TensorFlow 卷積神經網路系列案例（1）：貓狗識別

TensorFlow 卷積神經網路系列案例（1）：貓狗識別

TensorFlow 系列案例（2）：自然語言處理-TensorFlow + Word2Vec https://blog.csdn.net/duan_zhihua/article/details/81257323

使用卷積神經網路進行貓狗識別的步驟： 
1. 載入貓狗訓練圖片資料。 
2. 建立卷積神經網路層次。 
3. 對貓狗圖片進行訓練測試，儲存模型。
4. 載入訓練好的模型，對新的圖片進行貓、狗預測。

本節案例程式碼來源於網上資料，在此致謝網上各位AI大牛的貢獻！

•  資料處理：載入貓狗訓練圖片資料。 

      資料集來源於Kaggle，這裡從資料集中選取整理了1000只貓和1000只狗的圖片，圖片放在training_data目錄下，training_data目錄包括cats目錄和dogs目錄。連結：https://pan.baidu.com/s/124AY6eN2580eTWmryQ1-_A 密碼：9uqc

dataset.py程式碼：先安裝OpenCV、Pillow第三方模組。在程式碼中匯入cv2。

import numpy as np
import os
import glob
from sklearn.utils import shuffle
import cv2


def load_train(train_path, img_size, classes):
    images = []
    labels = []
    img_names = []
    cls = []
    print("讀取訓練圖片...")
    #classes傳入一個列表：<class 'list'>: ['dogs', 'cats']
    for fields in classes:
        index = classes.index(fields)
        print("Now going to read {} files (Index:{})".format(fields, index))
        #路徑讀入格式：'D:/PycharmProjects/Tensorflow_2018_test/catAndDog/training_data\\dogs\\*g'
        path = os.path.join(train_path, fields, '*g')
        #file內容：'D:/PycharmProjects/Tensorflow_2018_test/catAndDog/training_data\\dogs\\dog.0.jpg'
        files = glob.glob(path)
        print(files)
        for fl in files:
            print(fl)
            #img_size 圖片的大小（如 64），原始圖片可能是大圖片，也可能是小圖片，以下轉換為統一格式。
            image = cv2.imread(fl)
            #轉換為大小：<class 'tuple'>: (64, 64, 3)，這裡是彩色圖片RGB，通道數為3.
            image = cv2.resize(image, (img_size, img_size), 0, 0, cv2.INTER_LINEAR)
            image = image.astype(np.float32)
            # 歸一化處理，將資料乘以 1/255，轉換為（0，1）之間的範圍。
            image = np.multiply(image, 1.0 / 255.0)
            images.append(image)
            label = np.zeros(len(classes))
            # 貓狗二分類打標籤，如[1. 0.]
            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

    def images(self):
        return self._images

    def labels(self):
        return self._labels

    def img_names(self):
        return self._img_names

    def cls(self):
        return self._cls

    def num_examples(self):
        return self._num_examples

    def epochs_done(self):
        return self._epochs_done

    def next_batch(self, batch_size):
        start = self._index_in_epoch
        self._index_in_epoch += batch_size

        if self._index_in_epoch > self._num_examples:
            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)
    #呼叫sklearn.utils的 shuffle方法，打散貓狗圖片資料。
    images, labels, img_names, cls = shuffle(images, labels, img_names, cls)
    # 這裡讀入了2002張貓狗圖片，validation_size等於0.2，因此驗證集validation_size 為400個
    #images：<class 'tuple'>: (2002, 64, 64, 3)
    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

• 建立卷積神經網路層次。卷積神經網路（CNN）由輸入層、卷積層（啟用函式）、池化層、全連線層（神經網路的隱藏層）組成，即INPUT（輸入層）-CONV（卷積層 -RELU啟用函式）-POOL（池化層）-FC（全連線層）。在每一層轉換中需清楚shape的轉換，以下是卷積神經網路輸出size大小的計算：

卷積計算輸出的大小：
out_height =((input_height - filter_height + padding_top + padding_bottom) /stride_height) +1
out_width =((input_width - filter_width + padding_left + padding_right) /stride_width) +1

但我們通常都使用方陣：
out_height = out_width
input_height = input_width
filter_height filter_width 
padding_top =padding_bottom=padding_left=  padding_right
stride_height=stride_width

1）因此，如果padding 不等於“SAME”，
輸入圖片大小：W*W
filter核心大小：F*F
步長 S
padding畫素數P
卷積計算輸出的大小簡化為：N= （W-F+2P)/S +1

2）如果padding =“SAME”：
卷積計算公式輸出大小簡化為：W/S 向上取整。

計算例子：

① padding = "value", stride = 4, (227 - 11 + 2*0）/ 4 + 1 = 55

② padding = "value", stride = 2, (55 - 3 + 2*0）/  2 + 1 = 27

③ padding = "same", stride = 1, 27 / 1 = 27

④ padding = "value",stride = 2, (27 - 3 + 2*0) / 2 + 1 = 13

本案例中的計算示例：layer_conv1 的卷積計算，從輸入層到卷積層。

layer_conv1 = create_convolution_layer(input=x,
                                       num_input_channels=num_channels,
                                       conv_filter_size=filter_size_conv1,
                                       num_filters=num_filters_conv1)

def create_convolution_layer(input,
                             num_input_channels,
                             conv_filter_size,
                             num_filters):
    weights = create_weights(shape=[conv_filter_size, conv_filter_size, num_input_channels, num_filters])
    biases = create_biases(num_filters)
    layer = tf.nn.conv2d(input=input, filter=weights, strides=[1, 1, 1, 1], padding='SAME')

    layer += biases
    layer = tf.nn.relu(layer)
    layer = tf.nn.max_pool(value=layer, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

    return layer

 

layer_conv1 的卷積計算，從輸入層到卷積層。
在程式碼layer = tf.nn.conv2d(input=input, filter=weights, strides=[1, 1, 1, 1], padding='SAME')
tf.nn.conv2d方法的輸入tensor的shape形狀： `[batch, in_height, in_width, in_channels]`
tf.nn.conv2d方法的filter核心tensor 的shape形狀：`[filter_height, filter_width, in_channels, out_channels]`

 其中：
input=x 的shape：batchsize 64 64 3
weights核心的shape：  3 3 3 32 這裡的第三個引數3是 num_input_channels，需等於輸入層的通道數3 
stride步長：  1 1 1 1，這裡的strides[0] = strides[3] 需等於1
卷積計算輸出大小W/S=64/1=64 ，其shape為：  batchsize 64 64 32

這裡卷積層的池化計算中：
layer = tf.nn.max_pool(value=layer, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
輸入為layer，其shape為：  batchsize 64 64 32
輸出大小為W/S=64/2=32，其shape為：batchsize 32 32 32

其他的卷積層和池化層的大小計算依此類推。

資料訓練的整個程式碼train.py：

import  dataset
import tensorflow as tf
import numpy as np
from numpy.random import seed

seed(10)
from tensorflow import set_random_seed

set_random_seed(20)

batch_size = 32
classes = ['dogs', 'cats']
num_classes = len(classes)

validation_size = 0.2
img_size = 64
num_channels = 3
train_path = "D:/PycharmProjects/Tensorflow_2018_test/catAndDog/training_data"
data = dataset.read_train_sets(train_path, img_size, classes, validation_size)

session = tf.Session()
x = tf.placeholder(tf.float32, shape=[None, img_size, img_size, num_channels], name='x')
y_true = tf.placeholder(tf.float32, shape=[None, num_classes], name='y_true')
y_true_cls = tf.argmax(y_true, dimension=1)
filter_size_conv1 = 3
num_filters_conv1 = 32

filter_size_conv2 = 3
num_filters_conv2 = 32

filter_size_conv3 = 3
num_filters_conv3 = 64

# 全連線層的輸出
fc_layer_size = 1024


def create_weights(shape):
    return tf.Variable(tf.truncated_normal(shape, stddev=0.05))


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


def create_convolution_layer(input,
                             num_input_channels,
                             conv_filter_size,
                             num_filters):
    weights = create_weights(shape=[conv_filter_size, conv_filter_size, num_input_channels, num_filters])
    biases = create_biases(num_filters)
    layer = tf.nn.conv2d(input=input, filter=weights, strides=[1, 1, 1, 1], padding='SAME')

    layer += biases
    layer = tf.nn.relu(layer)
    layer = tf.nn.max_pool(value=layer, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

    return layer


def create_flatten_layer(layer):
    layer_shape = layer.get_shape()
    num_features = layer_shape[1:4].num_elements()
    layer = tf.reshape(layer, [-1, num_features])
    return layer


def create_fc_layer(input,
                    num_inputs,
                    num_outputs,
                    use_relu=True):
    weights = create_weights(shape=[num_inputs, num_outputs])
    biases = create_biases(num_outputs)

    layer = tf.matmul(input, weights) + biases
    layer = tf.nn.dropout(layer, keep_prob=0.7)
    if use_relu:
        layer = tf.nn.relu(layer)
    return layer

layer_conv1 = create_convolution_layer(input=x,
                                       num_input_channels=num_channels,
                                       conv_filter_size=filter_size_conv1,
                                       num_filters=num_filters_conv1)

layer_conv2 = create_convolution_layer(input=layer_conv1,
                                       num_input_channels=num_filters_conv1,
                                       conv_filter_size=filter_size_conv2,
                                       num_filters=num_filters_conv2)
layer_conv3 = create_convolution_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)
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, i):
    acc = session.run(accuracy, feed_dict=feed_dict_train)
    val_acc = session.run(accuracy, feed_dict=feed_dict_validate)
    print("epoch:", str(epoch + 1) + ",i:", str(i) +
          ",acc:", str(acc) + ",val_acc:", str(val_acc) + ",val_loss:", str(val_loss))


total_iterations = 0
saver = tf.train.Saver()


def train(num_iteration):
    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)
        examples = data.train.num_examples()
        if i % int(examples / batch_size) == 0:
            val_loss = session.run(cost, feed_dict=feed_dict_val)
            epoch = int(i / int(examples / batch_size))

            show_progress(epoch, feed_dict_tr, feed_dict_val, val_loss, i)
            saver.save(session, './dogs-cats-model/dog-cat.ckpt', global_step=i)
    total_iterations += num_iteration


train(num_iteration=100)

train(num_iteration=5000) 執行結果如下：

......
D:/PycharmProjects/Tensorflow_2018_test/catAndDog/training_data\cats\cat.998.jpg
D:/PycharmProjects/Tensorflow_2018_test/catAndDog/training_data\cats\cat.999.jpg
2018-07-22 19:47:35.412691: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 19:47:35.413146: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE2 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 19:47:35.413607: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE3 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 19:47:35.414036: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 19:47:35.414523: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 19:47:35.416029: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 19:47:35.416710: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 19:47:35.417239: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
epoch: 1,i: 0,acc: 0.625,val_acc: 0.4375,val_loss: 0.7436454
epoch: 2,i: 50,acc: 0.65625,val_acc: 0.40625,val_loss: 0.7284967
epoch: 3,i: 100,acc: 0.5625,val_acc: 0.40625,val_loss: 0.7191855
epoch: 4,i: 150,acc: 0.5625,val_acc: 0.5,val_loss: 0.7015336
epoch: 5,i: 200,acc: 0.625,val_acc: 0.59375,val_loss: 0.6757284
epoch: 6,i: 250,acc: 0.5625,val_acc: 0.5625,val_loss: 0.7258785
epoch: 7,i: 300,acc: 0.6875,val_acc: 0.5625,val_loss: 0.6614233
epoch: 8,i: 350,acc: 0.59375,val_acc: 0.59375,val_loss: 0.605597
epoch: 9,i: 400,acc: 0.625,val_acc: 0.53125,val_loss: 0.67095697
epoch: 10,i: 450,acc: 0.625,val_acc: 0.5,val_loss: 0.66302484
epoch: 11,i: 500,acc: 0.65625,val_acc: 0.53125,val_loss: 0.66088706
epoch: 12,i: 550,acc: 0.6875,val_acc: 0.59375,val_loss: 0.5763863
epoch: 13,i: 600,acc: 0.625,val_acc: 0.625,val_loss: 0.68232787
epoch: 14,i: 650,acc: 0.8125,val_acc: 0.78125,val_loss: 0.5594511
epoch: 15,i: 700,acc: 0.71875,val_acc: 0.75,val_loss: 0.6684464
epoch: 16,i: 750,acc: 0.75,val_acc: 0.59375,val_loss: 0.64405406
epoch: 17,i: 800,acc: 0.75,val_acc: 0.53125,val_loss: 0.63719827
epoch: 18,i: 850,acc: 0.71875,val_acc: 0.6875,val_loss: 0.5434316
epoch: 19,i: 900,acc: 0.78125,val_acc: 0.65625,val_loss: 0.7148928
epoch: 20,i: 950,acc: 0.75,val_acc: 0.78125,val_loss: 0.57082707
epoch: 21,i: 1000,acc: 0.8125,val_acc: 0.65625,val_loss: 0.63966876
epoch: 22,i: 1050,acc: 0.75,val_acc: 0.75,val_loss: 0.5933325
epoch: 23,i: 1100,acc: 0.75,val_acc: 0.625,val_loss: 0.6802487
epoch: 24,i: 1150,acc: 0.75,val_acc: 0.625,val_loss: 0.6392723
epoch: 25,i: 1200,acc: 0.78125,val_acc: 0.625,val_loss: 0.6674799
epoch: 26,i: 1250,acc: 0.78125,val_acc: 0.75,val_loss: 0.59081894
epoch: 27,i: 1300,acc: 0.71875,val_acc: 0.5,val_loss: 0.67688525
epoch: 28,i: 1350,acc: 0.8125,val_acc: 0.59375,val_loss: 0.6042088
epoch: 29,i: 1400,acc: 0.78125,val_acc: 0.53125,val_loss: 0.70344555
epoch: 30,i: 1450,acc: 0.75,val_acc: 0.75,val_loss: 0.5876309
epoch: 31,i: 1500,acc: 0.875,val_acc: 0.71875,val_loss: 0.6333902
......

執行結果將訓練好的模型進行持久化。

• 載入訓練好的模型，對新的圖片進行貓、狗預測。這裡仍使用training_data目錄的資料來進行圖片預測。

predict.py

import glob

import tensorflow as tf
import numpy as np
import os, cv2

image_size = 64
num_channels = 3
images = []

path = "D:/PycharmProjects/Tensorflow_2018_test/catAndDog/training_data"
direct = os.listdir(path)
for file in direct:
    path = os.path.join(path, file, '*g')
    files = glob.glob(path  )
    print(files)
    for fl in files:
         print(fl)
         image = cv2.imread(fl)
         image = cv2.resize(image, (image_size, image_size), 0, 0, cv2.INTER_LINEAR)
         images.append(image)

images = np.array(images, dtype=np.uint8)
images = images.astype('float32')
images = np.multiply(images, 1.0 / 255.0)

for img in images:
    x_batch = img.reshape(1, image_size, image_size, num_channels)

    sess = tf.Session()

    # step1網路結構圖
    saver = tf.train.import_meta_graph('./dogs-cats-model/dog-cat.ckpt-3050.meta')

    # step2載入權重引數
    saver.restore(sess, './dogs-cats-model/dog-cat.ckpt-3050')

    # 獲取預設的圖
    graph = tf.get_default_graph()

    y_pred = graph.get_tensor_by_name("y_pred:0")

    x = graph.get_tensor_by_name("x:0")
    y_true = graph.get_tensor_by_name("y_true:0")
    y_test_images = np.zeros((1, 2))

    feed_dict_testing = {x: x_batch, y_true: y_test_images}
    result = sess.run(y_pred, feed_dict_testing)

    res_label = ['dog', 'cat']
    print(res_label[result.argmax()])

預測結果如下：

......
2018-07-22 20:15:19.835048: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 20:15:19.835668: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE2 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 20:15:19.836264: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE3 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 20:15:19.836836: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 20:15:19.837650: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 20:15:19.838874: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 20:15:19.839484: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2018-07-22 20:15:19.906125: W c:\l\tensorflow_1501918863922\work\tensorflow-1.2.1\tensorflow\core\platform\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
cat
cat
cat
......

 

“我設想在未來，我們可能就相當於機器人的寵物狗狗，到那時我也會支援機器人的。”  ——克勞德·夏農

“全面化人工智慧可能意味著人類的終結...”機器可以自行啟動，並且自動對自身進行重新設計，速率也會越來越快。受到漫長的生物進化歷程的限制，人類無法與之競爭，終將被取代。“ ——史蒂芬·霍金