parallel_model.py

import tensorflow as tf
import keras.backend as K
import keras.layers as KL
import keras.models as KM


class ParallelModel(KM.Model):

    def __init__(self, keras_model, gpu_count):
        """Class constructor.
        keras_model: The Keras model to parallelize
        gpu_count: Number of GPUs. Must be > 1
        """
        self.inner_model = keras_model
        self.gpu_count = gpu_count
        merged_outputs = self.make_parallel()
        super(ParallelModel, self).__init__(inputs=self.inner_model.inputs,
                                            outputs=merged_outputs)

    def __getattribute__(self, attrname):
        """Redirect loading and saving methods to the inner model. That's where
        the weights are stored."""
        if 'load' in attrname or 'save' in attrname:
            return getattr(self.inner_model, attrname)
        return super(ParallelModel, self).__getattribute__(attrname)

    def summary(self, *args, **kwargs):
        """Override summary() to display summaries of both, the wrapper
        and inner models."""
        super(ParallelModel, self).summary(*args, **kwargs)
        self.inner_model.summary(*args, **kwargs)

    def make_parallel(self):
        """Creates a new wrapper model that consists of multiple replicas of
        the original model placed on different GPUs.
        """
        # Slice inputs. Slice inputs on the CPU to avoid sending a copy
        # of the full inputs to all GPUs. Saves on bandwidth and memory.
        input_slices = {name: tf.split(x, self.gpu_count)
                        for name, x in zip(self.inner_model.input_names,
                                           self.inner_model.inputs)}

        output_names = self.inner_model.output_names
        outputs_all = []
        for i in range(len(self.inner_model.outputs)):
            outputs_all.append([])

        # Run the model call() on each GPU to place the ops there
        for i in range(self.gpu_count):
            with tf.device('/gpu:%d' % i):
                with tf.name_scope('tower_%d' % i):
                    # Run a slice of inputs through this replica
                    zipped_inputs = zip(self.inner_model.input_names,
                                        self.inner_model.inputs)
                    inputs = [
                        KL.Lambda(lambda s: input_slices[name][i],
                                  output_shape=lambda s: (None,) + s[1:])(tensor)
                        for name, tensor in zipped_inputs]
                    # Create the model replica and get the outputs
                    outputs = self.inner_model(inputs)
                    if not isinstance(outputs, list):
                        outputs = [outputs]
                    # Save the outputs for merging back together later
                    for l, o in enumerate(outputs):
                        outputs_all[l].append(o)

        # Merge outputs on CPU
        with tf.device('/cpu:0'):
            merged = []
            for outputs, name in zip(outputs_all, output_names):
                # Concatenate or average outputs?
                # Outputs usually have a batch dimension and we concatenate
                # across it. If they don't, then the output is likely a loss
                # or a metric value that gets averaged across the batch.
                # Keras expects losses and metrics to be scalars.
                if K.int_shape(outputs[0]) == ():
                    # Average
                    m = KL.Lambda(lambda o: tf.add_n(o) / len(outputs), name=name)(outputs)
                else:
                    # Concatenate
                    m = KL.Concatenate(axis=0, name=name)(outputs)
                merged.append(m)
        return merged


if __name__ == "__main__":
    # Testing code below. It creates a simple model to train on MNIST and
    # tries to run it on 2 GPUs. It saves the graph so it can be viewed
    # in TensorBoard. Run it as:
    #
    # python3 parallel_model.py

    import os
    import numpy as np
    import keras.optimizers
    from keras.datasets import mnist
    from keras.preprocessing.image import ImageDataGenerator

    GPU_COUNT = 2

    # Root directory of the project
    ROOT_DIR = os.path.abspath("../")

    # Directory to save logs and trained model
    MODEL_DIR = os.path.join(ROOT_DIR, "logs")

    def build_model(x_train, num_classes):
        # Reset default graph. Keras leaves old ops in the graph,
        # which are ignored for execution but clutter graph
        # visualization in TensorBoard.
        tf.reset_default_graph()

        inputs = KL.Input(shape=x_train.shape[1:], name="input_image")
        x = KL.Conv2D(32, (3, 3), activation='relu', padding="same",
                      name="conv1")(inputs)
        x = KL.Conv2D(64, (3, 3), activation='relu', padding="same",
                      name="conv2")(x)
        x = KL.MaxPooling2D(pool_size=(2, 2), name="pool1")(x)
        x = KL.Flatten(name="flat1")(x)
        x = KL.Dense(128, activation='relu', name="dense1")(x)
        x = KL.Dense(num_classes, activation='softmax', name="dense2")(x)

        return KM.Model(inputs, x, "digit_classifier_model")

    # Load MNIST Data
    (x_train, y_train), (x_test, y_test) = mnist.load_data()
    x_train = np.expand_dims(x_train, -1).astype('float32') / 255
    x_test = np.expand_dims(x_test, -1).astype('float32') / 255

    print('x_train shape:', x_train.shape)
    print('x_test shape:', x_test.shape)

    # Build data generator and model
    datagen = ImageDataGenerator()
    model = build_model(x_train, 10)

    # Add multi-GPU support.
    model = ParallelModel(model, GPU_COUNT)

    optimizer = keras.optimizers.SGD(lr=0.01, momentum=0.9, clipnorm=5.0)

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

    model.summary()

    # Train
    model.fit_generator(
        datagen.flow(x_train, y_train, batch_size=64),
        steps_per_epoch=50, epochs=10, verbose=1,
        validation_data=(x_test, y_test),
        callbacks=[keras.callbacks.TensorBoard(log_dir=MODEL_DIR,
                                               write_graph=True)]
    )