Closed leon-w closed 1 year ago
I figured it out: Instead of using the config that get_config()
in config_diffusion_cifar10.py
returns, one should use the hyperparameters that are contained in get_sweep()
in the same file. With that i managed to get very convincing results, here are some samples for the first 67k steps:
If anyone is looking for ready-to-use code, feel free to use this: https://github.com/leon-w/pix2seq/tree/cifar10
Hi,
I'm currently trying to train RIN on CIFAR-10 using the code and config that was provided in the repository.
I did some minor changes to the code to get it to work:
cifar10_stats_real.npy
file (Diff)image_diffusion_model.py
where theModel.sample
calls a method that is only defined inModelT
(Diff)The entire diff can be seen here: https://github.com/google-research/pix2seq/compare/main..leon-w:b6609 This was the exact code I used for training and evaluation.
Training Setup
I trained the model using this command:
train_log.txt
The training finishes after around 2 hours. These are the training curves logged to Tensorboard:![image](https://github.com/google-research/pix2seq/assets/29836160/229a2336-daff-4434-81eb-742a34070f1f)
Eval Setup
I then run the trained model in evaluation mode to create a few samples using:
eval_log.txt
Unfortunately, the generated samples don't seem to contain anything meaningful and only look like pure noise:
I also tried training the model 10x longer but still got only noise.
Has anyone successfully trained a RIN model using this code base before and has any idea how I can get this to work? Any help would be highly appreciated!