Is upscale_two possible?

[whyboris]

Could you comment whether it's possible to modify only a bit of the code and continue using your pre-trained model to "upscale_two"?

Thank you so much for TecoGAN 🙇
I've successfully upscaled sequences of images with your pre-trained model 🚀

The current code hard-codes upscale_four - quadrupling the width and quadrupling the height.

I have attempted several code changes, starting with changing output_shape to be *2 (not *4) of the input shape.

Needless to say, minor tinkering keeps resulting in incompatible shapes, e.g:

Dimension 1 in both shapes must be equal, but are 288 and 576. Shapes are [1,288,360,3] and [1,576,720,3]. for 'generator/Assign_1' (op: 'Assign') with input shapes: [1,288,360,3], [1,576,720,3].

I suspect I might be trying a fool's errand (since the model was trained to upsample x4 rather than x2). Please let me know 🙇

ps - In your paper you write:

In our case, the outputs have four times the resolution of the inputs.

But based on the code, it seems that you may have increased the resolution 16 times?

pps - minor grammatical error in the paper:

In order to avoid the this undesirable case

[RachelCmy]

Hi,

right, although the algorithm is not for a fixed resolution, the implementation here is now fixed to a factor of 4 in each dimension and increases the resolution for 16 times.

The code has to be changed consistently for many places for a factor of 2 and the published model cannot be used. An easier solution is to get the 4x result first, and down-scale it to 2x resolution.

Thank you for the ps and pps! We will update the paper in a future version.