CODEWITHSUNDEEP
pythonmachine-learningtensorflowconv-neural-networkdeconvolution
Devin Haslam
Devin Haslam

Reputation: 759

Transpose convolution (deconvolution) arithmetic

I am using tensorflow to construct a convolution neural network. Given a tensor of the shape (none, 16, 16, 4, 192) I want to perform a transpose convolution that results in the shape (none, 32, 32, 7, 192).

Would a filter size of [2,2,4,192,192] and stride of [2,2,1,1,1] produce the output shape that I want?

Upvotes: 1

Views: 617

Answers (1)

Maxim
Maxim

Reputation: 53758

Yes, you are almost right.

One minor correction is that tf.nn.conv3d_transpose expects NCDHW or NDHWC input format (yours appears to be NHWDC) and the filter shape is expected to be [depth, height, width, output_channels, in_channels]. This affects the order of dimensions in the filter and stride:

# Original format: NHWDC.
original = tf.placeholder(dtype=tf.float32, shape=[None, 16, 16, 4, 192])
print original.shape

# Convert to NDHWC format.
input = tf.reshape(original, shape=[-1, 4, 16, 16, 192])
print input.shape

# input shape:  [batch, depth, height, width, in_channels].
# filter shape: [depth, height, width, output_channels, in_channels].
# output shape: [batch, depth, height, width, output_channels].
filter = tf.get_variable('filter', shape=[4, 2, 2, 192, 192], dtype=tf.float32)
conv = tf.nn.conv3d_transpose(input,
                              filter=filter,
                              output_shape=[-1, 7, 32, 32, 192],
                              strides=[1, 1, 2, 2, 1],
                              padding='SAME')
print conv.shape

final = tf.reshape(conv, shape=[-1, 32, 32, 7, 192])
print final.shape

Which outputs:

(?, 16, 16, 4, 192)
(?, 4, 16, 16, 192)
(?, 7, 32, 32, 192)
(?, 32, 32, 7, 192)

Upvotes: 1

Related Questions