Inconsistent results of each training under the same parameter configuration even when seed is fixed.

EricGuo5513 / momask-codes

Official implementation of "MoMask: Generative Masked Modeling of 3D Human Motions (CVPR2024)"

https://ericguo5513.github.io/momask/

MIT License

856 stars 73 forks source link

Inconsistent results of each training under the same parameter configuration even when seed is fixed. #44

Closed fanqinshan0114 closed 3 weeks ago

fanqinshan0114 commented 6 months ago

Hello! I want to ask why the model evaluation results of each training under the same parameter configuration are inconsistent. My environment is exactly the same as the environment.yaml file provided, and I have not modified any code. I want to know if this is normal when the random seed is fixed.

Murrol commented 6 months ago

Hi, thanks for using our code.

We found that the random seed for train_vq is not set and just fixed it. Please pull the latest code to use it.

Feel free to create a pull request to contribute to MoMask.

fanqinshan0114 commented 6 months ago

Hi, thanks for using our code.

We found that the random seed for train_vq is not set and just fixed it. Please pull the latest code to use it.

Feel free to create a pull request to contribute to MoMask.

It seems that the eval_val_loader in train_vq.py should use the validation set instead of the test set?