Chapter 37: Integrating TensorFlow with Hugging Face

“When TensorFlow meets Hugging Face, models gain both muscle and memory.”

---

Introduction: The Best of Both Worlds

• TensorFlow is an industrial-grade engine. Hugging Face is a model zoo with superpowers. Combine the two and you get:

• Access to thousands of pretrained models (BERT, GPT-2, ViT, T5, etc.)

• Seamless tokenization, loading, and fine-tuning

• Exportable .h5, .pb, or .tflite models for real-world deployment

In this chapter, we’ll walk through how to use Hugging Face Transformers with TensorFlow, including:

• Loading pretrained models

• Tokenizing data

• Fine-tuning for NLP tasks

• Saving & exporting for inference

---

Step 1: Install the Libraries

pip install -q transformers datasets

---

Step 2: Load a Pretrained BERT Model

Let’s classify sentiment using bert-base-uncased.

from transformers import TFAutoModelForSequenceClassification, AutoTokenizer
import tensorflow as tf

model_name = "bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = TFAutoModelForSequenceClassification.from_pretrained(model_name, num_labels=2)

---

Step 3: Prepare Data (e.g., IMDb Sentiment)

from datasets import load_dataset

dataset = load_dataset("imdb")
train_texts = dataset["train"]["text"][:3000]
train_labels = dataset["train"]["label"][:3000]

# Tokenize
train_encodings = tokenizer(train_texts, truncation=True, padding=True, return_tensors='tf')
train_dataset = tf.data.Dataset.from_tensor_slices((
    dict(train_encodings),
    train_labels
)).batch(16)

---

Step 4: Compile and Fine-Tune

optimizer = tf.keras.optimizers.Adam(learning_rate=5e-5)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)

model.compile(optimizer=optimizer, loss=loss, metrics=["accuracy"])
model.fit(train_dataset, epochs=2)

---

Step 5: Save and Export

Save the model (TensorFlow format)

model.save_pretrained("bert-finetuned-imdb")
tokenizer.save_pretrained("bert-finetuned-imdb")

You can now:

• Reload it via TFAutoModelForSequenceClassification.from_pretrained()

• Convert to .tflite using TFLiteConverter

• Deploy on Hugging Face Spaces or use transformers.pipelines for inference

---

Bonus: Quick Prediction with pipeline()

from transformers import pipeline

classifier = pipeline("sentiment-analysis", model="bert-finetuned-imdb", tokenizer="bert-finetuned-imdb")
print(classifier("I loved this movie! So well-acted and emotional."))

---

Summary

In this chapter, you learned:

• How to load and fine-tune Hugging Face models using TensorFlow

• How to tokenize text and structure datasets for training

• How to export models and deploy them in real-world apps

TensorFlow’s maturity in deployment combined with Hugging Face’s innovation in pretrained models gives you the ultimate toolkit — powerful models with plug-and-play simplicity.

Whether you’re building sentiment analyzers, question-answering bots, or vision transformers, this integration helps you stand on the shoulders of AI giants.

---