總的流程看這兩張圖就差不多了
參考：
https://blog.csdn.net/weixin_43624538/article/details/83988998
https://blog.csdn.net/qq_28123095/article/details/79787782
代碼如下：
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
class AlexNet():
def __init__(self):
self.learning_rate = 0.001
self.training_iters = 10000 # 我這裏只迭代一萬次
self.batch_size = 64 # 每個batch的大小
self.display_step = 20 # 每20步展示一下結果
# 定義網絡參數
self.n_input = 784 # 輸入的維度
self.n_classes = 10 # 標籤的維度
self.dropout = 0.8 # Dropout 的概率
def conv2d(self, x, filter, k_size, stride=[1, 1], padding='same', dilation=[1, 1], activation=tf.nn.relu,
scope='conv2d'): # dilation 卷積層膨脹
return tf.layers.conv2d(inputs=x, filters=filter, kernel_size=k_size,
strides=stride, dilation_rate=dilation, padding=padding,
name=scope, activation=activation)
# def max_pool2d(self, x, pool_size=[2, 2], stride=[2, 2], scope='max_pool2d'):
# return tf.layers.max_pooling2d(inputs=x, pool_size=pool_size, strides=stride, name=scope)
def max_pool(self,name, l_input, k):
return tf.nn.max_pool(l_input, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME', name=name)
def dropoutx(self, x, d_rate=0.5):
return tf.layers.dropout(inputs=x, rate=d_rate)
def norm(self, x, l_size, bias=1.0, alpha=0.001 / 9.0, beta=0.75, scope='norm'):
return tf.nn.lrn(x, l_size, bias=bias, alpha=alpha, beta=beta, name=scope)
def set_net(self, x):
check_points = {}
x = tf.reshape(x, shape=[-1, 28, 28, 1])
net = self.conv2d(x, filter=64, k_size=[3, 3], stride=[1,1], scope='conv1_1')
#net = self.max_pool2d(net, pool_size=[2, 2], stride=[2,2], scope='pool1')
net = self.max_pool('pool1', net, k=2)
net = self.norm(net, l_size=4, scope='norm_1')
net = self.dropoutx(net, d_rate=self.dropout)
check_points['block1'] = net
net = self.conv2d(net, filter=128, k_size=[3, 3], stride=[1,1], scope='conv2_1')
# net = self.max_pool2d(net, pool_size=[2, 2], stride=[2,2], scope='pool2')
net = self.max_pool('pool2', net, k=2)
net = self.norm(net, l_size=4, scope='norm_2')
net = self.dropoutx(net, d_rate=self.dropout)
check_points['block2'] = net
net = self.conv2d(net, filter=256, k_size=[3, 3], stride=[1,1], scope='conv5_1')
#net = self.max_pool2d(net, pool_size=[2, 2], stride=[2,2], scope='pool3')
net = self.max_pool('pool3', net, k=2)
net = self.norm(net, l_size=4, scope='norm_3')
net = self.dropoutx(net, d_rate=self.dropout)
check_points['block3'] = net
net = tf.reshape(net, [-1, 4 * 4 * 256])
net = tf.layers.dense(net, units=1024, activation=tf.nn.relu, use_bias=True)
net = tf.layers.dense(net, units=1024, activation=tf.nn.relu, use_bias=True)
net = tf.layers.dense(net, units=10, activation=tf.nn.relu, use_bias=True)
check_points['block4'] = net
return net
def alex_prediction(self, mnist, scope='alexnet'):
with tf.name_scope('inputs'):
x = tf.placeholder(tf.float32, [None, self.n_input])
y = tf.placeholder(tf.float32, [None, self.n_classes])
pred = self.set_net(x) # pred是計算完的值，此時還沒歸一化
a = tf.nn.softmax(pred) # a是歸一化後的值。
# 定義損失函數和學習步驟
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y)) # 這個是損失loss
optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(cost) # 最小化loss
# 測試網絡
correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
# 初始化所有的共享變量
init = tf.initialize_all_variables()
with tf.Session() as sess:
sess.run(init)
step = 1
writer = tf.summary.FileWriter("E://TensorBoard//test", sess.graph)
# Keep training until reach max iterations
while step * self.batch_size < self.training_iters: # 直到達到最大迭代次數，沒考慮梯度！！！
batch_xs, batch_ys = mnist.train.next_batch(self.batch_size)
# 獲取批數據
sess.run(optimizer, feed_dict={x: batch_xs, y: batch_ys})
if step % self.display_step == 0: # 每一步裏有64batch，64*20=1280
# 計算精度
acc = sess.run(accuracy, feed_dict={x: batch_xs, y: batch_ys})
# 計算損失值
loss = sess.run(cost, feed_dict={x: batch_xs, y: batch_ys})
print("Iter " + str(step * self.batch_size) + ", Minibatch Loss= " + "{:.6f}".format(
loss) + ", Training Accuracy = " + "{:.5f}".format(acc))
step += 1
print("Optimization Finished!")
# 計算測試精度
print("Testing Accuracy:", sess.run(accuracy,
feed_dict={x: mnist.test.images[:256], y: mnist.test.labels[:256],
})) # 拿前256個來測試
print("Testing Result:", sess.run(a, feed_dict={x: mnist.test.images[63:64], y: mnist.test.labels[63:64],
})) # 數組範圍，從0開始，含左不含右
writer.close()
if __name__ == '__main__':
alex = AlexNet()
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
alex.alex_prediction(mnist)
#{'a': <tf.Tensor 'dropout/mul:0' shape=(?, 14, 14, 64) dtype=float32>, 'b': <tf.Tensor 'dropout_1/mul:0' shape=(?, 7, 7, 128) dtype=float32>, 'c': <tf.Tensor 'dropout_2/mul:0' shape=(?, 4, 4, 256) dtype=float32>, 'd': <tf.Tensor 'fc2:0' shape=(?, 1024) dtype=float32>, 'e': <tf.Tensor 'add_2:0' shape=(?, 10) dtype=float32>}
#{'block1': <tf.Tensor 'norm_1:0' shape=(?, 14, 14, 64) dtype=float32>, 'block2': <tf.Tensor 'norm_2:0' shape=(?, 7, 7, 128) dtype=float32>, 'block3': <tf.Tensor 'norm_3:0' shape=(?, 3, 3, 256) dtype=float32>, 'block4': <tf.Tensor 'fc2:0' shape=(?, 1024) dtype=float32>, 'block5': <tf.Tensor 'add_2:0' shape=(?, 10) dtype=float32>}
使用tf.layers.max_pooling2d的時候會產生下面那個shape（3,3,256）的情形導致out跟y的shape不匹配
測試代碼：
import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
weights = {
'wd1': tf.Variable(tf.random_normal([4 * 4 * 256, 1024])),
'wd2': tf.Variable(tf.random_normal([1024, 1024])),
'out': tf.Variable(tf.random_normal([1024, 10]))
}
biases = {
'bd1': tf.Variable(tf.random_normal([1024])),
'bd2': tf.Variable(tf.random_normal([1024])),
'out': tf.Variable(tf.random_normal([10]))
}
def conv2d( x, filter, k_size, stride=[1, 1], padding='same', dilation=[1, 1], activation=tf.nn.relu,
scope='conv2d'): # dilation 卷積層膨脹
return tf.layers.conv2d(inputs=x, filters=filter, kernel_size=k_size,
strides=stride, dilation_rate=dilation, padding=padding,
name=scope, activation=activation)
def max_pool( name, l_input, k):
return tf.nn.max_pool(l_input, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME', name=name)
def dropoutx( x, d_rate=0.5):
return tf.layers.dropout(inputs=x, rate=d_rate)
def norm( x, l_size, bias=1.0, alpha=0.001 / 9.0, beta=0.75, scope='norm'):
return tf.nn.lrn(x, l_size, bias=bias, alpha=alpha, beta=beta, name=scope)
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
x = tf.placeholder('float',[None,784])
net = tf.reshape(x, shape=[-1, 28, 28, 1])
net = conv2d(net, filter=64, k_size=[3, 3], stride=[1, 1], scope='conv1_1')
# net = max_pool2d(net, pool_size=[2, 2], stride=[2,2], scope='pool1')
net = max_pool('pool1', net, k=2)
net_1 = norm(net, l_size=4, scope='norm_1')
net = conv2d(net_1, filter=128, k_size=[3, 3], stride=[1, 1], scope='conv2_1')
# net = max_pool2d(net, pool_size=[2, 2], stride=[2,2], scope='pool2')
net = max_pool('pool2', net, k=2)
net_2 = norm(net, l_size=4, scope='norm_2')
net = conv2d(net_2, filter=256, k_size=[3, 3], stride=[1, 1], scope='conv5_1')
# net = max_pool2d(net, pool_size=[2, 2], stride=[2,2], scope='pool3')
net = max_pool('pool3', net, k=3)
net_3 = norm(net, l_size=4, scope='norm_3')
net_final_1 = tf.reshape(net_3, [-1, weights['wd1'].get_shape().as_list()[0]])
net_final_2 = tf.nn.relu(tf.matmul(net_final_1,weights['wd1']) + biases['bd1'], name='fc1')
net_final_3 = tf.nn.relu(tf.matmul(net_final_2,weights['wd2']) + biases['bd2'], name='fc2') # Relu activation
out = tf.matmul(net_final_3, weights['out']) + biases['out']
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
batch_xs, batch_ys = mnist.train.next_batch(64)
net1 = sess.run(net_1,feed_dict={x:batch_xs})
print(net1.shape)
net2 = sess.run(net_2,feed_dict={x:batch_xs})
print(net2.shape)
net3 = sess.run(net_3,feed_dict={x:batch_xs})
print(net3.shape)
result_1 = sess.run(net_final_1,feed_dict={x:batch_xs})
print(result_1.shape)
result_2 = sess.run(net_final_2,feed_dict={x:batch_xs})
print(result_2.shape)
result_3 = sess.run(net_final_3,feed_dict={x:batch_xs})
print(result_3.shape)
out = sess.run(out,feed_dict={x:batch_xs})
print(out.shape)
剛開始嘗試用tf寫，太菜了。。。