Fine-tuning GPT2 for text-generation with TensorFlow

I’m trying to fine-tune gpt2 with TensorFlow on my apple m1:

Here’s my code, following the guide on the course:

import os
import psutil
import kaggle
import tensorflow as tf
from itertools import chain
from datasets import load_dataset
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.losses import SparseCategoricalCrossentropy
from tensorflow.keras.optimizers.schedules import PolynomialDecay
from transformers import AutoTokenizer, TFAutoModelForCausalLM, DataCollatorWithPadding

CPU_COUNT = psutil.cpu_count()
MODEL_CHECKPOINT = 'distilgpt2'
KAGGLE_DS_DIR = 'kaggle_dataset'

print('Number of CPUs:', CPU_COUNT)
print('Available GPUs:', tf.config.experimental.list_physical_devices('GPU'))

# download dataset if it doesn't exists
if not os.path.exists(KAGGLE_DS_DIR):
        'simiotic/github-code-snippets-development-sample', path=KAGGLE_DS_DIR, unzip=True)

# load raw dataset from sqlite3
raw_dataset = load_dataset('./')
if "validation" not in raw_dataset.keys():
    raw_dataset["validation"] = load_dataset(
    raw_dataset["train"] = load_dataset(

# initiate tokenizer and model on cuda
tokenizer = AutoTokenizer.from_pretrained(MODEL_CHECKPOINT)

tokenizer.pad_token = tokenizer.eos_token
max_seq_length = tokenizer.model_max_length

model = TFAutoModelForCausalLM.from_pretrained(MODEL_CHECKPOINT)
data_collator = DataCollatorWithPadding(tokenizer, return_tensors="tf")


def tokenize_funcion(examples):
    return tokenizer(examples['text'], truncation=True)

def group_texts(examples):
    # Concatenate all texts.
    concatenated_examples = {
        k: list(chain(*examples[k])) for k in examples.keys()}
    total_length = len(concatenated_examples[list(examples.keys())[0]])
    # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
    # customize this part to your needs.
    if total_length >= max_seq_length:
        total_length = (total_length // max_seq_length) * max_seq_length
    # Split by chunks of max_len.
    result = {
        k: [t[i: i + max_seq_length]
            for i in range(0, total_length, max_seq_length)]
        for k, t in concatenated_examples.items()
    return result

# tokenize the raw dataset
tokenized_ds =, batched=True, num_proc=CPU_COUNT, remove_columns=["text"])
tokenized_ds =, batched=True, num_proc=CPU_COUNT)

# convert training dataset to tf dataset
tf_ds = tokenized_ds['train'].to_tf_dataset(
    columns=['input_ids', 'attention_mask', 'token_type_ids'],
    # columns=[col for col in tokenized_ds['train'].features if col != "special_tokens_mask"],

eval_ds = tokenized_ds['validation'].to_tf_dataset(
    columns=['input_ids', 'attention_mask', 'token_type_ids'],
    # columns=[col for col in tokenized_ds['validation'].features if col != "special_tokens_mask"],

num_train_steps = len(tf_ds) * NUM_EPOCHS
lr_scheduler = PolynomialDecay(
opt = Adam(learning_rate=lr_scheduler)

	loss=SparseCategoricalCrossentropy(from_logits=True, reduction=tf.keras.losses.Reduction.NONE),
), validation_data=eval_ds, epochs=NUM_EPOCHS, steps_per_epoch=len(tf_ds) // BATCH_SIZE)

EDIT - Here’s the local loading script:

import os
import sqlite3
import datasets

class SqlDsLoader(datasets.GeneratorBasedBuilder):
    """Code Corpus Dataset."""

    VERSION = datasets.Version("1.0.0")

    def _info(self):
        return datasets.DatasetInfo(
            features=datasets.Features({"text": datasets.Value("string")}),

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""
        input_path = "kaggle_dataset/snippets-dev/snippets-dev.db"
        return [
                gen_kwargs={"filepath": input_path},

    def _generate_examples(self, filepath):
        """Yields examples."""
        con = sqlite3.connect(filepath)
        cur = con.cursor()

        for id, snippet in cur.execute('SELECT id, snippet FROM snippets'):
            yield str(id), {
                "text": snippet,

However, once I add the loss function SparseCategoricalCrossentropy to the model.compile(), I get the following error:

Traceback (most recent call last):
  File "/Users/elonsalfati/devel/metissio/research/", line 115, in <module>, validation_data=eval_ds, epochs=NUM_EPOCHS, steps_per_epoch=len(tf_ds) // BATCH_SIZE)
  File "/opt/homebrew/Caskroom/miniforge/base/envs/nlp/lib/python3.9/site-packages/keras/utils/", line 67, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/opt/homebrew/Caskroom/miniforge/base/envs/nlp/lib/python3.9/site-packages/tensorflow/python/framework/", line 1147, in autograph_handler
    raise e.ag_error_metadata.to_exception(e)
TypeError: in user code:

    File "/opt/homebrew/Caskroom/miniforge/base/envs/nlp/lib/python3.9/site-packages/keras/engine/", line 1021, in train_function  *
        return step_function(self, iterator)
    File "/opt/homebrew/Caskroom/miniforge/base/envs/nlp/lib/python3.9/site-packages/keras/engine/", line 1010, in step_function  **
        outputs =, args=(data,))
    File "/opt/homebrew/Caskroom/miniforge/base/envs/nlp/lib/python3.9/site-packages/keras/engine/", line 1000, in run_step  **
        outputs = model.train_step(data)
    File "/opt/homebrew/Caskroom/miniforge/base/envs/nlp/lib/python3.9/site-packages/transformers/", line 916, in train_step
        self.optimizer.minimize(loss, self.trainable_variables, tape=tape)
    File "/opt/homebrew/Caskroom/miniforge/base/envs/nlp/lib/python3.9/site-packages/keras/optimizer_v2/", line 530, in minimize
        grads_and_vars = self._compute_gradients(
    File "/opt/homebrew/Caskroom/miniforge/base/envs/nlp/lib/python3.9/site-packages/keras/optimizer_v2/", line 583, in _compute_gradients
        grads_and_vars = self._get_gradients(tape, loss, var_list, grad_loss)
    File "/opt/homebrew/Caskroom/miniforge/base/envs/nlp/lib/python3.9/site-packages/keras/optimizer_v2/", line 464, in _get_gradients
        grads = tape.gradient(loss, var_list, grad_loss)

    TypeError: Target should be a list or nested structure of Tensors or Variables to be differentiated, but received None

I’ve tried to look for some examples on how to fine-tune gpt2 with TensorFlow for text generation, but I couldn’t find much. Any suggestions on how to solve this TypeError and what does it mean?

1 Like

Hi! Your code depended on some local loading scripts, so I wasn’t able to run it, but it mostly looks correct, and I’m not totally sure what the bug is. We do provide some notebooks for language modeling (both fine-tuned and from-scratch) here, so you could try starting with those and then adapting them to use your input data instead! notebooks/examples at master · huggingface/notebooks · GitHub

We also specifically cover language modeling for code generation in the course - take a look at Main NLP tasks - Hugging Face Course . There is a link at the top to a Colab notebook that you can try out, and it should be possible to swap in your own data for the data we use there.

I’ve added the local loading script to the question, and I’ll take a look into the notebooks you provided - I wasn’t aware of this repo.

Still, I’m trying to figure out what exactly is the problem to avoid/learn from that. It will be great to understand what went wrong.

I think you need to have the labels in the column list in your dataset

You’ve declared

    columns=['input_ids', 'attention_mask', 'token_type_ids'],

but if I understand your data source right, you only have the text in there. The data should contain labels, which are the tokens shifted left by one. So if you use the following tokenize function:

def tokenize_function(example):
    tokens = tokenizer(example["text"], truncation=True)

    ids = tokens['input_ids']

    return {'input_ids': ids[:,:-1].numpy(), 
            'labels': ids[:,1:].numpy(), 
            'attention_mask': tokens['attention_mask'][:,1:].numpy()}

and you drop the token_type_ids from columns (they are used by BERT, not the GPTs), I’d expect it to work.