Bad accuracy at black/white images

[TheCrazyT]

I just noticed that on black/white images the process fails. Is there a reason why the model was trained on colored images? Reason I'm asking is because you would use much less memory(and maybe computation time) on black/white images.

[TheRustlessSummer]

Hello, I'm sorry to bother you, but I have some questions when I run make . Can you tell me the version of tensorflow and cmake and cuda you use?

[TheCrazyT]

Here are the versions I'm using: tensorflow-gpu: 2.3.0 cmake: 3.12.0 CUDA: 10.1

Modified the code slightly with the help of "tf_upgrade_v2" and some other modifications(with "sed") to be able to use compat v1 mode of tensorflow 2.3.0. Also changed some code of "lmbspecialops" (else i wasn't able to compile it ...)

The whole code can be found here: https://colab.research.google.com/drive/14DzqseiuQ0ohh8VnTXaz-mE4kqwGzbWI?usp=sharing