Classifying CIFAR-10 with XLA

This tutorial trains a TensorFlow model to classify the CIFAR-10 dataset, and we compile it using XLA.

Load and normalize the dataset using the Keras API:



In [0]:

    
import tensorflow as tf

# Check that GPU is available: cf. https://colab.research.google.com/notebooks/gpu.ipynb
assert(tf.test.is_gpu_available())

tf.keras.backend.clear_session()
tf.config.optimizer.set_jit(False) # Start with XLA disabled.

def load_data():
  (x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()
  x_train = x_train.astype('float32') / 256
  x_test = x_test.astype('float32') / 256

  # Convert class vectors to binary class matrices.
  y_train = tf.keras.utils.to_categorical(y_train, num_classes=10)
  y_test = tf.keras.utils.to_categorical(y_test, num_classes=10)
  return ((x_train, y_train), (x_test, y_test))

(x_train, y_train), (x_test, y_test) = load_data()

We define the model, adapted from the Keras CIFAR-10 example:



In [0]:

    
def generate_model():
  return tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(32, (3, 3), padding='same', input_shape=x_train.shape[1:]),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.Conv2D(32, (3, 3)),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),
    tf.keras.layers.Dropout(0.25),

    tf.keras.layers.Conv2D(64, (3, 3), padding='same'),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.Conv2D(64, (3, 3)),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),
    tf.keras.layers.Dropout(0.25),

    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(512),
    tf.keras.layers.Activation('relu'),
    tf.keras.layers.Dropout(0.5),
    tf.keras.layers.Dense(10),
    tf.keras.layers.Activation('softmax')
  ])

model = generate_model()

We train the model using the RMSprop optimizer:



In [0]:

    
def compile_model(model):
  opt = tf.keras.optimizers.RMSprop(lr=0.0001, decay=1e-6)
  model.compile(loss='categorical_crossentropy',
                optimizer=opt,
                metrics=['accuracy'])
  return model

model = compile_model(model)

def train_model(model, x_train, y_train, x_test, y_test, epochs=25):
  model.fit(x_train, y_train, batch_size=256, epochs=epochs, validation_data=(x_test, y_test), shuffle=True)

def warmup(model, x_train, y_train, x_test, y_test):
  # Warm up the JIT, we do not wish to measure the compilation time.
  initial_weights = model.get_weights()
  train_model(model, x_train, y_train, x_test, y_test, epochs=1)
  model.set_weights(initial_weights)

warmup(model, x_train, y_train, x_test, y_test)
%time train_model(model, x_train, y_train, x_test, y_test)

scores = model.evaluate(x_test, y_test, verbose=1)
print('Test loss:', scores[0])
print('Test accuracy:', scores[1])

Now let's train the model again, using the XLA compiler. To enable the compiler in the middle of the application, we need to reset the Keras session.



In [0]:

    
# We need to clear the session to enable JIT in the middle of the program.
tf.keras.backend.clear_session()
tf.config.optimizer.set_jit(True) # Enable XLA.
model = compile_model(generate_model())
(x_train, y_train), (x_test, y_test) = load_data()

warmup(model, x_train, y_train, x_test, y_test)
%time train_model(model, x_train, y_train, x_test, y_test)

On a machine with a Titan V GPU and an Intel Xeon E5-2690 CPU the speed up is ~1.17x.