如何在张量流中制作反卷积层？

Question

我写了一个反卷积层的代码，

def deconv2d(x, W,stride):
  x_shape = tf.shape(x)
  output_shape = tf.stack([x_shape[0], x_shape[1]*2, x_shape[2]*2, x_shape[3]//2])
  decon = tf.nn.conv2d_transpose(x, W, output_shape, strides=[1, stride, stride, 1], padding='SAME')
  layer_shape = get_layer_shape(decon)
  print('DECONV Shape : ', layer_shape)
  return  decon

上面的函数我是这样调用的，

deconvolution1 = deconv2d(x=cnn_layer10, W=[2,2,512,1024], stride=2)

收到此错误，

File "u-net.py", line 84, in obj.computation() File "u-net.py", line 41, in computation deconvolution1 = deconv2d(x=cnn_layer10, W=[2,2,512,1024], stride=2) File "/home/shuvo/u-net/architecture.py", line 35, in deconv2d decon = tf.nn.conv2d_transpose(x, W, output_shape, strides=[1, stride, stride, 1], padding='SAME') File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/ops/nn_ops.py", line 1019, in conv2d_transpose if not value.get_shape()[axis].is_compatible_with(filter.get_shape()[3]):
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/tensor_shape.py", line 500, in getitem return self._dims[key] IndexError: list index out of range

我的目的是在形状应该是的地方做一个反卷积层， [batch_size, 36,36,1024]=>[batch_size,72,72,512]。请帮我解决这个错误，

Answer 1

输入参数 filter 到 tf.nn.conv2d_transpose 是权重矩阵本身，而不仅仅是过滤器的大小。

修复上述问题的修改代码如下所示：

cnn_layer10 = tf.placeholder(tf.float32, (10, 36, 36, 1024))

def deconv2d(x, W,stride):
   x_shape = tf.shape(x)    
   weights = tf.Variable(tf.random_normal(W))
   output_shape = tf.stack([x_shape[0], x_shape[1]*2, x_shape[2]*2, x_shape[3]//2])
   decon = tf.nn.conv2d_transpose(x, weights, output_shape, strides=[1, stride, stride, 1], padding='SAME')

   return  decon

deconvolution1 = deconv2d(x=cnn_layer10, W=[2,2,512,1024], stride=2)

with tf.Session() as sess:
   sess.run(tf.global_variables_initializer())
   print('Decon Shape:',sess.run(deconvolution1, {cnn_layer10: np.random.random((10, 36,36,1024))}).shape)
   #Output
   #Decon Shape: (10, 72, 72, 512)

注意：最好使用 tf.layers.conv2d_transpose API，其中 filters arg 是过滤器大小，权重初始化发生在其中。

如何在张量流中制作反卷积层？

How to make a de-convolution layer in tensorflow?

python

convolution

deep-learning

tensorflow

deconvolution