lemon-cmd / energy-transformer-torch Goto Github PK

Thank you for providing the PyTorch version of the Energy Transformer originally implemented in JAX. This is really helpful for me, but I have not yet been able to reproduce the results like example1.png or images in visuals directory (gif files).

Let me confirm that the example model model.pth is trained with the config file model_config.txt in the same directory, not with the config file of the same name in the parent directory energy-transformer-torch/model_config.txt? If this is correct, some parameters ('alpha': 5.0, 'attn_beta': 16.0, 'epochs': 10000) are much bigger than the hyperparameter settings reported in the paper p.16 of 2302.07253, and some ('b1': 0.99, 'b2': 0.999, 'batch_size': 512, 'learning_rate': 5e-05, 'weight_decay': 0.001) are still somewhat different.

I know that the best hyperparameters would be different since the paper uses ImageNet-1k to train the model and here uses CIFAR-10/100. I just would like to know in what configuration the example model is trained and how I can reproduce the results of example1.png or images in visuals directory.

Thanks for any additional details you can provide, I will be trying to reproduce the results in this repo.