使用 horovod 和 slurm 扩展 keras 训练

Question

我在 keras 库中有这段代码，用于在 MNIST 数据集上训练 alexnet 模型。

我想在运行 Slurm 作为工作负载管理器和 horovod (https://github.com/uber/horovod) 的集群上扩展训练以进行分布式训练。

代码包含一个定义 alexnet 层的函数和一个加载 MNIST 数据并准备在 ALexNet 模型上进行训练的主函数。 该代码仅包含 keras 代码。

import ....

def alexnet_model(img_shape=(28, 28, 1), n_classes=10, l2_reg=0.):

    alexnet = Sequential()
    alexnet.add(Conv2D(96, (11, 11), input_shape=img_shape,
        padding='same', kernel_regularizer=l2(l2_reg)))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(MaxPooling2D(pool_size=(2, 2)))
    alexnet.add(Conv2D(256, (5, 5), padding='same'))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(MaxPooling2D(pool_size=(2, 2)))
    alexnet.add(ZeroPadding2D((1, 1)))
    alexnet.add(Conv2D(512, (3, 3), padding='same'))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(MaxPooling2D(pool_size=(2, 2)))
    alexnet.add(ZeroPadding2D((1, 1)))
    alexnet.add(Conv2D(1024, (3, 3), padding='same'))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))

    alexnet.add(ZeroPadding2D((1, 1)))
    alexnet.add(Conv2D(1024, (3, 3), padding='same'))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(MaxPooling2D(pool_size=(2, 2)))
    alexnet.add(Flatten())
    alexnet.add(Dense(3072))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(Dropout(0.5))
    alexnet.add(Dense(4096))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(Dropout(0.5))
    alexnet.add(Dense(n_classes))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('softmax'))
    return alexnet

if __name__ == "__main__":


    batch_size = 32
    num_classes = 10
    epochs = 3

    # input image dimensions
    img_rows, img_cols = 28, 28

    # the data, split between train and test sets
    (x_train, y_train), (x_test, y_test) = mnist.load_data()

    if K.image_data_format() == 'channels_first':
        x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
        x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
        input_shape = (1, img_rows, img_cols)
    else:
        x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
        x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
        input_shape = (img_rows, img_cols, 1)

    x_train = x_train.astype('float32')
    x_test = x_test.astype('float32')
    x_train /= 255
    x_test /= 255
    print('x_train shape:', x_train.shape)
    print(x_train.shape[0], 'train samples')
    print(x_test.shape[0], 'test samples')

    # convert class vectors to binary class matrices
    y_train = keras.utils.to_categorical(y_train, num_classes)
    y_test = keras.utils.to_categorical(y_test, num_classes)

    model = alexnet_model()

    model.compile(loss=keras.losses.categorical_crossentropy,
                optimizer=keras.optimizers.Adadelta(),
                metrics=['accuracy'])

    checkpoint = ModelCheckpoint(filepath='alexnet_mnist_checkpoint.hdf5')

    history= model.fit(x_train, y_train,
            batch_size=batch_size,
            epochs=epochs,
            verbose=1,
            validation_data=(x_test, y_test),
            callbacks=[tbCallback, checkpoint])
    score = model.evaluate(x_test, y_test, verbose=0)
}

我想知道我应该添加哪些行来在 Slurm 和 Horovod 上扩展此代码？

Answer 1

链接https://github.com/horovod/horovod中已经提到了需要进行的更改。它基本上归结为时期、学习率、回调、优化器、检查点保存以及一些初始化和配置设置的变化。请在下面找到启用 horovod 的代码版本：

import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D
from keras import backend as K
from keras.regularizers import l2
from keras.layers import BatchNormalization, Activation, ZeroPadding2D
import math
import tensorflow as tf
import horovod.keras as hvd

# Horovod: initialize Horovod.
hvd.init()

# Horovod: pin GPU to be used to process local rank (one GPU per process)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = str(hvd.local_rank())
K.set_session(tf.Session(config=config))

def alexnet_model(img_shape=(28, 28, 1), n_classes=10, l2_reg=0.):

    alexnet = Sequential()
    alexnet.add(Conv2D(96, (11, 11), input_shape=img_shape,
        padding='same', kernel_regularizer=l2(l2_reg)))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(MaxPooling2D(pool_size=(2, 2)))
    alexnet.add(Conv2D(256, (5, 5), padding='same'))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(MaxPooling2D(pool_size=(2, 2)))
    alexnet.add(ZeroPadding2D((1, 1)))
    alexnet.add(Conv2D(512, (3, 3), padding='same'))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(MaxPooling2D(pool_size=(2, 2)))
    alexnet.add(ZeroPadding2D((1, 1)))
    alexnet.add(Conv2D(1024, (3, 3), padding='same'))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))

    alexnet.add(ZeroPadding2D((1, 1)))
    alexnet.add(Conv2D(1024, (3, 3), padding='same'))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(MaxPooling2D(pool_size=(2, 2)))
    alexnet.add(Flatten())
    alexnet.add(Dense(3072))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(Dropout(0.5))
    alexnet.add(Dense(4096))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('relu'))
    alexnet.add(Dropout(0.5))
    alexnet.add(Dense(n_classes))
    alexnet.add(BatchNormalization())
    alexnet.add(Activation('softmax'))
    return alexnet

if __name__ == "__main__":


    batch_size = 32
    num_classes = 10
    epochs = int(math.ceil(3.0 / hvd.size()))

    # input image dimensions
    img_rows, img_cols = 28, 28

    # the data, split between train and test sets
    (x_train, y_train), (x_test, y_test) = mnist.load_data()

    if K.image_data_format() == 'channels_first':
        x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
        x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
        input_shape = (1, img_rows, img_cols)
    else:
        x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
        x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
        input_shape = (img_rows, img_cols, 1)

    x_train = x_train.astype('float32')
    x_test = x_test.astype('float32')
    x_train /= 255
    x_test /= 255
    print('x_train shape:', x_train.shape)
    print(x_train.shape[0], 'train samples')
    print(x_test.shape[0], 'test samples')

    # convert class vectors to binary class matrices
    y_train = keras.utils.to_categorical(y_train, num_classes)
    y_test = keras.utils.to_categorical(y_test, num_classes)

    model = alexnet_model()

    model.compile(loss=keras.losses.categorical_crossentropy,
                optimizer=hvd.DistributedOptimizer(keras.optimizers.Adadelta(1.0 * hvd.size())),
                metrics=['accuracy'])

    callbacks=[hvd.callbacks.BroadcastGlobalVariablesCallback(0)]

# Horovod: save checkpoints only on worker 0 to prevent other workers from corrupting them.
    if hvd.rank() == 0:
        callbacks.append(keras.callbacks.ModelCheckpoint(filepath='alexnet_mnist_checkpoint.hdf5'))

    history= model.fit(x_train, y_train,
            batch_size=batch_size,
            epochs=epochs,
            verbose=1,
            validation_data=(x_test, y_test),
            callbacks=callbacks)
    score = model.evaluate(x_test, y_test, verbose=0)

更改代码后，请使用 mpirun/mpiexec 运行它，以便 horovod 可以有效地与其进程通信。例如：mpirun -hosts <hostname0,hostname1> -np 2 -genv OMP_NUM_THREADS 4 python alexnet_mnist_horovod.py。请注意，您需要在集群上已经安装 OpenMPI/IntelMPI。给出的示例适用于 IntelMPI。如果您使用的是 OpenMPI，请进行相应更改