tf.nn.conv2d_transpose 是否转置过滤器？

Question

I have a trained feedforward CNN. The shape of filter is [height, width, in_channels, out_channels]. And i want to use those filter to do deconv, we know that the deconv process needs the transpose of the filter.Do i need to transpose the filter mannuly, or the TF will do it inside the tf.nn.conv2d_transpose and all we need to do is pass the trained filter to tf.nn.conv2d_transpose ?

Answer 1

We needn't to transpose the filter manually.In general, wo organize our code in the following way.

stride = [1,1,1,1]
conv1W = tf.Variable(tf.random.normal[4,4,3,20])
conv1 = tf.nn.conv2d(input, conv1W, strides=stride, padding='SAME')
conv1 = tf.nn.relu(conv1)

Then, do the deconv process

deconv1 = tf.nn.conv2d_transpose(conv1, conv1W, output_shape=[batch_size,output_height, output_width, output_channels],strides=stride)
res = tf.nn.relu(deconv1)

The res is the result of deconv process.

In a word, the filter and stride using in deconv process is the same as the filter and stride using in conv process.

Answer 2

According to tensorflow documentation, you have to change the filter shape. Here I will put how this fact has described in latest TF documentation 1.9.

For the tf.nn.conv_2d the filter variable should be - A 4-D tensor of shape [filter_height, filter_width, in_channels, out_channels ]

For the tf.nn.conv_2d_transpose the filter variable should be - A 4-D tensor of shape [filter_height, filter_width, out_channels,in_channels ]

• Here clearly we can see we need to change the change the shape with respect to out_channels