最近在看残差网络的论文,然后看了很多网上实现的代码,我发现很多人写代码是没有逻辑的,其实那个代码写得压根就不对,只是可能恰巧结果对,然后我不明白明明池化很简单的道理,非要说成什么降采样,给我整的看论文看得我一脸蒙逼,现在的模型适合大多数数据集的几乎不存在,我参考论文网上的帖子,实现了resnet50,但是我没训练,因为没有好的224*224的数据集,硬盘太小,大的程序也跑不起来,今天把代码贴出来,然后如果需要的话拿去参考。还有就是复原模型最重要的就是搞清楚网络的结构,最好是看着结构图来做模型,这样你就会很清楚每一层的tensor是如何进行变化的,然后就不会那么蒙圈了,在网上找了个不错的网络结构图,分享给大家 https://blog.csdn.net/haoji007/article/details/90259359
然后下面就是代码了,我相信只要你仔细看过残差网络的论文,你都会理解的,然后我打算今晚再做一个18层的残差网络用来训练cifar10,看一下训练的效果,下面是代码
import tensorflow as tf import tensorflow.contrib.slim as slim WEIGHT_DECAY = 0.01 # 这段是我之前看别人的帖子上做的,但是后来我做的过程中我发现其实并没有什么用, # 改变featuremap的大小用卷积也完全可以实现,所以我把它注释掉,发现也是正常的 # 但是由于tensorflow老大哥最近又开始迭代更新了,所以可能会有warning,但是 # 没什么大影响 # def sampling(input_tensor, # ksize=1, # stride=2): # data = input_tensor # if stride > 1: # data = slim.max_pool2d(data, ksize, stride=stride) # print('sampling', 2) # return data def conv2d_same(input_tensor, num_outputs, kernel_size, stride, is_train=True, activation_fn=tf.nn.relu, normalizer_fc=True ): data = input_tensor if stride is 1: data = slim.conv2d(inputs=data, num_outputs=num_outputs, kernel_size=kernel_size, stride=stride, weights_regularizer=slim.l2_regularizer(WEIGHT_DECAY), activation_fn=None, padding='SAME', ) else: pad_total = kernel_size - 1 pad_begin = pad_total // 2 pad_end = pad_total - pad_begin data = tf.pad(data, [[0, 0], [pad_begin, pad_end], [pad_begin, pad_end], [0, 0]]) data = slim.conv2d(data, num_outputs=num_outputs, kernel_size=kernel_size, stride=stride, weights_regularizer=slim.l2_regularizer(WEIGHT_DECAY), activation_fn=None, padding='VALID', ) if normalizer_fc: data = tf.layers.batch_normalization(data, training=is_train) if activation_fn is not None: data = activation_fn(data) return data def bottle_net(input_tensor, output_depth, is_train, stride=1): data = input_tensor depth = input_tensor.get_shape().as_list()[-1] if depth == output_depth: shortcut_tensor = input_tensor else: shortcut_tensor = conv2d_same(input_tensor, output_depth, 1, stride, is_train=is_train, activation_fn=None, normalizer_fc=True) data = conv2d_same(data, output_depth // 4, 1, 1, is_train=is_train) data = conv2d_same(data, output_depth // 4, 3, stride, is_train=is_train) data = conv2d_same(data, output_depth, 1, 1, is_train=is_train, activation_fn=None, normalizer_fc=False) # 生成残差 data = data + shortcut_tensor data = tf.nn.relu(data) return data def create_block(input_tensor, output_depth, block_nums, init_stride=1, is_train=True, scope='block'): with tf.variable_scope(scope): data = bottle_net(input_tensor, output_depth, is_train=is_train, stride=init_stride) for i in range(1, block_nums): data = bottle_net(data, output_depth, is_train=is_train) return data def ResNet(input_tensor, num_output, is_train, scope='resnet50'): data = input_tensor with tf.variable_scope(scope): data = conv2d_same(data, 64, 7, 2, is_train=is_train, normalizer_fc=True) data = slim.max_pool2d(data, 3, 2, padding='SAME', scope='pool_1') # 第一个残差块组 data = create_block(data, 256, 3, init_stride=1, is_train=is_train, scope='block1') # 第二个残差块组 data = create_block(data, 512, 4, init_stride=2, is_train=is_train, scope='block2') # 第三个残差块组 data = create_block(data, 1024, 6, init_stride=2, is_train=is_train, scope='block3') # 第四个残差块组 data = create_block(data, 2048, 3, init_stride=2, is_train=is_train, scope='block4') # 接下来就是池化层和全连接层 data = slim.avg_pool2d(data, 7) data = slim.conv2d(data, num_output, 1, activation_fn=None, scope='final_conv') data_shape = data.get_shape().as_list() nodes = data_shape[1] * data_shape[2] * data_shape[3] data = tf.reshape(data, [-1, nodes]) return data if __name__ == '__main__': x = tf.random_normal([32, 224, 224, 3]) data = ResNet(x, 1000, True) print(data)
能力极其有限,有错误的地方希望各位同学指正