我试图实现一个卷积自动编码器，该编码器以人脸作为输入，从中提取特征，并使用这些特征的反卷积返回图像。该程序是用TensorFlow编写的，并在GeForce GTX1080 GPU上运行。输出是纯谷物（https://imgur.com/a/KxxQUco-输出图像和脸部图像）。以下是相关功能：
def res_block(self, inp, num, in_c, mid_c, f, t):
with tf.variable_scope(t+'_res_'+str(num)):
w1 = tf.Variable(tf.truncated_normal(shape=[f, f, in_c, mid_c], stddev=0.1),
name=t+'_res_w1_'+str(num))
conv1 = tf.nn.conv2d(inp,
filter=w1,
strides=(1,1,1,1),
padding="SAME",
name=t+'_res_conv1_'+str(num))
b1 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv1), stddev=0.1),
name=t+'_res_b1_'+str(num))
conv1_ = tf.nn.relu(conv1+b1, name=t+'_res_conv1_'+str(num)+'_')
w2 = tf.Variable(tf.truncated_normal(shape=[f, f, mid_c, in_c], stddev=0.1),
name=t+'_res_w2_'+str(num))
conv2 = tf.nn.conv2d(conv1_,
filter=w2,
strides=(1,1,1,1),
padding="SAME",
name=t+'_res_conv2_'+str(num))
conv_out = tf.add(inp, conv2)
return conv_out
def encoder(self, image, reuse_in):
# Start of feature extraction network
with tf.variable_scope('e_conv'):
w1 = tf.Variable(tf.truncated_normal(shape=[5, 5, 3, 40], stddev=1.0),
name='e_w1')
conv1 = tf.nn.conv2d(image,
filter=w1,
strides=(1,2,2,1),
padding="VALID",
name='e_conv1')
b1 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv1), stddev=0.1),
name='e_b1')
conv1_ = tf.nn.relu(conv1+b1, name='e_conv1_')
res11 = self.res_block(conv1_, 1, 40, 40, 5, 'e')
res12 = self.res_block(res11, 2, 40, 40, 5, 'e')
with tf.variable_scope('e_conv'):
w2 = tf.Variable(tf.truncated_normal(shape=[5, 5, 40, 64], stddev=0.1),
name='e_w2')
conv2 = tf.nn.conv2d(res12,
filter=w2,
strides=(1,2,2,1),
padding="VALID",
name='e_conv2')
b2 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv2), stddev=0.1),
name='e_b2')
conv2_ = tf.nn.relu(conv2+b2, name='e_conv2_')
res21 = self.res_block(conv2_, 3, 64, 64, 5, 'e')
res22 = self.res_block(res21, 4, 64, 64, 5, 'e')
with tf.variable_scope('e_conv'):
w3 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 64], stddev=0.1),
name='e_w3')
conv3 = tf.nn.conv2d(res22,
filter=w3,
strides=(1,2,2,1),
padding="VALID",
name='e_conv3')
b3 = tf.Variable(tf.truncated_normal(shape=tf.shape(conv3), stddev=0.1),
name='e_b3')
conv3_ = tf.nn.relu(conv3+b3, name='conv3_')
res31 = self.res_block(conv3_, 5, 64, 64, 3, 'e')
return res31
# End of encoder
def decoder(self, img, reuse_in):
img1 = self.res_block(img, 8, 64, 64, 3, 'd')
with tf.variable_scope('d_conv'):
w2 = tf.Variable(tf.truncated_normal(shape=[2, 2, 64, 64], stddev=0.1),
name='d_w2')
img2 = tf.nn.relu(tf.nn.conv2d_transpose(img1,
filter=w2,
output_shape=[self.batch_size, 15, 15, 64],
strides=(1,1,1,1),
padding="VALID",
name='d_deconv2'))
img31 = self.res_block(img2, 9, 64, 64, 3, 'd')
img32 = self.res_block(img31, 10, 64, 64, 3, 'd')
with tf.variable_scope('d_conv'):
w4 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 64], stddev=0.1),
name='d_w4')
img4 = tf.nn.relu(tf.nn.conv2d_transpose(img32,
filter=w4,
output_shape=[self.batch_size, 31, 31, 64],
strides=(1,2,2,1),
padding="VALID",
name='d_deconv4'))
img51 = self.res_block(img4, 11, 64, 64, 3, 'd')
img52 = self.res_block(img51, 12, 64, 64, 3, 'd')
with tf.variable_scope('d_conv'):
w6 = tf.Variable(tf.truncated_normal(shape=[3, 3, 64, 64], stddev=0.1),
name='d_w6')
img6 = tf.nn.relu(tf.nn.conv2d_transpose(img52,
filter=w6,
output_shape=[self.batch_size, 63, 63, 64],
strides=(1,2,2,1),
padding="VALID",
name='d_deconv6'))
with tf.variable_scope('d_conv'):
w7 = tf.Variable(tf.truncated_normal(shape=[3, 3, 3, 64], stddev=0.1),
name='d_w7')
img7 = tf.nn.conv2d_transpose(img6,
filter=w7,
output_shape=[self.batch_size, 128, 128, 3],
strides=(1,2,2,1),
padding="VALID",
name='d_deconv7')
return tf.sigmoid(img7)
这是培训详细信息。
self.fc1 = self.encoder(self.img_noisy, False)
self.img_dec = self.decoder(self.fc1, False)
self.img_loss = tf.reduce_mean(tf.losses.mean_squared_error(self.img_dec, self.img_front))
self.imge_graph = tf.summary.scalar("encoder-decoder MSE", self.img_loss)
self.loss = self.img_loss
self.optim = tf.train.AdamOptimizer(learning_rate=self.lr).minimize(self.loss)
我尝试使用tf.layers.conv2d_transpose()，并使用tf.image.image_gradients()包括梯度损失。
出于类似的目的，我在带有反卷积层的其他一些代码中也注意到了这种现象。预先感谢。