Dynamic image size [batch, None, None, 3] error.

[zsdonghao]

Hi thank you for releasing this implementation.

I try to evaluate the model on images with different size, so I would like to define a placeholder as:

t_image = tf.placeholder('float32', [1, None, None, 3], name='input_image')

then when I use the subpixel layer, I will get:

    X = tf.reshape(I, (bsize, a, b, r, r))
  File "/ssd2/Workspace/env3/lib/python3.5/site-packages/tensorflow/python/ops/gen_array_ops.py", line 2451, in reshape
    name=name)
  File "/ssd2/Workspace/env3/lib/python3.5/site-packages/tensorflow/python/framework/op_def_library.py", line 508, in apply_op
    (input_name, err))
ValueError: Tried to convert 'shape' to a tensor and failed. Error: None values not supported.

Do you have any idea to solve this problem? if we cannot use dynamic size, we need to define different inference for different image, which is resource and time consuming.

Many thanks in advance.

[riokt]

I got same problem.

[zsdonghao]

@riokt I solved this problem, please check the SRGAN implementation https://github.com/zsdonghao/SRGAN

[riokt]

@zsdonghao Nice. Thank you

[SystemErrorWang]

@zsdonghao Hello, I tried to contact you through tencent QQ, but mention again here if you didn't see it. Hope you can give me some help, thanks.
I tried to check your tensorlayer code but failed to figure out how you solve this problem. The subpixel conv2d is not capable of dealing with placeholder with shape [batch_size, None, None, 3] as tf.reshape is used, and I also tried to save and restore network parameters in npz file format, but tensorflow still give me this: "You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [?,48,48,3]". However, the tensorlayer based SRGAN code can test on a whole DIV2K image.
So I would like to know what method you used in tensorlayer that solved this problem.

[zsdonghao]

@SystemErrorWang please update your TensorLayer version, it will support dynamic image size.

[SystemErrorWang]

@zsdonghao I tried the tensorlayer script and it succeeded, but still I would like to know if there is any method to deal with it in native tensorflow, thanks.

[zsdonghao]

@SystemErrorWang hi, you can just simply copy the core codes from TensorLayer's layer.

[SystemErrorWang]

@zsdonghao Thank, I solved this problem
@riokt You can simply use tf.shape(X) to substitute X.get_shape() in PS() function, and set input placeholder to be shaped [None, None, None, 3]

[lhanappa]

@zsdonghao @SystemErrorWang
Hi, good day. As the implementation of Tensorlayer, if I understand correctly, the function of_phase_shift() is equal to "depth_to_space"?
https://www.tensorflow.org/api_docs/python/tf/nn/depth_to_space

[SystemErrorWang]

@lhanappa I didn't test this function myself, but I think they are theorically the same. Guess you can do some experiments.

[lhanappa]

@SystemErrorWang Thanks, yes, I already did.
"depth_to_space" and "space_to_depth" are two reciprocal functions for expanding/compressing channels in native tensorflow.