vggf.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import tensorflow as tf
import scipy.misc
import scipy.io
import numpy as np

_R_MEAN = 123.68
_G_MEAN = 116.78
_B_MEAN = 103.94


def construct_net(data_path, input_image,codelens):
  
    data = scipy.io.loadmat(data_path)
    layers = (
        'conv1', 'relu1', 'norm1', 'pool1',
        'conv2', 'relu2', 'norm2', 'pool2',
        'conv3', 'relu3', 'conv4', 'relu4',
        'conv5', 'relu5', 'pool5',
        'fc6', 'relu6', 'fc7', 'relu7','fc8'

    )
    weights = data['layers'][0]
    mean = np.ones([224,224,3],dtype=np.float32)
    mean[:,:,0] = _R_MEAN
    mean[:,:,1] = _G_MEAN
    mean[:,:,2] = _B_MEAN

    net = {}
    ops = []  #variable_list

    #current = tf.convert_to_tensor(input_image,dtype='float')
    current = input_image - mean
    for i, name in enumerate(layers[:-1]):
        if name.startswith('conv'):
            kernels, bias = weights[i][0][0][2][0]
            # matconvnet: weights are [width, height, in_channels, out_channels]
            # tensorflow: weights are [height, width, in_channels, out_channels]
            #kernels = np.transpose(kernels, (1, 0, 2, 3))

            bias = bias.reshape(-1)
            pad = weights[i][0][0][4]
            stride = weights[i][0][0][5]
            current = _conv_layer(current,kernels,bias,pad,stride,i,ops,net)
        elif name.startswith('relu'):
            current = tf.nn.relu(current)
        elif name.startswith('pool'):
            stride = weights[i][0][0][4]
            pad = weights[i][0][0][5]
            area = weights[i][0][0][3]
            current = _pool_layer(current,stride,pad,area)
        elif name.startswith('fc'):
            kernels, bias = weights[i][0][0][2][0]
            # matconvnet: weights are [width, height, in_channels, out_channels]
            # tensorflow: weights are [height, width, in_channels, out_channels]
            #kernels = np.transpose(kernels, (1, 0, 2, 3))

            bias = bias.reshape(-1)
            current = _full_conv(current,kernels,bias,i,ops,net)
            #current = tf.matmul(tf.reshape(current, [-1, 1 * 1 * 4096]), kernels) + bias
        elif name.startswith('norm'):
            current = tf.nn.local_response_normalization(current, depth_radius=2.5, bias=2.000, alpha=0.0001, beta=0.75)
        net[name] = current
    
    W_fc8 = tf.truncated_normal([4096,codelens], stddev=0.01)
    #偏置值
    b_fc8 = tf.truncated_normal([codelens],stddev = 0.01)

    w = tf.Variable(W_fc8, name='w' + str(20))
    b = tf.Variable(b_fc8, name='bias' + str(20))

    ops.append(w)
    ops.append(b)
    #将卷积的产出展开
    fc8 = tf.matmul(tf.squeeze(current),w) + b
    net['weigh21'] = w
    net['b21'] = b
    # conv = tf.nn.conv2d(current, w, strides=[1, 1, 1, 1], padding='VALID') + tf.Variable(b)
    net[layers[-1]] = fc8
    return net,mean,ops
               
def _conv_layer(input, weights, bias,pad,stride,i,ops,net):
    pad = pad[0]
    stride= stride[0]
    input = tf.pad(input, [[0, 0], [pad[0], pad[1]], [pad[2], pad[3]], [0, 0]], "CONSTANT")
    w = tf.Variable(weights,name='w'+str(i),dtype='float32')
    b = tf.Variable(bias,name='bias'+str(i),dtype='float32')
    ops.append(w)
    ops.append(b)
    net['weights' + str(i)] = w
    net['b' + str(i)] = b
    conv = tf.nn.conv2d(input, w, strides=[1,stride[0],stride[1],1],padding='VALID',name='conv'+str(i))
    return tf.nn.bias_add(conv, b,name='add'+str(i))

def _full_conv(input, weights, bias,i,ops,net):
    w = tf.Variable(weights, name='w' + str(i),dtype='float32')
    b = tf.Variable(bias, name='bias' + str(i),dtype='float32')
    ops.append(w)
    ops.append(b)
    net['weights' + str(i)] = w
    net['b' + str(i)] = b
    conv = tf.nn.conv2d(input, w,strides=[1,1,1,1],padding='VALID',name='fc'+str(i))
    return tf.nn.bias_add(conv, b,name='add'+str(i))

def _pool_layer(input,stride,pad,area):
    pad = pad[0]
    area = area[0]
    stride = stride[0]
    input = tf.pad(input, [[0, 0], [pad[0], pad[1]], [pad[2], pad[3]], [0, 0]], "CONSTANT")
    return tf.nn.max_pool(input, ksize=[1, area[0], area[1], 1], strides=[1,stride[0],stride[1],1],padding='VALID')