python tensorflow 实现VGG16

以下是使用TensorFlow实现VGG16的示例代码：

import tensorflow as tf
import numpy as np
import os

def load_weights(weights_file, sess):
    weights_dict = np.load(weights_file, encoding='bytes').item()
    for op_name in weights_dict:
        if op_name not in ['fc6', 'fc7', 'fc8']:
            with tf.variable_scope(op_name, reuse=True):
                for data in weights_dict[op_name]:
                    if len(data.shape) == 1:
                        var = tf.get_variable('biases', trainable=False)
                        sess.run(var.assign(data))
                    else:
                        var = tf.get_variable('weights', trainable=False)
                        sess.run(var.assign(data))

def VGG16(x):
    # conv1_1
    with tf.variable_scope("conv1_1") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 3, 64], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[64], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(x, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv1_1 = tf.nn.relu(out, name=scope.name)

    # conv1_2
    with tf.variable_scope("conv1_2") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 64, 64], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[64], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(conv1_1, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv1_2 = tf.nn.relu(out, name=scope.name)

    # pool1
    pool1 = tf.nn.max_pool(conv1_2, [1, 2, 2, 1], [1, 2, 2, 1], padding="SAME")

    # conv2_1
    with tf.variable_scope("conv2_1") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 64, 128], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[128], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv2_1 = tf.nn.relu(out, name=scope.name)

    # conv2_2
    with tf.variable_scope("conv2_2") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 128, 128], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[128], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(conv2_1, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv2_2 = tf.nn.relu(out, name=scope.name)

    # pool2
    pool2 = tf.nn.max_pool(conv2_2, [1, 2, 2, 1], [1, 2, 2, 1], padding="SAME")

    # conv3_1
    with tf.variable_scope("conv3_1") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 128, 256], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[256], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv3_1 = tf.nn.relu(out, name=scope.name)

    # conv3_2
    with tf.variable_scope("conv3_2") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 256, 256], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[256], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(conv3_1, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv3_2 = tf.nn.relu(out, name=scope.name)

    # conv3_3
    with tf.variable_scope("conv3_3") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 256, 256], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[256], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(conv3_2, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv3_3 = tf.nn.relu(out, name=scope.name)

    # pool3
    pool3 = tf.nn.max_pool(conv3_3, [1, 2, 2, 1], [1, 2, 2, 1], padding="SAME")

    # conv4_1
    with tf.variable_scope("conv4_1") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 256, 512], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[512], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(pool3, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv4_1 = tf.nn.relu(out, name=scope.name)

    # conv4_2
    with tf.variable_scope("conv4_2") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 512, 512], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[512], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(conv4_1, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv4_2 = tf.nn.relu(out, name=scope.name)

    # conv4_3
    with tf.variable_scope("conv4_3") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 512, 512], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[512], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(conv4_2, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv4_3 = tf.nn.relu(out, name=scope.name)

    # pool4
    pool4 = tf.nn.max_pool(conv4_3, [1, 2, 2, 1], [1, 2, 2, 1], padding="SAME")

    # conv5_1
    with tf.variable_scope("conv5_1") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 512, 512], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[512], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(pool4, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv5_1 = tf.nn.relu(out, name=scope.name)

    # conv5_2
    with tf.variable_scope("conv5_2") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 512, 512], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[512], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(conv5_1, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv5_2 = tf.nn.relu(out, name=scope.name)

    # conv5_3
    with tf.variable_scope("conv5_3") as scope:
        kernel = tf.get_variable("weights", shape=[3, 3, 512, 512], dtype=tf.float32, trainable=False)
        biases = tf.get_variable("biases", shape=[512], dtype=tf.float32, trainable=False)
        conv = tf.nn.conv2d(conv5_2, kernel, [1, 1, 1, 1], padding="SAME")
        out = tf.nn.bias_add(conv, biases)
        conv5_3 = tf.nn.relu(out, name=scope.name)

    # pool5
    pool5 = tf.nn.max_pool(conv5_3, [1, 2, 2, 1], [1, 2, 2, 1], padding="SAME")

    # fc6
    with tf.variable_scope("fc6") as scope:
        shape = int(np.prod(pool5.get_shape()[1:]))
        fc6w = tf.get_variable("weights", shape=[shape, 4096], dtype=tf.float32, trainable=False)
        fc6b = tf.get_variable("biases", shape=[4096], dtype=tf.float32, trainable=False)
        pool5_flat = tf.reshape(pool5, [-1, shape])
        fc6 = tf.nn.bias_add(tf.matmul(pool5_flat, fc6w), fc6b)
        fc6 = tf.nn.relu(fc6)

    # fc7
    with tf.variable_scope("fc7") as scope:
        fc7w = tf.get_variable("weights", shape=[4096, 4096], dtype=tf.float32, trainable=False)
        fc7b = tf.get_variable("biases", shape=[4096], dtype=tf.float32, trainable=False)
        fc7 = tf.nn.bias_add(tf.matmul(fc6, fc7w), fc7b)
        fc7 = tf.nn.relu(fc7)

    # fc8
    with tf.variable_scope("fc8") as scope:
        fc8w = tf.get_variable("weights", shape=[4096, 1000], dtype=tf.float32, trainable=False)
        fc8b = tf.get_variable("biases", shape=[1000], dtype=tf.float32, trainable=False)
        fc8 = tf.nn.bias_add(tf.matmul(fc7, fc8w), fc8b)

    return fc8

if __name__ == '__main__':
    img = tf.placeholder(tf.float32, [None, 224, 224, 3])
    with tf.Session() as sess:
        vgg16_npy_path = 'vgg16.npy'
        if os.path.isfile(vgg16_npy_path):
            print('Loading the weights of VGG16')
            load_weights(vgg16_npy_path, sess)
        else:
            print('Please download the pre-trained VGG16 weights')
        preds = VGG16(img)
        sess.run(tf.global_variables_initializer())
        img_data = np.zeros((10, 224, 224, 3))
        predictions = sess.run(preds, feed_dict={img: img_data})
        print(predictions.shape) # (10, 1000)

其中，`load_weights`函数用于从预训练模型文件中读取权重值，`VGG16`函数定义了VGG16的网络结构，`if __name__ == '__main__'`中的代码用于测试模型，生成一个10x1000的预测结果。需要注意的是，在运行代码之前，需要下载预训练模型的权重文件`vgg16.npy`，并将其放在代码所在目录下。