How to use Hugging Face Transformers library in Tensorflow for text classification on custom data?

Question

I am trying to do binary text classification on custom data (which is in csv format) using different transformer architectures that Hugging Face 'Transformers' library offers. I am using this Tensorflow blog post as reference.

I am loading the custom dataset into 'tf.data.Dataset' format using the following code:

def get_dataset(file_path, **kwargs):
   dataset = tf.data.experimental.make_csv_dataset(
     file_path,
     batch_size=5, # Artificially small to make examples easier to show.
     na_value="",
     num_epochs=1,
     ignore_errors=True, 
     **kwargs)
   return dataset 

After this when I tried using the 'glue_convert_examples_to_features' method to tokenize as below:

train_dataset = glue_convert_examples_to_features(
                           examples = train_data,
                           tokenizer = tokenizer, 
                           task = None,
                           label_list = ['0', '1'],
                           max_length = 128
                           )

which throws an error "UnboundLocalError: local variable 'processor' referenced before assignment" at:

 if is_tf_dataset:
    example = processor.get_example_from_tensor_dict(example)
    example = processor.tfds_map(example)

In all the examples, I see that they are using the tasks like 'mrpc' etc which are pre-defined and have a glue_processor to handle. Error raises at the 'line 85' in source code.

Can anyone help with solving this issue using with 'custom data' ?

Answer 1

I had the same starting problem.

This Kaggle submission helped me a lot. There you can see how you can tokenize the data according to the chosen pre-trained model:

from transformers import BertTokenizer
from keras.preprocessing.sequence import pad_sequences

bert_model_name = 'bert-base-uncased'

tokenizer = BertTokenizer.from_pretrained(bert_model_name, do_lower_case=True)
MAX_LEN = 128

def tokenize_sentences(sentences, tokenizer, max_seq_len = 128):
    tokenized_sentences = []

    for sentence in tqdm(sentences):
        tokenized_sentence = tokenizer.encode(
                            sentence,                  # Sentence to encode.
                            add_special_tokens = True, # Add '[CLS]' and '[SEP]'
                            max_length = max_seq_len,  # Truncate all sentences.
                    )
        
        tokenized_sentences.append(tokenized_sentence)

    return tokenized_sentences

def create_attention_masks(tokenized_and_padded_sentences):
    attention_masks = []

    for sentence in tokenized_and_padded_sentences:
        att_mask = [int(token_id > 0) for token_id in sentence]
        attention_masks.append(att_mask)

    return np.asarray(attention_masks)

input_ids = tokenize_sentences(df_train['comment_text'], tokenizer, MAX_LEN)
input_ids = pad_sequences(input_ids, maxlen=MAX_LEN, dtype="long", value=0, truncating="post", padding="post")
attention_masks = create_attention_masks(input_ids)

After that you should split ids and masks:

from sklearn.model_selection import train_test_split

labels =  df_train[label_cols].values

train_ids, validation_ids, train_labels, validation_labels = train_test_split(input_ids, labels, random_state=0, test_size=0.1)
train_masks, validation_masks, _, _ = train_test_split(attention_masks, labels, random_state=0, test_size=0.1)

train_size = len(train_inputs)
validation_size = len(validation_inputs)

Furthermore, I looked into the source of glue_convert_examples_to_features. There you can see how a tf.data.dataset compatible with the BERT model can be created. I created a function for this:

def create_dataset(ids, masks, labels):
    def gen():
        for i in range(len(train_ids)):
            yield (
                {
                    "input_ids": ids[i],
                    "attention_mask": masks[i]
                },
                labels[i],
            )

    return tf.data.Dataset.from_generator(
        gen,
        ({"input_ids": tf.int32, "attention_mask": tf.int32}, tf.int64),
        (
            {
                "input_ids": tf.TensorShape([None]),
                "attention_mask": tf.TensorShape([None])
            },
            tf.TensorShape([None]),
        ),
    )

train_dataset = create_dataset(train_ids, train_masks, train_labels)

I then use the dataset like this:

from transformers import TFBertForSequenceClassification, BertConfig

model = TFBertForSequenceClassification.from_pretrained(
    bert_model_name, 
    config=BertConfig.from_pretrained(bert_model_name, num_labels=20)
)

# Prepare training: Compile tf.keras model with optimizer, loss and learning rate schedule
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
loss = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
metric = tf.keras.metrics.CategoricalAccuracy('accuracy')
model.compile(optimizer=optimizer, loss=loss, metrics=[metric])

# Train and evaluate using tf.keras.Model.fit()
history = model.fit(train_dataset, epochs=1, steps_per_epoch=115, validation_data=val_dataset, validation_steps=7)