Wrong results after Tensorflow 2x model conversion #70

[sergiocasaspastor]

Hi,

I hope you still give support for this repository. I tried converting this repository model to Tensorflow 2x step by step (also following the original repository by shepnerd). I think everything was fine, but when training and testing the model the results were not very satisfactory. So, maybe there is a problem with my conversion that I can't find it. I tried to train the model on a small subset of OpenImages V6. The principal parameters that I used (sorry, because names can be slightly different than yours):

--img_size 256x256x3 --batch_size 4 --learning_rate 1e-4 --gaussian_steps 7 --gaussian_kernel_size 32 --gaussian_kernel_std 20.0 --reconstruction_loss_weight 1.2 --adversarial_loss_weight 0.001 --gradient_penalty_loss_weight 10 --id_mrf_loss_weight' 0.03 --nn_stretch_sigma 0.5 --id_mrf_style_weight 1.0 --id_mrf_content_weight 1.0

I pretrained the model with only confidence reconstruction loss for the recommended steps, and results for this phase seem fine.


However, in the training phase, after some steps, the results do not seem to converge and eliminate the mask.

I would like to know if you have any idea of what could be happening, or you have experienced any similar issue. I reviewed the code, networks, and losses many times, but could not find any solution.

Thank you very much.

[github-actions[bot]]

Message that will be displayed on users' first issue