本文借鑑已有cat-vs-dog模型，在此模型上進行修改。該模型可在以下網址下載，後續將對模型進行解析及進一步修改。https://download.csdn.net/download/twinkle_star1314/10414568。今天先對模型進行分析：

一、模型訓練

內容見train.py及dataset.py.

1、資料集

資料集不需另外定義標籤，只需在training_data資料夾下，構建類別資料夾，本文為dogs和cats.

具體為training_data/dogs/dog.0.jpg,dog.1.jpg...........

           training_data/cats/cat.0.jpg,cat.1.jpg...........        

2、train.py解析

import dataset
import tensorflow as tf
import time
from datetime import timedelta
import math
import random
import numpy as np
import os

#Adding Seed so that random initialization is consistent
from numpy.random import seed
seed(1)
from tensorflow import set_random_seed
set_random_seed(2)

batch_size = 32//每次輸入的圖片數目

#Prepare input data
classes = os.listdir('training_data')//獲取分類的總類別
num_classes = len(classes)//本文為2

# 20% of the data will automatically be used for validation
validation_size = 0.2//在training_data總資料集中選取20%作為驗證集
img_size = 128//圖片的大小，相當於caffe中的crop_size
num_channels = 3
train_path='training_data'

# We shall load all the training and validation images and labels into memory using openCV and use that during training
data = dataset.read_train_sets(train_path, img_size, classes, validation_size=validation_size)


print("Complete reading input data. Will Now print a snippet of it")
print("Number of files in Training-set:\t\t{}".format(len(data.train.labels)))
print("Number of files in Validation-set:\t{}".format(len(data.valid.labels)))



session = tf.Session()
x = tf.placeholder(tf.float32, shape=[None, img_size,img_size,num_channels], name='x')

## labels
y_true = tf.placeholder(tf.float32, shape=[None, num_classes], name='y_true')
y_true_cls = tf.argmax(y_true, dimension=1)



##Network graph params
filter_size_conv1 = 3
num_filters_conv1 = 32

filter_size_conv2 = 3
num_filters_conv2 = 32

filter_size_conv3 = 3
num_filters_conv3 = 64
    
fc_layer_size = 128

def create_weights(shape):
    return tf.Variable(tf.truncated_normal(shape, stddev=0.05))//權重大小,初始值

def create_biases(size):
    return tf.Variable(tf.constant(0.05, shape=[size]))



def create_convolutional_layer(input,
               num_input_channels,
               conv_filter_size,        
               num_filters):  
    
    ## We shall define the weights that will be trained using create_weights function.
    weights = create_weights(shape=[conv_filter_size, conv_filter_size, num_input_channels, num_filters])
    ## We create biases using the create_biases function. These are also trained.
    biases = create_biases(num_filters)

    ## Creating the convolutional layer
    layer = tf.nn.conv2d(input=input,
                     filter=weights,
                     strides=[1, 1, 1, 1],
                     padding='SAME')

    layer += biases

    ## We shall be using max-pooling.  

    layer = tf.nn.max_pool(value=layer, //這樣設定pool層特徵圖大小是上一層conv層的一半w2= (上一層大小-核的大小)/步長 +1                        

                            ksize=[1, 2, 2, 1],

                            strides=[1, 2, 2, 1],
                            padding='SAME')
    ## Output of pooling is fed to Relu which is the activation function for us.
    layer = tf.nn.relu(layer)

    return layer

    

def create_flatten_layer(layer):
    #We know that the shape of the layer will be [batch_size img_size img_size num_channels]
    # But let's get it from the previous layer.
    layer_shape = layer.get_shape()

    ## Number of features will be img_height * img_width* num_channels. But we shall calculate it in place of hard-coding it.
    num_features = layer_shape[1:4].num_elements()

    ## Now, we Flatten the layer so we shall have to reshape to num_features
    layer = tf.reshape(layer, [-1, num_features])

    return layer


def create_fc_layer(input,          
             num_inputs,    
             num_outputs,
             use_relu=True):
    
    #Let's define trainable weights and biases.
    weights = create_weights(shape=[num_inputs, num_outputs])
    biases = create_biases(num_outputs)

    # Fully connected layer takes input x and produces wx+b.Since, these are matrices, we use matmul function in Tensorflow
    layer = tf.matmul(input, weights) + biases
    if use_relu:
        layer = tf.nn.relu(layer)

    return layer


layer_conv1 = create_convolutional_layer(input=x,
               num_input_channels=num_channels,
               conv_filter_size=filter_size_conv1,
               num_filters=num_filters_conv1)
layer_conv2 = create_convolutional_layer(input=layer_conv1,
               num_input_channels=num_filters_conv1,
               conv_filter_size=filter_size_conv2,
               num_filters=num_filters_conv2)

layer_conv3= create_convolutional_layer(input=layer_conv2,
               num_input_channels=num_filters_conv2,
               conv_filter_size=filter_size_conv3,
               num_filters=num_filters_conv3)
          
layer_flat = create_flatten_layer(layer_conv3)

layer_fc1 = create_fc_layer(input=layer_flat,
                     num_inputs=layer_flat.get_shape()[1:4].num_elements(),
                     num_outputs=fc_layer_size,
                     use_relu=True)

layer_fc2 = create_fc_layer(input=layer_fc1,
                     num_inputs=fc_layer_size,
                     num_outputs=num_classes,
                     use_relu=False)

y_pred = tf.nn.softmax(layer_fc2,name='y_pred')

y_pred_cls = tf.argmax(y_pred, dimension=1)
session.run(tf.global_variables_initializer())
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=layer_fc2,
                                                    labels=y_true)
cost = tf.reduce_mean(cross_entropy)
optimizer = tf.train.AdamOptimizer(learning_rate=1e-4).minimize(cost)//學習率
correct_prediction = tf.equal(y_pred_cls, y_true_cls)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))


session.run(tf.global_variables_initializer())


def show_progress(epoch, feed_dict_train, feed_dict_validate, val_loss):
    acc = session.run(accuracy, feed_dict=feed_dict_train)
    val_acc = session.run(accuracy, feed_dict=feed_dict_validate)
    msg = "Training Epoch {0} --- Training Accuracy: {1:>6.1%}, Validation Accuracy: {2:>6.1%},  Validation Loss: {3:.3f}"
    print(msg.format(epoch + 1, acc, val_acc, val_loss))

total_iterations = 0

saver = tf.train.Saver()
def train(num_iteration)://總迭代數相當於caffe中max_iter
    global total_iterations
    
    for i in range(total_iterations,
                   total_iterations + num_iteration):

        x_batch, y_true_batch, _, cls_batch = data.train.next_batch(batch_size)
        x_valid_batch, y_valid_batch, _, valid_cls_batch = data.valid.next_batch(batch_size)

        
        feed_dict_tr = {x: x_batch,
                           y_true: y_true_batch}
        feed_dict_val = {x: x_valid_batch,
                              y_true: y_valid_batch}

        session.run(optimizer, feed_dict=feed_dict_tr)

        if i % int(data.train.num_examples/batch_size) == 0:
            val_loss = session.run(cost, feed_dict=feed_dict_val)
            epoch = int(i / int(data.train.num_examples/batch_size))    
            
            show_progress(epoch, feed_dict_tr, feed_dict_val, val_loss)
            saver.save(session, 'dogs-cats-model')


    total_iterations += num_iteration

train(num_iteration=3000)#總迭代數相當於caffe中max_iter

2 dataset.py

import cv2
import os
import glob
from sklearn.utils import shuffle
import numpy as np
 
 
def load_train(train_path, image_size, classes):
    images = []
    labels = []
    img_names = []
    cls = []
 
    print('Going to read training images')
    for fields in classes:    
        index = classes.index(fields)
        print('Now going to read {} files (Index: {})'.format(fields, index))
        path = os.path.join(train_path, fields, '*g')
        files = glob.glob(path)
        for fl in files:
            image = cv2.imread(fl)
            image = cv2.resize(image, (image_size, image_size),0,0, cv2.INTER_LINEAR)
            image = image.astype(np.float32)
            image = np.multiply(image, 1.0 / 255.0)
            images.append(image)
            label = np.zeros(len(classes))
            label[index] = 1.0
            labels.append(label)
            flbase = os.path.basename(fl)
            img_names.append(flbase)
            cls.append(fields)
    images = np.array(images)
    labels = np.array(labels)
    img_names = np.array(img_names)
    cls = np.array(cls)
 
    return images, labels, img_names, cls
 
 
class DataSet(object):
 
  def __init__(self, images, labels, img_names, cls):
    self._num_examples = images.shape[0]
 
    self._images = images
    self._labels = labels
    self._img_names = img_names
    self._cls = cls
    self._epochs_done = 0
    self._index_in_epoch = 0
 
  @property
  def images(self):
    return self._images
 
  @property
  def labels(self):
    return self._labels
 
  @property
  def img_names(self):
    return self._img_names
 
  @property
  def cls(self):
    return self._cls
 
  @property
  def num_examples(self):
    return self._num_examples
 
  @property
  def epochs_done(self):
    return self._epochs_done
 
  def next_batch(self, batch_size):
    """Return the next `batch_size` examples from this data set."""
    start = self._index_in_epoch
    self._index_in_epoch += batch_size
 
    if self._index_in_epoch > self._num_examples:
      # After each epoch we update this
      self._epochs_done += 1
      start = 0
      self._index_in_epoch = batch_size
      assert batch_size <= self._num_examples
    end = self._index_in_epoch
 
    return self._images[start:end], self._labels[start:end], self._img_names[start:end], self._cls[start:end]
 
 
def read_train_sets(train_path, image_size, classes, validation_size):
  class DataSets(object):
    pass
  data_sets = DataSets()
 
  images, labels, img_names, cls = load_train(train_path, image_size, classes)
  images, labels, img_names, cls = shuffle(images, labels, img_names, cls)   
 
  if isinstance(validation_size, float):
    validation_size = int(validation_size * images.shape[0])
 
  validation_images = images[:validation_size]
  validation_labels = labels[:validation_size]
  validation_img_names = img_names[:validation_size]
  validation_cls = cls[:validation_size]
 
  train_images = images[validation_size:]
  train_labels = labels[validation_size:]
  train_img_names = img_names[validation_size:]
  train_cls = cls[validation_size:]
 
  data_sets.train = DataSet(train_images, train_labels, train_img_names, train_cls)
  data_sets.valid = DataSet(validation_images, validation_labels, validation_img_names, validation_cls)
 
  return data_sets