TensorFlow

TensorFlow is Google’s production ML platform, covering the full pipeline from research to deployment. TensorFlow Serving, TensorFlow Lite, and TensorFlow.js handle server, mobile, and browser inference respectively. Since TF2, Keras is TensorFlow’s high-level API, making model definition intuitive while TensorFlow handles the underlying computation graph and hardware acceleration. TensorFlow’s production tooling (TFX, TF Serving, TFLite) remains the most mature in the ecosystem.

Quick start

pip install tensorflow

import tensorflow as tf
from tensorflow import keras

# Build a simple image classifier
model = keras.Sequential([
    keras.layers.Input(shape=(28, 28, 1)),
    keras.layers.Conv2D(32, (3, 3), activation='relu'),
    keras.layers.MaxPooling2D(),
    keras.layers.Flatten(),
    keras.layers.Dense(64, activation='relu'),
    keras.layers.Dense(10, activation='softmax'),
])

model.compile(
    optimizer='adam',
    loss='sparse_categorical_crossentropy',
    metrics=['accuracy'],
)

# Load MNIST dataset
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0
x_train = x_train[..., tf.newaxis]
x_test = x_test[..., tf.newaxis]

model.fit(x_train, y_train, epochs=5, validation_data=(x_test, y_test))

# Save the model
model.save('mnist_model.keras')

When to use

TensorFlow is the right choice when production deployment is the primary concern — TF Serving, TFLite for mobile, and TFX pipelines are unmatched. Google’s TPU infrastructure also runs TensorFlow natively. For research and experimentation, PyTorch has become the dominant choice due to its Pythonic API and HuggingFace ecosystem. If you’re primarily using pretrained models (LLMs, vision transformers), PyTorch + HuggingFace Transformers is more practical.