:
In [ ]:
glue_train = glue['train']
for key, value in glue_train.items():
print(f"{key:9s}: {value[0].numpy()}")
To fine tune a pre-trained model you need to be sure that you're using exactly the same tokenization, vocabulary, and index mapping as you used during training.
The BERT tokenizer used in this tutorial is written in pure Python (It's not built out of TensorFlow ops). So you can't just plug it into your model as a keras.layer like you can with preprocessing.TextVectorization.
The following code rebuilds the tokenizer that was used by the base model:
In [ ]:
# Set up tokenizer to generate Tensorflow dataset
tokenizer = bert.tokenization.FullTokenizer(
vocab_file=os.path.join(gs_folder_bert, "vocab.txt"),
do_lower_case=True)
print("Vocab size:", len(tokenizer.vocab))
Tokenize a sentence:
In [ ]:
tokens = tokenizer.tokenize("Hello TensorFlow!")
print(tokens)
ids = tokenizer.convert_tokens_to_ids(tokens)
print(ids)
The section manually preprocessed the dataset into the format expected by the model.
This dataset is small, so preprocessing can be done quickly and easily in memory. For larger datasets the tf_models library includes some tools for preprocessing and re-serializing a dataset. See Appendix: Re-encoding a large dataset for details.
In [ ]:
tokenizer.convert_tokens_to_ids(['[CLS]', '[SEP]'])
Start by encoding all the sentences while appending a [SEP] token, and packing them into ragged-tensors:
In [ ]:
def encode_sentence(s):
tokens = list(tokenizer.tokenize(s.numpy()))
tokens.append('[SEP]')
return tokenizer.convert_tokens_to_ids(tokens)
sentence1 = tf.ragged.constant([
encode_sentence(s) for s in glue_train["sentence1"]])
sentence2 = tf.ragged.constant([
encode_sentence(s) for s in glue_train["sentence2"]])
In [ ]:
print("Sentence1 shape:", sentence1.shape.as_list())
print("Sentence2 shape:", sentence2.shape.as_list())
Now prepend a [CLS] token, and concatenate the ragged tensors to form a single input_word_ids tensor for each example. RaggedTensor.to_tensor() zero pads to the longest sequence.
In [ ]:
cls = [tokenizer.convert_tokens_to_ids(['[CLS]'])]*sentence1.shape[0]
input_word_ids = tf.concat([cls, sentence1, sentence2], axis=-1)
_ = plt.pcolormesh(input_word_ids.to_tensor())
The model expects two additional inputs:
The mask allows the model to cleanly differentiate between the content and the padding. The mask has the same shape as the input_word_ids, and contains a 1 anywhere the input_word_ids is not padding.
In [ ]:
input_mask = tf.ones_like(input_word_ids).to_tensor()
plt.pcolormesh(input_mask)
The "input type" also has the same shape, but inside the non-padded region, contains a 0 or a 1 indicating which sentence the token is a part of.
In [ ]:
type_cls = tf.zeros_like(cls)
type_s1 = tf.zeros_like(sentence1)
type_s2 = tf.ones_like(sentence2)
input_type_ids = tf.concat([type_cls, type_s1, type_s2], axis=-1).to_tensor()
plt.pcolormesh(input_type_ids)
In [ ]:
def encode_sentence(s, tokenizer):
tokens = list(tokenizer.tokenize(s))
tokens.append('[SEP]')
return tokenizer.convert_tokens_to_ids(tokens)
def bert_encode(glue_dict, tokenizer):
num_examples = len(glue_dict["sentence1"])
sentence1 = tf.ragged.constant([
encode_sentence(s, tokenizer)
for s in np.array(glue_dict["sentence1"])])
sentence2 = tf.ragged.constant([
encode_sentence(s, tokenizer)
for s in np.array(glue_dict["sentence2"])])
cls = [tokenizer.convert_tokens_to_ids(['[CLS]'])]*sentence1.shape[0]
input_word_ids = tf.concat([cls, sentence1, sentence2], axis=-1)
input_mask = tf.ones_like(input_word_ids).to_tensor()
type_cls = tf.zeros_like(cls)
type_s1 = tf.zeros_like(sentence1)
type_s2 = tf.ones_like(sentence2)
input_type_ids = tf.concat(
[type_cls, type_s1, type_s2], axis=-1).to_tensor()
inputs = {
'input_word_ids': input_word_ids.to_tensor(),
'input_mask': input_mask,
'input_type_ids': input_type_ids}
return inputs
In [ ]:
glue_train = bert_encode(glue['train'], tokenizer)
glue_train_labels = glue['train']['label']
glue_validation = bert_encode(glue['validation'], tokenizer)
glue_validation_labels = glue['validation']['label']
glue_test = bert_encode(glue['test'], tokenizer)
glue_test_labels = glue['test']['label']
Each subset of the data has been converted to a dictionary of features, and a set of labels. Each feature in the input dictionary has the same shape, and the number of labels should match:
In [ ]:
for key, value in glue_train.items():
print(f'{key:15s} shape: {value.shape}')
print(f'glue_train_labels shape: {glue_train_labels.shape}')
The first step is to download the configuration for the pre-trained model.
In [ ]:
import json
bert_config_file = os.path.join(gs_folder_bert, "bert_config.json")
config_dict = json.loads(tf.io.gfile.GFile(bert_config_file).read())
bert_config = bert.configs.BertConfig.from_dict(config_dict)
config_dict
The config defines the core BERT Model, which is a Keras model to predict the outputs of num_classes from the inputs with maximum sequence length max_seq_length.
This function returns both the encoder and the classifier.
In [ ]:
bert_classifier, bert_encoder = bert.bert_models.classifier_model(
bert_config, num_labels=2)
The classifier has three inputs and one output:
In [ ]:
tf.keras.utils.plot_model(bert_classifier, show_shapes=True, dpi=48)
Run it on a test batch of data 10 examples from the training set. The output is the logits for the two classes:
In [ ]:
glue_batch = {key: val[:10] for key, val in glue_train.items()}
bert_classifier(
glue_batch, training=True
).numpy()
The TransformerEncoder in the center of the classifier above is the bert_encoder.
Inspecting the encoder, we see its stack of Transformer layers connected to those same three inputs:
In [ ]:
tf.keras.utils.plot_model(bert_encoder, show_shapes=True, dpi=48)
In [ ]:
checkpoint = tf.train.Checkpoint(model=bert_encoder)
checkpoint.restore(
os.path.join(gs_folder_bert, 'bert_model.ckpt')).assert_consumed()
Note: The pretrained TransformerEncoder is also available on TensorFlow Hub. See the Hub appendix for details.
BERT adopts the Adam optimizer with weight decay (aka "AdamW"). It also employs a learning rate schedule that firstly warms up from 0 and then decays to 0.
In [ ]:
# Set up epochs and steps
epochs = 3
batch_size = 32
eval_batch_size = 32
train_data_size = len(glue_train_labels)
steps_per_epoch = int(train_data_size / batch_size)
num_train_steps = steps_per_epoch * epochs
warmup_steps = int(epochs * train_data_size * 0.1 / batch_size)
# creates an optimizer with learning rate schedule
optimizer = nlp.optimization.create_optimizer(
2e-5, num_train_steps=num_train_steps, num_warmup_steps=warmup_steps)
This returns an AdamWeightDecay optimizer with the learning rate schedule set:
In [ ]:
type(optimizer)
To see an example of how to customize the optimizer and it's schedule, see the Optimizer schedule appendix.
The metric is accuracy and we use sparse categorical cross-entropy as loss.
In [ ]:
metrics = [tf.keras.metrics.SparseCategoricalAccuracy('accuracy', dtype=tf.float32)]
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
bert_classifier.compile(
optimizer=optimizer,
loss=loss,
metrics=metrics)
bert_classifier.fit(
glue_train, glue_train_labels,
validation_data=(glue_validation, glue_validation_labels),
batch_size=32,
epochs=epochs)
Now run the fine-tuned model on a custom example to see that it works.
Start by encoding some sentence pairs:
In [ ]:
my_examples = bert_encode(
glue_dict = {
'sentence1':[
'The rain in Spain falls mainly on the plain.',
'Look I fine tuned BERT.'],
'sentence2':[
'It mostly rains on the flat lands of Spain.',
'Is it working? This does not match.']
},
tokenizer=tokenizer)
The model should report class 1 "match" for the first example and class 0 "no-match" for the second:
In [ ]:
result = bert_classifier(my_examples, training=False)
result = tf.argmax(result).numpy()
result
In [ ]:
np.array(info.features['label'].names)[result]
In [ ]:
export_dir='./saved_model'
tf.saved_model.save(bert_classifier, export_dir=export_dir)
In [ ]:
reloaded = tf.saved_model.load(export_dir)
reloaded_result = reloaded([my_examples['input_word_ids'],
my_examples['input_mask'],
my_examples['input_type_ids']], training=False)
original_result = bert_classifier(my_examples, training=False)
# The results are (nearly) identical:
print(original_result.numpy())
print()
print(reloaded_result.numpy())
This tutorial you re-encoded the dataset in memory, for clarity.
This was only possible because glue/mrpc is a very small dataset. To deal with larger datasets tf_models library includes some tools for processing and re-encoding a dataset for efficient training.
The first step is to describe which features of the dataset should be transformed:
In [ ]:
processor = nlp.data.classifier_data_lib.TfdsProcessor(
tfds_params="dataset=glue/mrpc,text_key=sentence1,text_b_key=sentence2",
process_text_fn=bert.tokenization.convert_to_unicode)
Then apply the transformation to generate new TFRecord files.
In [ ]:
# Set up output of training and evaluation Tensorflow dataset
train_data_output_path="./mrpc_train.tf_record"
eval_data_output_path="./mrpc_eval.tf_record"
max_seq_length = 128
batch_size = 32
eval_batch_size = 32
# Generate and save training data into a tf record file
input_meta_data = (
nlp.data.classifier_data_lib.generate_tf_record_from_data_file(
processor=processor,
data_dir=None, # It is `None` because data is from tfds, not local dir.
tokenizer=tokenizer,
train_data_output_path=train_data_output_path,
eval_data_output_path=eval_data_output_path,
max_seq_length=max_seq_length))
Finally create tf.data input pipelines from those TFRecord files:
In [ ]:
training_dataset = bert.run_classifier.get_dataset_fn(
train_data_output_path,
max_seq_length,
batch_size,
is_training=True)()
evaluation_dataset = bert.run_classifier.get_dataset_fn(
eval_data_output_path,
max_seq_length,
eval_batch_size,
is_training=False)()
The resulting tf.data.Datasets return (features, labels) pairs, as expected by keras.Model.fit:
In [ ]:
training_dataset.element_spec
If you need to modify the data loading here is some code to get you started:
In [ ]:
def create_classifier_dataset(file_path, seq_length, batch_size, is_training):
"""Creates input dataset from (tf)records files for train/eval."""
dataset = tf.data.TFRecordDataset(file_path)
if is_training:
dataset = dataset.shuffle(100)
dataset = dataset.repeat()
def decode_record(record):
name_to_features = {
'input_ids': tf.io.FixedLenFeature([seq_length], tf.int64),
'input_mask': tf.io.FixedLenFeature([seq_length], tf.int64),
'segment_ids': tf.io.FixedLenFeature([seq_length], tf.int64),
'label_ids': tf.io.FixedLenFeature([], tf.int64),
}
return tf.io.parse_single_example(record, name_to_features)
def _select_data_from_record(record):
x = {
'input_word_ids': record['input_ids'],
'input_mask': record['input_mask'],
'input_type_ids': record['segment_ids']
}
y = record['label_ids']
return (x, y)
dataset = dataset.map(decode_record,
num_parallel_calls=tf.data.experimental.AUTOTUNE)
dataset = dataset.map(
_select_data_from_record,
num_parallel_calls=tf.data.experimental.AUTOTUNE)
dataset = dataset.batch(batch_size, drop_remainder=is_training)
dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
return dataset
In [ ]:
# Set up batch sizes
batch_size = 32
eval_batch_size = 32
# Return Tensorflow dataset
training_dataset = create_classifier_dataset(
train_data_output_path,
input_meta_data['max_seq_length'],
batch_size,
is_training=True)
evaluation_dataset = create_classifier_dataset(
eval_data_output_path,
input_meta_data['max_seq_length'],
eval_batch_size,
is_training=False)
In [ ]:
training_dataset.element_spec
You can get the BERT model off the shelf from TFHub. It would not be hard to add a classification head on top of this hub.KerasLayer
In [ ]:
# Note: 350MB download.
import tensorflow_hub as hub
In [ ]:
hub_model_name = "bert_en_uncased_L-12_H-768_A-12" #@param ["bert_en_uncased_L-24_H-1024_A-16", "bert_en_wwm_cased_L-24_H-1024_A-16", "bert_en_uncased_L-12_H-768_A-12", "bert_en_wwm_uncased_L-24_H-1024_A-16", "bert_en_cased_L-24_H-1024_A-16", "bert_en_cased_L-12_H-768_A-12", "bert_zh_L-12_H-768_A-12", "bert_multi_cased_L-12_H-768_A-12"]
In [ ]:
hub_encoder = hub.KerasLayer(f"https://tfhub.dev/tensorflow/{hub_model_name}",
trainable=True)
print(f"The Hub encoder has {len(hub_encoder.trainable_variables)} trainable variables")
Test run it on a batch of data:
In [ ]:
result = hub_encoder(
inputs=[glue_train['input_word_ids'][:10],
glue_train['input_mask'][:10],
glue_train['input_type_ids'][:10],],
training=False,
)
print("Pooled output shape:", result[0].shape)
print("Sequence output shape:", result[1].shape)
At this point it would be simple to add a classification head yourself.
The bert_models.classifier_model function can also build a classifier onto the encoder from TensorFlow Hub:
In [ ]:
hub_classifier, hub_encoder = bert.bert_models.classifier_model(
# Caution: Most of `bert_config` is ignored if you pass a hub url.
bert_config=bert_config, hub_module_url=hub_url_bert, num_labels=2)
The one downside to loading this model from TFHub is that the structure of internal keras layers is not restored. So it's more difficult to inspect or modify the model. The TransformerEncoder model is now a single layer:
In [ ]:
tf.keras.utils.plot_model(hub_classifier, show_shapes=True, dpi=64)
In [ ]:
try:
tf.keras.utils.plot_model(hub_encoder, show_shapes=True, dpi=64)
assert False
except Exception as e:
print(f"{type(e).__name__}: {e}")
If you need a more control over the construction of the model it's worth noting that the classifier_model function used earlier is really just a thin wrapper over the nlp.modeling.networks.TransformerEncoder and nlp.modeling.models.BertClassifier classes. Just remember that if you start modifying the architecture it may not be correct or possible to reload the pre-trained checkpoint so you'll need to retrain from scratch.
Build the encoder:
In [ ]:
transformer_config = config_dict.copy()
# You need to rename a few fields to make this work:
transformer_config['attention_dropout_rate'] = transformer_config.pop('attention_probs_dropout_prob')
transformer_config['activation'] = tf_utils.get_activation(transformer_config.pop('hidden_act'))
transformer_config['dropout_rate'] = transformer_config.pop('hidden_dropout_prob')
transformer_config['initializer'] = tf.keras.initializers.TruncatedNormal(
stddev=transformer_config.pop('initializer_range'))
transformer_config['max_sequence_length'] = transformer_config.pop('max_position_embeddings')
transformer_config['num_layers'] = transformer_config.pop('num_hidden_layers')
transformer_config
In [ ]:
manual_encoder = nlp.modeling.networks.TransformerEncoder(**transformer_config)
Restore the weights:
In [ ]:
checkpoint = tf.train.Checkpoint(model=manual_encoder)
checkpoint.restore(
os.path.join(gs_folder_bert, 'bert_model.ckpt')).assert_consumed()
Test run it:
In [ ]:
result = manual_encoder(my_examples, training=True)
print("Sequence output shape:", result[0].shape)
print("Pooled output shape:", result[1].shape)
Wrap it in a classifier:
In [ ]:
manual_classifier = nlp.modeling.models.BertClassifier(
bert_encoder,
num_classes=2,
dropout_rate=transformer_config['dropout_rate'],
initializer=tf.keras.initializers.TruncatedNormal(
stddev=bert_config.initializer_range))
In [ ]:
manual_classifier(my_examples, training=True).numpy()
In [ ]:
optimizer = nlp.optimization.create_optimizer(
2e-5, num_train_steps=num_train_steps, num_warmup_steps=warmup_steps)
That high level wrapper sets up the learning rate schedules and the optimizer.
The base learning rate schedule used here is a linear decay to zero over the training run:
In [ ]:
epochs = 3
batch_size = 32
eval_batch_size = 32
train_data_size = len(glue_train_labels)
steps_per_epoch = int(train_data_size / batch_size)
num_train_steps = steps_per_epoch * epochs
In [ ]:
decay_schedule = tf.keras.optimizers.schedules.PolynomialDecay(
initial_learning_rate=2e-5,
decay_steps=num_train_steps,
end_learning_rate=0)
plt.plot([decay_schedule(n) for n in range(num_train_steps)])
This, in turn is wrapped in a WarmUp schedule that linearly increases the learning rate to the target value over the first 10% of training:
In [ ]:
warmup_steps = num_train_steps * 0.1
warmup_schedule = nlp.optimization.WarmUp(
initial_learning_rate=2e-5,
decay_schedule_fn=decay_schedule,
warmup_steps=warmup_steps)
# The warmup overshoots, because it warms up to the `initial_learning_rate`
# following the original implementation. You can set
# `initial_learning_rate=decay_schedule(warmup_steps)` if you don't like the
# overshoot.
plt.plot([warmup_schedule(n) for n in range(num_train_steps)])
Then create the nlp.optimization.AdamWeightDecay using that schedule, configured for the BERT model:
In [ ]:
optimizer = nlp.optimization.AdamWeightDecay(
learning_rate=warmup_schedule,
weight_decay_rate=0.01,
epsilon=1e-6,
exclude_from_weight_decay=['LayerNorm', 'layer_norm', 'bias'])