带张量流的自动编码器

Question

我尝试使用tensorflow实现Stacked autoencoder。我使用了mnist数据集，并尝试将维数从784减小到2。我已经使用keras进行了处理，其结果很好（火车误差几乎为0.04）。但是，使用此tesorflow代码时效果不佳（火车误差几乎为0.4）。我不知道为什么这些结果如此不同。你能告诉我如何修改吗？

from keras.datasets import mnist
import tensorflow as tf


(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')

x_train = x_train.reshape(60000,28*28)
x_test = x_test.reshape(10000,28*28)

x_train  = x_train/255.0
x_test = x_test/255.0

x_ = tf.placeholder('float32', [None,784])



#encoder
w1 = tf.Variable(tf.random_normal([784,512]))
b1 = tf.Variable(tf.random_normal([512]))
h1 = tf.nn.relu(tf.matmul(x_, w1) + b1)

w2 = tf.Variable(tf.random_normal([512,128]))
b2 = tf.Variable(tf.random_normal([128]))
h2 = tf.nn.relu(tf.matmul(h1, w2) + b2)

w0 = tf.Variable(tf.random_normal([128,2]))
b0 = tf.Variable(tf.random_normal([2]))
h0 = tf.matmul(h2, w0) + b0



#decoder
w_1 = tf.Variable(tf.random_normal([2,128]))
b_1 = tf.Variable(tf.random_normal([128]))
h_1 = tf.nn.relu(tf.matmul(h0, w_1) + b_1)

w_2 = tf.Variable(tf.random_normal([128,512]))
b_2 = tf.Variable(tf.random_normal([512]))
h_2 = tf.nn.relu(tf.matmul(h_1, w_2) + b_2)

w_0 = tf.Variable(tf.random_normal([512,784]))
b_0 = tf.Variable(tf.random_normal([784]))
h_0 = tf.nn.sigmoid(tf.matmul(h_2, w_0) + b_0)


#cost
cost = tf.reduce_mean(tf.pow(x_-h_0,2))
train_step = tf.train.AdamOptimizer(0.05).minimize(cost)
init = tf.global_variables_initializer()


batch_count = int(x_train.shape[0]/256)

sess = tf.Session()
sess.run(init)
for i in range(2) :
    total_cost = 0
    for j in range(batch_count) :
        batch_xs= x_train[i*256:i*256 + 256,:]
        _, cost_val = sess.run([train_step, cost], feed_dict = {x_ : batch_xs})
        total_cost = total_cost + cost_val
        if j % 20 == 0 :
            print('epoch : %d ,%d / %d , loss : %f , average_loss : %f' %(i+1, j+1, batch_count, cost_val, total_cost/(j+1)))


hidden1, hidden2, encoder = sess.run([h1,h2,h0],feed_dict={x_ : x_test})

Answer 1

您的最后一个编码器步骤没有激活h0 = tf.matmul(h2, w0) + b0，因此在此处进行relu激活（即替换）后，损耗保持在0.48。

h0 = tf.matmul(h2, w0) + b0与h0 = tf.nn.relu(tf.matmul(h2, w0) + b0)，损失仅在两个时期内下降到0.06。

我还更改了在其中显示输出以适合图像中更多数据的频率。

如果还有其他疑问，请在下面评论。