基於tensorflow2.0和cifar100的VGG13網絡訓練

VGG是2014年ILSVRC圖像分類競賽的第二名，相比當年的冠軍GoogleNet在可擴展性方面更勝一籌，此外，它也是從圖像中提取特征的CNN首選算法，VGG的各種網絡模型結構如下：

 

今天代碼的原型是基於VGG13，也就是上圖的B類，可以看到它的參數量是很可觀的。

因為設備和時間問題，網絡並沒有訓練完成，但是已經看到參數變化的效果。（畢竟VGG團隊在最初訓練時使用4塊顯卡並行計算還訓練了2-3周，雖然當今顯卡性能已經有了明顯的提升，但是只能CPU訓練的小可憐實在不敢繼續下去了）

直接上代碼吧

import tensorflow as tf
from tensorflow import keras
import os

os.environ['TF_CPP_MIN_LOG'] = '2'

conv_layers = [
    # part 1
    keras.layers.Conv2D(64,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.Conv2D(64,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.MaxPool2D(pool_size=[2,2],strides=2,padding='same'),

    # part 2
    keras.layers.Conv2D(128,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.Conv2D(128,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.MaxPool2D(pool_size=[2,2],strides=2,padding='same'),

    # part 3
    keras.layers.Conv2D(256,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.Conv2D(256,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.MaxPool2D(pool_size=[2,2],strides=2,padding='same'),

    # part 4
    keras.layers.Conv2D(512,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.Conv2D(512,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.MaxPool2D(pool_size=[2,2],strides=2,padding='same'),

    # part 5
    keras.layers.Conv2D(512,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.Conv2D(512,kernel_size=[3,3],padding='same',activation=tf.nn.relu),
    keras.layers.MaxPool2D(pool_size=[2,2],strides=2,padding='same'),
]

fc_layers =[
    keras.layers.Dense(4096,activation = tf.nn.relu),
    keras.layers.Dense(4096,activation = tf.nn.relu),
    keras.layers.Dense(10)
]


def preprocess(x,y):
    x = tf.cast(x,dtype=tf.float32)/255.
    y = tf.cast(y,dtype=tf.int32)
    return x,y

(x,y),(x_test,y_test) = keras.datasets.cifar100.load_data()
y = tf.squeeze(y,axis=1)
y_test = tf.squeeze(y_test,axis=1)
print(x.shape,y.shape,x_test.shape,y_test.shape)

train_db = tf.data.Dataset.from_tensor_slices((x,y))
train_db = train_db.shuffle(1000).map(preprocess).batch(64)

test_db = tf.data.Dataset.from_tensor_slices((x_test,y_test))
test_db = train_db.map(preprocess).batch(64)

def main():
    conv_net = keras.Sequential(conv_layers)
    conv_net.build(input_shape=[None,32,32,3])
    fc_net = keras.Sequential(fc_layers)
    fc_net.build(input_shape=[None,512])
    optimizer = keras.optimizers.Adam(lr=1e-4)

    for epoch in range(50):
        for step,(x,y) in enumerate(train_db):
            with tf.GradientTape() as tape:
                out = conv_net(x)
                out = tf.reshape(out,[-1,512])
                logits = fc_net(out)
                y_onehot = tf.one_hot(y,depth=10)
                loss = tf.losses.categorical_crossentropy(y_onehot,logits,from_logits=True)
                loss = tf.reduce_mean(loss)

            gradient = tape.gradient(loss,conv_net.trainable_variables + fc_net.trainable)
            optimizer.apply_gradients(zip(gradient,conv_net.trainable_variables + fc_net.trainable))

            if step % 100 == 0:
                print(epoch,step,'loss:',float(loss))

        total_num = 0
        total_correct = 0
        for x,y in test_db:
            out = conv_net(x)
            out = tf.reshape(out,[-1,512])
            logits = fc_net(out)
            prob = tf.nn.softmax(logits,axis=1)
            pred = tf.argmax(prob,axis=1)
            pred = tf.cast(pred,dtype=tf.int32)

            correct = tf.cast(tf.equal(pred,y),dtype=tf.int32)
            correct = tf.reduce_sum(correct)

            total_num += x.shape[0]
            total_correct += correct
            acc = total_correct/total_num

            print("acc:",acc)


if __name__ == '__main__':
    main()

通過這樣一個網絡模型的搭建，確實又加深了我對神經網絡的認識以及tensorflow使用的熟練度，果然上機才是最佳學習方式！