CODEWITHSUNDEEP
pythontensorflowconv-neural-network
Jon Rein
Jon Rein

Reputation: 938

Arbitrary/unknown output shape of Tensorflow conv2d_transpose

Suppose I have a tensor of varying height, i.e. shape [batch_size=32, height=None, width=25, n_channels=128]. I'd like to upsample this tensor with the conv2d_transpose op, but I'm not sure how to generate the required output_shape argument. With a known height, I'd do something like

def get_conv_transpose_shape(input, out_channels):
    out_shape = input.get_shape().as_list()
    out_shape[1] *= 2
    out_shape[2] *= 2
    out_shape[3] = out_channels
    return out_shape

But when height=None, this produces the following error

TypeError: unsupported operand type(s) for *= 'NoneType' and 'int'

Is there a solution to this other than zero-padding all of my inputs to a standard size? That's a computation cost I'd like to avoid.

Upvotes: 0

Views: 288

Answers (1)

rafaelvalle
rafaelvalle

Reputation: 7063

When you call .get_shape().as_list(), you end up in "static python land" trying to multiply a None with an int.

The operation should be carried in the symbolic domain, that is multiplying tensorflow.get_shape(input) with another symbolic variable of type int.

Upvotes: 1

Related Questions