使用tf-slim的ResNet V1 152和ResNet V2 152预训练模型进行图像分类

本文使用tf-slim的ResNet V1 152和ResNet V2 152预训练模型进行图像分类，并研究slim网络的scope命名等。

tf-slim文档不太多，实现过程中多参考官网的源码： https://github.com/tensorflow/models/tree/master/research/slim
注意resnet v2的预处理有点不一样，输入是299而不是224
ResNet V2 152
(tf-slim: ResNet V2 models use Inception pre-processing and input image size of 299 (use –preprocessing_name inception –eval_image_size 299 when using eval_image_classifier.py). Performance numbers for ResNet V2 models are reported on the ImageNet validation set.)

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Fri Sep 29 16:25:16 2017

@author: wayne


我们用的是tf1.2，最新的tf1.3地址是
https://github.com/tensorflow/models/tree/master/research/slim

http://geek.csdn.net/news/detail/126133
如何用TensorFlow和TF-Slim实现图像分类与分割

https://www.2cto.com/kf/201706/649266.html
【Tensorflow】辅助工具篇——tensorflow slim(TF-Slim)介绍

https://stackoverflow.com/questions/39582703/using-pre-trained-inception-resnet-v2-with-tensorflow
The Inception networks expect the input image to have color channels scaled from [-1, 1]. As seen here.
You could either use the existing preprocessing, or in your example just scale the images yourself: im = 2*(im/255.0)-1.0 before feeding them to the network.
Without scaling the input [0-255] is much larger than the network expects and the biases all work to very strongly predict category 918 (comic books).

TensorFlow实现ResNet（ResNet 152网络结构的forward耗时检测）
http://blog.csdn.net/superman_xxx/article/details/65452735
ResNet原理及其在TF-Slim中的实现
http://www.jianshu.com/p/3af06422c768
"""

import tensorflow as tf
slim = tf.contrib.slim
from PIL import Image
import numpy as np
import matplotlib.pyplot as plt
import imagenet  #注意需要用最新版tf中的对应文件，否则http地址是不对的

from inception_resnet_v2 import *
from resnet_v1 import *
from resnet_v2 import *

import inception_preprocessing
import vgg_preprocessing


'''
inception_resnet_v2
  Returns:
    tensor_out: output tensor corresponding to the final_endpoint.
    end_points: a set of activations for external use, for example summaries or
                losses.
'''
tf.reset_default_graph()

checkpoint_file = 'inception_resnet_v2_2016_08_30.ckpt'
image = tf.image.decode_jpeg(tf.read_file('dog.jpeg'), channels=3) #['dog.jpg', 'panda.jpg']

image_size = inception_resnet_v2.default_image_size #  299

'''这个函数做了裁剪，缩放和归一化等'''
processed_image = inception_preprocessing.preprocess_image(image, 
                                                        image_size, 
                                                        image_size,
                                                        is_training=False,)
processed_images  = tf.expand_dims(processed_image, 0)

'''Creates the Inception Resnet V2 model.'''
arg_scope = inception_resnet_v2_arg_scope()
with slim.arg_scope(arg_scope):
  logits, end_points = inception_resnet_v2(processed_images, is_training=False)   

probabilities = tf.nn.softmax(logits)

saver = tf.train.Saver()


with tf.Session() as sess:
    saver.restore(sess, checkpoint_file)

    #predict_values, logit_values = sess.run([end_points['Predictions'], logits])
    logits2, image2, network_inputs, probabilities2 = sess.run([logits,
                                                                image,
                                                       processed_images,
                                                       probabilities])

    print(logits2)  
    print(logits2.shape) #(1, 1001)

    print(network_inputs.shape)
    print(probabilities2.shape)
    probabilities2 = probabilities2[0,:]
    sorted_inds = [i[0] for i in sorted(enumerate(-probabilities2),
                                        key=lambda x:x[1])]    


# 显示下载的图片
plt.figure()
plt.imshow(image2)#.astype(np.uint8))
plt.suptitle("Original image", fontsize=14, fontweight='bold')
plt.axis('off')
plt.show()

# 显示最终传入网络模型的图片
plt.imshow(network_inputs[0,:,:,:])
plt.suptitle("Resized, Cropped and Mean-Centered inputs to network",
             fontsize=14, fontweight='bold')
plt.axis('off')
plt.show()

names = imagenet.create_readable_names_for_imagenet_labels()
for i in range(5):
    index = sorted_inds[i]
    print(index)
    # 打印top5的预测类别和相应的概率值。
    print('Probability %0.2f => [%s]' % (probabilities2[index], names[index+1]))





'''https://github.com/tensorflow/models/blob/master/research/slim/train_image_classifier.py'''
def _get_variables_to_train():
    """Returns a list of variables to train.
    Returns:
      A list of variables to train by the optimizer.
    """
    trainable_scopes = 'InceptionResnetV2/Logits,InceptionResnetV2/AuxLogits'

    if trainable_scopes is None:
      return tf.trainable_variables()
    else:
      scopes = [scope.strip() for scope in trainable_scopes.split(',')]

    variables_to_train = []
    for scope in scopes:
      variables = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope)
      variables_to_train.extend(variables)
    return variables_to_train

'''
一些关于inception_resnet_v2变量的测试，在理解模型代码和迁移学习中很有用
'''
exx = tf.trainable_variables()
print(type(exx))
print(exx[0])
print(exx[-1])
print(exx[-2])
print(exx[-3])
print(exx[-4])
print(exx[-5])
print(exx[-6])
print(exx[-7])
print(exx[-8])
print(exx[-9])
print(exx[-10])

print('###############################################################')
variables_to_train = _get_variables_to_train()
print(variables_to_train)

print('###############################################################')
exclude = ['InceptionResnetV2/Logits', 'InceptionResnetV2/AuxLogits']
variables_to_restore = slim.get_variables_to_restore(exclude = exclude)
print(variables_to_restore[0])
print(variables_to_restore[-1])

print('###############################################################')
exclude = ['InceptionResnetV2/Logits']
variables_to_restore = slim.get_variables_to_restore(exclude = exclude)
print(variables_to_restore[0])
print(variables_to_restore[-1])



'''
resnet_v2 152
    num_classes: Number of predicted classes for classification tasks. If None
      we return the features (2048) before the logit layer.

  Returns:
    net: A rank-4 tensor of size [batch, height_out, width_out, channels_out].
      If global_pool is False, then height_out and width_out are reduced by a
      factor of output_stride compared to the respective height_in and width_in,
      else both height_out and width_out equal one. If num_classes is None, then
      net is the output of the last ResNet block, potentially after global
      average pooling. If num_classes is not None, net contains the pre-softmax
      activations.
    end_points: A dictionary from components of the network to the corresponding
      activation.
'''
tf.reset_default_graph()

checkpoint_file = 'resnet_v2_152.ckpt'
image = tf.image.decode_jpeg(tf.read_file('dog.jpeg'), channels=3) #['dog.jpg', 'panda.jpg']

image_size = inception_resnet_v2.default_image_size #  299

'''这个函数做了裁剪，缩放和归一化等'''
processed_image = inception_preprocessing.preprocess_image(image, 
                                                        image_size, 
                                                        image_size,
                                                        is_training=False,)
processed_images  = tf.expand_dims(processed_image, 0)

'''Creates the Resnet V2 model.'''
arg_scope = resnet_arg_scope()
with slim.arg_scope(arg_scope):
    net, end_points = resnet_v2_152(processed_images, 1001, is_training=False)   


probabilities = tf.nn.softmax(net)

saver = tf.train.Saver()


with tf.Session() as sess:
    saver.restore(sess, checkpoint_file)

    net2, image2, network_inputs, end_points2, probabilities2= sess.run([net, 
                                                             image,
                                                       processed_images,
                                                       end_points,