open-genie
open-genie copied to clipboard
myscience
Pytorch implementation of "Genie: Generative Interactive Environments", Bruce et al. (2024).
'''usage: tokenizer.py [-h] [-c CONFIG] [--print_config[=flags]] {fit,validate,test,predict} ... error: argument subcommand: invalid choice: 'train' (choose from 'fit', 'validate', 'test', 'predict')''' there's bug with the script
Whenever I try to use the latent action model using the default descriptions given, I get an error saying there's no 'n_embed' argument. I tried using specifically the script test_action...
Thanks for the codebase! Do you plan to share the training results of your implementation?
There is no config file for training the latent action model. Also, in the example code of training the latent action model, there is no "n_embed" for space-time_attn: `('space-time_attn', {...
https://github.com/myscience/open-genie/blob/732b9f9b746f18fff1a0fb22f83638224f2f7cc6/genie/module/quantization.py#L125 ```python entropy_loss = inp_ent + self.diversity_weight * avg_ent ``` This line should be ```python entropy_loss = inp_ent - self.diversity_weight * avg_ent ```
https://github.com/myscience/open-genie/blob/732b9f9b746f18fff1a0fb22f83638224f2f7cc6/genie/module/attention.py#L290C26-L290C37 ```python b, *t, h, w, c = video.shape ``` should be: ```python b, *t, h, w, c = inp.shape ```