tensorflow版的bvlc模型

研究相关的图片分类，偶然看到bvlc模型，但是没有tensorflow版本的，所以将caffe版本的改成了tensorflow的:

关于模型这个图：

下面贴出通用模板：

  1 from __future__ import print_function
  2 import tensorflow as tf
  3 import numpy as np
  4 from scipy.misc import imread, imresize
  5
  6
  7 class BVLG:
  8     def __init__(self, imgs, weights=None, sess=None):
  9         self.imgs = imgs
 10         self.convlayers()
 11         self.fc_layers()
 12
 13         self.probs = tf.nn.softmax(self.fc3l)
 14         if weights is not None and sess is not None:
 15             self.load_weights(weights,sess)
 16
 17     def convlayers(self):
 18         self.parameters = []
 19
 20         # zero-mean input
 21         with tf.name_scope(‘preprocess‘) as scope:
 22             mean = tf.constant([123.68, 116.779, 103.939], dtype=tf.float32, shape=[1, 1, 1, 3], name=‘img_mean‘)
 23             images = self.imgs - mean
 24
 25         # conv1
 26         with tf.name_scope(‘conv1‘) as scope:
 27             kernel = tf.Variable(tf.truncated_normal([7, 7, 3, 96], dtype=tf.float32,
 28                                                      stddev=1e-1), name=‘weights‘)
 29             conv = tf.nn.conv2d(images, kernel, [3, 3, 1, 1], padding=‘SAME‘)
 30             biases = tf.Variable(tf.constant(0.0, shape=[96], dtype=tf.float32),
 31                                  trainable=True, name=‘biases‘)
 32             out = tf.nn.bias_add(conv, biases)
 33             self.conv1 = tf.nn.relu(out, name=scope)
 34             self.parameters += [kernel, biases]
 35
 36         # pool1
 37         self.pool1 = tf.nn.max_pool(self.conv1,
 38                                     ksize=[1, 3, 3, 1],
 39                                     strides=[1, 2, 2, 1],
 40                                     padding=‘SAME‘,
 41                                     name=‘pool1‘)
 42
 43         # conv2
 44         with tf.name_scope(‘conv2‘) as scope:
 45             kernel = tf.Variable(tf.truncated_normal([4, 4, 96, 256], dtype=tf.float32,
 46                                                      stddev=1e-1), name=‘weights‘)
 47             conv = tf.nn.conv2d(self.pool1, kernel, [1, 1, 1, 1], padding=‘SAME‘)
 48             biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),
 49                                  trainable=True, name=‘biases‘)
 50             out = tf.nn.bias_add(conv, biases)
 51             self.conv2_1 = tf.nn.relu(out, name=scope)
 52             self.parameters += [kernel, biases]
 53
 54
 55         # pool2
 56         self.pool2 = tf.nn.max_pool(self.conv2,
 57                                     ksize=[1, 3, 3, 1],
 58                                     strides=[1, 2, 2, 1],
 59                                     padding=‘SAME‘,
 60                                     name=‘pool2‘)
 61
 62         # conv5
 63         with tf.name_scope(‘conv5‘) as scope:
 64             kernel = tf.Variable(tf.truncated_normal([3, 3, 256, 256], dtype=tf.float32,
 65                                                      stddev=1e-1), name=‘weights‘)
 66             conv = tf.nn.conv2d(self.pool2, kernel, [1, 1, 1, 1], padding=‘SAME‘)
 67             biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),
 68                                  trainable=True, name=‘biases‘)
 69             out = tf.nn.bias_add(conv, biases)
 70             self.conv5 = tf.nn.relu(out, name=scope)
 71             self.parameters += [kernel, biases]
 72
 73         # pool5
 74         self.pool5 = tf.nn.max_pool(self.conv5,
 75                                     ksize=[1, 2, 2, 1],
 76                                     strides=[1, 2, 2, 1],
 77                                     padding=‘SAME‘,
 78                                     name=‘pool4‘)
 79
 80     def fc_layers(self):
 81         # fc1
 82         with tf.name_scope(‘fc1‘) as scope:
 83             shape = int(np.prod(self.pool5.get_shape()[1:]))
 84             fc1w = tf.Variable(tf.truncated_normal([shape, 4096],
 85                                                    dtype=tf.float32,
 86                                                    stddev=1e-1), name=‘weights‘)
 87             fc1b = tf.Variable(tf.constant(1.0, shape=[4096], dtype=tf.float32),
 88                                trainable=True, name=‘biases‘)
 89             pool5_flat = tf.reshape(self.pool5, [-1, shape])
 90             fc1l = tf.nn.bias_add(tf.matmul(pool5_flat, fc1w), fc1b)
 91             self.fc1 = tf.nn.relu(fc1l)
 92             self.parameters += [fc1w, fc1b]
 93
 94         # fc3
 95         with tf.name_scope(‘fc3‘) as scope:
 96             fc3w = tf.Variable(tf.truncated_normal([4096, 587],
 97                                                    dtype=tf.float32,
 98                                                    stddev=1e-1), name=‘weights‘)
 99             fc3b = tf.Variable(tf.constant(1.0, shape=[587], dtype=tf.float32),
100                                trainable=True, name=‘biases‘)
101             self.fc3l = tf.nn.bias_add(tf.matmul(self.fc2, fc3w), fc3b)
102             self.parameters += [fc3w, fc3b]

caffe版本的ImageNet模型地址：https://github.com/BVLC/caffe/tree/master/models/bvlc_reference_caffenet