tf.keras在 cifar 上训练 AlexNet，数据集过大导致 OOM

Posted 2022-04-16 wuliyttaotao

cifar-10 每张图片的大小为 32×32，而 AlexNet 要求图片的输入是 224×224（也有说 227×227 的，这是 224×224 的图片进行大小为 2 的 zero padding 的结果），所以一种做法是将 cifar-10 数据集的图片 resize 到 224×224。

此时遇到的问题是，cifar-10 resize 到 224×224 时，32G 内存都将无法完全加载所有数据，在归一化那一步（即每个像素点除以 255）就将发生 OOM（out of memory）。

那么此时的做法有：
1）将 resize 作为模型的一部分，如设置一个 layer 来对一个 batch 的图像进行 resize，这样 32×32 的 cifar-10 仍然可以完全加载到内存中；
2）一种通用的方法，每次只加载一部分数据到内存中，其余数据等到需要的时候再加载到内存。

注：本文 AlexNet 结构与 PyTorch 中一致。AlexNet in pytorch/vision

方法 1：加上一个 Lambda 层，对输入图片进行 resize

import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.python.keras import backend as K

K.clear_session()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True  # 不全部占满显存, 按需分配
K.set_session(tf.Session(config=config))

# 超参数
learning_rate = 0.001
epochs = 120
batch_size = 32

cifar10 = tf.keras.datasets.cifar10
(x_train, y_train), (x_test, y_test) = cifar10.load_data()

x_train = x_train.astype(np.float32)
x_test = x_test.astype(np.float32)

x_train = x_train / 255
x_test = x_test / 255

model = tf.keras.models.Sequential([
    # Lambda 层，对输入图片进行 resize，以下是将图片扩大了 7 倍
    # resize 时，默认使用最近邻插值，想要用其它插值方式，需要直接修改 K.resize_images 方法的源代码。
    layers.Lambda(lambda img: K.resize_images(img, 7, 7, data_format='channels_last'), input_shape=(32, 32, 3)),
    layers.ZeroPadding2D(padding=(2, 2)),
    layers.Conv2D(64, (11, 11), strides=(4, 4), padding='valid', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.MaxPooling2D(pool_size=(3, 3), strides=(2, 2)),

    layers.Conv2D(192, (5, 5), strides=(1, 1), padding='same', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.MaxPooling2D(pool_size=(3, 3), strides=(2, 2)),

    layers.Conv2D(384, (3, 3), strides=(1, 1), padding='same', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.Conv2D(256, (3, 3), strides=(1, 1), padding='same', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.Conv2D(256, (3, 3), strides=(1, 1), padding='same', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.MaxPooling2D(pool_size=(3, 3), strides=(2, 2)),

    layers.Flatten(),

    layers.Dense(4096, activation='relu', kernel_initializer='he_uniform'),
    layers.Dropout(drop_rate),
    layers.Dense(4096, activation='relu', kernel_initializer='he_uniform'),
    layers.Dropout(drop_rate),
    layers.Dense(num_classes, activation='softmax', kernel_initializer='he_uniform')
])

model.summary()

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

model.fit(x_train, y_train,
          epochs=epochs,
          batch_size=batch_size,
          verbose=2,
          validation_data=(x_val, y_val))

方法 2：使用 tensorflow.keras.utils.Sequence，构造一个 data generator

import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.python.keras import backend as K
from tensorflow.keras.utils import Sequence

from sklearn.model_selection import StratifiedShuffleSplit

import cv2
import os
import numpy as np
import h5py
import time

class CIFAR10Sequence(Sequence):
    def __init__(self, x_set, y_set, batch_size):
        """
        :param x_set: hdf5
        :param y_set: hdf5
        :param batch_size: int
        """
        self.x, self.y = x_set, y_set
        self.batch_size = batch_size

    def __len__(self):
        return int(np.ceil(len(self.x) / float(self.batch_size)))

    def __getitem__(self, idx):
        batch_x = self.x[idx * self.batch_size:(idx + 1) * self.batch_size]
        batch_y = self.y[idx * self.batch_size:(idx + 1) * self.batch_size]

        batch_x = batch_x.astype(np.float32)
        batch_x = batch_x / 255

        return batch_x, batch_y


def _resized_data():
    """
    将 resize 后的 cifar-10 保存到 'data/cifar-10.h5'
    图片大小: [224, 224, 3]
    :return: None
    """
    cifar10 = tf.keras.datasets.cifar10

    (x_train, y_train), (x_test, y_test) = cifar10.load_data()

    start_time = time.clock()

    x_train = np.array([cv2.resize(img, (224, 224), interpolation=cv2.INTER_CUBIC) for img in x_train])
    x_test = np.array([cv2.resize(img, (224, 224), interpolation=cv2.INTER_CUBIC) for img in x_test])

    # initialize
    x_val = np.array([])
    y_val = np.array([])

    sss = StratifiedShuffleSplit(n_splits=1, test_size=0.1, random_state=32)
    for train_index, val_index in sss.split(x_train, y_train):
        print("TRAIN:", train_index, "VAL:", val_index)
        x_train, x_val = x_train[train_index], x_train[val_index]
        y_train, y_val = y_train[train_index], y_train[val_index]

    end_time = time.clock()
    print('Time consuming of resizing: ', (end_time - start_time))

    # 写文件
    filename = 'data/cifar-10.h5'
    h5f = h5py.File(filename, 'w')
    h5f.create_dataset('x_train', data=x_train)
    h5f.create_dataset('y_train', data=y_train)
    h5f.create_dataset('x_val', data=x_val)
    h5f.create_dataset('y_val', data=y_val)
    h5f.create_dataset('x_test', data=x_test)
    h5f.create_dataset('y_test', data=y_test)
    h5f.close()


def load_resized_data(filename='data/cifar-10.h5'):
    if not os.path.exists(filename):
        _resized_data()
    
    # 不要关闭 h5 文件，否则将无法读取数据，这一步并不会直接将数据加载到内存中
    # h5 文件支持切片读取，而且也很快
    h5f = h5py.File(filename, 'r')
    x_train = h5f['x_train']
    y_train = h5f['y_train']
    x_val = h5f['x_val']
    y_val = h5f['y_val']
    x_test = h5f['x_test']
    y_test = h5f['y_test']

    return (x_train, y_train), (x_val, y_val), (x_test, y_test)


K.clear_session()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True  # 不全部占满显存, 按需分配
K.set_session(tf.Session(config=config))

# 超参数
learning_rate = 0.001
epochs = 120
batch_size = 32

(x_train, y_train), (x_val, y_val), (x_test, y_test) = load_resized_data()

x_val = x_val.astype(np.float32)
x_test = x_test.astype(np.float32)

x_val = x_val / 255
x_test = x_test / 255

model = tf.keras.models.Sequential([
    layers.ZeroPadding2D(padding=(2, 2), input_shape=(224, 224, 3)),
    layers.Conv2D(64, (11, 11), strides=(4, 4), padding='valid', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.MaxPooling2D(pool_size=(3, 3), strides=(2, 2)),

    layers.Conv2D(192, (5, 5), strides=(1, 1), padding='same', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.MaxPooling2D(pool_size=(3, 3), strides=(2, 2)),

    layers.Conv2D(384, (3, 3), strides=(1, 1), padding='same', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.Conv2D(256, (3, 3), strides=(1, 1), padding='same', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.Conv2D(256, (3, 3), strides=(1, 1), padding='same', activation='relu',
                  kernel_initializer='he_uniform'),
    layers.MaxPooling2D(pool_size=(3, 3), strides=(2, 2)),

    layers.Flatten(),

    layers.Dense(4096, activation='relu', kernel_initializer='he_uniform'),
    layers.Dropout(drop_rate),
    layers.Dense(4096, activation='relu', kernel_initializer='he_uniform'),
    layers.Dropout(drop_rate),
    layers.Dense(num_classes, activation='softmax', kernel_initializer='he_uniform')
])

model.summary()

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

# shuffle 默认为 True, 意味着在训练一个 epoch 之后，CIFAR10Sequence 的 idx 会随机选择，而不是顺序选择，这样在 batch-level 进行了随机，一个 batch 内的样本顺序是固定的
model.fit_generator(CIFAR10Sequence(x_train, y_train, batch_size=batch_size),
                    # steps_per_epoch=int(np.ceil(len(x_train)/batch_size)),
                    epochs=epochs,
                    verbose=2,
                    callbacks=None,
                    validation_data=(x_val[:], y_val[:]))

References

class CIFAR10Sequence(Sequence) -- github
keras.utils.Sequence()
AlexNet in pytorch/vision