卷积神经网络实践
本节介绍如何构造一个简单的CNN模型进行手写数字识别,
但在现实场景中,往往使用imagenet预训练的深度CNN模型进行迁移学习,能极大地提升预测准确率,
可参考我在百度大数据竞赛中开源的模型: keras-dog
数据处理
- dataset处理成四维的,label仍然作为one-hot encoding
def reformat(dataset, labels, image_size, num_labels, num_channels):
dataset = dataset.reshape(
(-1, image_size, image_size, num_channels)).astype(np.float32)
labels = (np.arange(num_labels) == labels[:, None]).astype(np.float32)
return dataset, labels
- 将lesson2的dnn转为cnn很简单,只要把WX+b改为conv2d(X)+b即可
- 关键在于conv2d
tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, data_format=None, name=None) {#conv2d}
给定四维的input和filter tensor,计算一个二维卷积
Args:
input: ATensor. type必须是以下几种类型之一:half,float32,float64.filter: ATensor. type和input必须相同strides: A list ofints.一维,长度4, 在input上切片采样时,每个方向上的滑窗步长,必须和format指定的维度同阶padding: Astringfrom:"SAME", "VALID". padding 算法的类型use_cudnn_on_gpu: An optionalbool. Defaults toTrue.data_format: An optionalstringfrom:"NHWC", "NCHW", 默认为"NHWC"。 指定输入输出数据格式,默认格式为"NHWC", 数据按这样的顺序存储:[batch, in_height, in_width, in_channels]也可以用这种方式:"NCHW", 数据按这样的顺序存储:[batch, in_channels, in_height, in_width]name: 操作名,可选.
Returns:
A Tensor. type与input相同
Given an input tensor of shape [batch, in_height, in_width, in_channels]
and a filter / kernel tensor of shape
[filter_height, filter_width, in_channels, out_channels]
conv2d实际上执行了以下操作:
- 将filter转为二维矩阵,shape为
[filter_height * filter_width * in_channels, output_channels]. - 从input tensor中提取image patches,每个patch是一个virtual tensor,shape
[batch, out_height, out_width, filter_height * filter_width * in_channels]. - 将每个filter矩阵和image patch向量相乘
具体来讲,当data_format为NHWC时:
output[b, i, j, k] =
sum_{di, dj, q} input[b, strides[1] * i + di, strides[2] * j + dj, q] *
filter[di, dj, q, k]
input 中的每个patch都作用于filter,每个patch都能获得其他patch对filter的训练
需要满足strides[0] = strides[3] = 1. 大多数水平步长和垂直步长相同的情况下:strides = [1, stride, stride, 1].
- 然后再接一个WX+b连Relu连WX+b的全连接神经网络即可
Max Pooling
在tf.nn.conv2d后面接tf.nn.max_pool,将卷积层输出减小,从而减少要调整的参数
tf.nn.max_pool(value, ksize, strides, padding, data_format='NHWC', name=None) {#max_pool}
Performs the max pooling on the input.
Args:
value: A 4-DTensorwith shape[batch, height, width, channels]and typetf.float32.ksize: A list of ints that has length >= 4. 要执行取最值的切片在各个维度上的尺寸strides: A list of ints that has length >= 4. 取切片的步长padding: A string, either'VALID'or'SAME'. padding算法data_format: A string. 'NHWC' and 'NCHW' are supported.name: 操作名,可选
Returns:
A Tensor with type tf.float32. The max pooled output tensor.
优化
仿照lesson2,添加learning rate decay 和 drop out,可以将准确率提高到90.6%
补充
- 最近在用GPU版本的TensorFlow,发现,如果import tensorflow放在代码第一行,运行会报段错误(pycharm debug模式下不会),因此最好在import tensorflow前import numpy或者其他的module