残差网络的理解及其基于cifar10数据集的代码实现

最新推荐文章于 2023-02-20 10:14:10 发布

热烈的南风

最新推荐文章于 2023-02-20 10:14:10 发布

阅读量426

点赞数 1

文章标签： python 分类算法

本文链接：https://blog.csdn.net/weixin_46574699/article/details/119608903

版权

本文介绍了深度残差学习如何解决网络深度增加导致的性能下降问题，并提供了基于CIFAR10数据集的残差网络实现。通过残差块中的卷积、BN层和激活层进行图像分类。

摘要由CSDN通过智能技术生成

这是我的第一篇帖子，就用它来写何凯明大神的深度残差学习吧，下面内容都是我总结修改的成果。

Deep Residual Learning for Image Recognition

文章的重要几点如下：
1.解决了之前对于网络越深，反而效果越不好的问题，也就是所谓的退化问题；
2. 残差块中有卷积层，BN层，和激活层，另外对其进行求和。

代码实现基于cifar10数据集

cifar10数据集共有60000张图片，共有十类数据，每一类有6000张图片。其中50000张训练集，10000张测试集。残差网络主要用于cifar10数据集的分类。

// resnet.py用于构建具体的残差网络
from __future__ import division

import six
from keras.models import Model
from keras.layers import (
    Input,
    Activation,
    Dense,
    Flatten
)
from keras.layers.convolutional import (
    Conv2D,
    MaxPooling2D,
    AveragePooling2D
)
from keras.layers.merge import add
from keras.layers.normalization import BatchNormalization
from keras.regularizers import l2
from keras import backend as K


def _bn_relu(input):
    """Helper to build a BN -> relu block
    """
    norm = BatchNormalization(axis=CHANNEL_AXIS)(input)
    return Activation("relu")(norm)


def _conv_bn_relu(**conv_params):
    """Helper to build a conv -> BN -> relu block
    """
    filters = conv_params["filters"]
    kernel_size = conv_params["kernel_size"]
    strides = conv_params.setdefault("strides", (1, 1))
    kernel_initializer = conv_params.setdefault("kernel_initializer", "he_normal")
    padding = conv_params.setdefault("padding", "same")
    kernel_regularizer = conv_params.setdefault("kernel_regularizer", l2(1.e-4))

    def f(input):
        conv = Conv2D(filters=filters, kernel_size=kernel_size,
                      strides=strides, padding=padding,
                      kernel_initializer=kernel_initializer,
                      kernel_regularizer=kernel_regularizer)(input)
        return _bn_relu(conv)

    return f


def _bn_relu_conv(**conv_params):
    """Helper to build a BN -> relu -> conv block.
    This is an improved scheme proposed in http://arxiv.org/pdf/1603.05027v2.pdf
    """
    filters = conv_params["filters"]
    kernel_size = conv_params["kernel_size"]
    strides = conv_params.setdefault("strides", (1, 1))
    kernel_initializer = conv_params.setdefault("kernel_initializer", "he_normal")
    padding = conv_params.setdefault("padding", "same")
    kernel_regularizer = conv_params.setdefault("kernel_regularizer", l2(1.e-4))

    def f(input):
        activation = _bn_relu(input)
        return Conv2D(filters=filters, kernel_size=kernel_size,
                      strides=strides, padding=padding,
                      kernel_initializer=kernel_initializer,
                      kernel_regularizer=kernel_regularizer)(activation)

    return f


def _shortcut(input, residual):
    """Adds a shortcut between input and residual block and merges them with "sum"
    """
    # Expand channels of shortcut to match residual.
    # Stride appropriately to match residual (width, height)
    # Should be int if network architecture is correctly configured.
    input_shape = K.int_shape(input)
    residual_shape = K.int_shape(residual)
    stride_width = int(round(input_shape[