Created
April 2, 2020 05:08
-
-
Save papapabi/124c6ac406e6bbd1f28df732e953ac6d to your computer and use it in GitHub Desktop.
Using TFBertForSequenceClassification in a custom training loop
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import pandas as pd | |
import tensorflow as tf | |
from collections import namedtuple | |
from typing import List, Tuple | |
from transformers import ( | |
BertConfig, | |
BertTokenizer, | |
TFBertForSequenceClassification | |
) | |
EPOCHS = 3 | |
BATCH_SIZE = 16 | |
TO_FINETUNE = 'bert-case-based' | |
# InputExample is just an intermediary consruct to pair strings with their labels | |
InputExample = namedtuple('InputExample', ['text', 'category_index']) | |
# InputFeatures is just an intermediary construct to easily convert to a tf.data.Dataset | |
InputFeatures = namedtuple('InputFeatures', ['input_ids', 'attention_mask', 'token_type_ids', 'label']) | |
# pd.DataFrame with 'text' and 'category_index' columns as per your example | |
# i'm assuming values in text are str, category_index are int | |
df = pd.read_csv("foo.csv") | |
# Get total number of labels in the df | |
num_labels = df['category_index'].nunique() | |
num_examples = df.shape[0] # row count | |
examples = [] | |
for row in df.itertuples(index=False): | |
examples.append(InputExample(text=row.text, category_index=row.category_index)) | |
config = BertConfig.from_pretrained(TO_FINETUNE, num_labels=num_labels) | |
tokenizer = BertTokenizer.from_pretrained(TO_FINETUNE) | |
def convert_examples_to_tf_dataset( | |
examples: List[Tuple[str, int]], | |
tokenizer, | |
max_length=512, | |
): | |
""" | |
Loads data into a tf.data.Dataset for finetuning a given model. | |
Args: | |
examples: List of tuples representing the examples to be fed | |
tokenizer: Instance of a tokenizer that will tokenize the examples | |
max_length: Maximum string length | |
Returns: | |
a ``tf.data.Dataset`` containing the condensed features of the provided sentences | |
""" | |
features = [] # -> will hold InputFeatures to be converted later | |
for e in examples: | |
# Documentation is really strong for this method, so please take a look at it | |
input_dict = tokenizer.encode_plus( | |
e.text, | |
add_special_tokens=True, | |
max_length=max_length, # truncates if len(s) > max_length | |
return_token_type_ids=True, | |
return_attention_mask=True, | |
pad_to_max_length=True, # pads to the right by default | |
) | |
# input ids = token indices in the tokenizer's internal dict | |
# token_type_ids = binary mask identifying different sequences in the model | |
# attention_mask = binary mask indicating the positions of padded tokens so the model does not attend to them | |
input_ids, token_type_ids, attention_mask = (input_dict["input_ids"], | |
input_dict["token_type_ids"], input_dict['attention_mask']) | |
features.append( | |
InputFeatures( | |
input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, label=e.category_index | |
) | |
) | |
def gen(): | |
for f in features: | |
yield ( | |
{ | |
"input_ids": f.input_ids, | |
"attention_mask": f.attention_mask, | |
"token_type_ids": f.token_type_ids, | |
}, | |
f.label, | |
) | |
return tf.data.Dataset.from_generator( | |
gen, | |
({"input_ids": tf.int32, "attention_mask": tf.int32, "token_type_ids": tf.int32}, tf.int64), | |
( | |
{ | |
"input_ids": tf.TensorShape([None]), | |
"attention_mask": tf.TensorShape([None]), | |
"token_type_ids": tf.TensorShape([None]), | |
}, | |
tf.TensorShape([]), | |
), | |
) | |
# Make the CPU do all data pre-processing steps, not the GPU | |
with tf.device('/cpu:0'): | |
train_data = convert_examples_to_tf_dataset(examples, tokenizer) | |
train_data = train_data.shuffle(buffer_size=num_examples, reshuffle_each_iteration=True) \ | |
.batch(BATCH_SIZE) \ | |
.repeat(-1) | |
config = BertConfig.from_pretrained(TO_FINETUNE) | |
model = TFBertForSequenceClassification.from_pretrained(TO_FINETUNE, config=config) | |
# train_data is then a tf.data.Dataset we can pass to model.fit() | |
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-05, epsilon=1e-08) | |
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True) | |
metric = tf.keras.metrics.SparseCategoricalCrossentropy(name='accuracy') | |
model.compile(optimizer=optimizer, | |
loss=loss, | |
metrics=[metric]) | |
train_steps = num_examples // BATCH_SIZE | |
history = model.fit(train_data, | |
epochs=EPOCHS, | |
steps_per_epoch=train_steps) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I am trying to understand your example of using Bert for sequence classification with a custom example.
My question is fairly simple: what is the need for:
If I don't shuffle the data, I get an error:
So, to me, it seems that shuffling is not only shuffling but doing something else and I don't exactly understand what.