Edoardo Botta

The output images in your example are actually slightly different. Checking how many entries match between the two tensors: ``` import torchvision.transforms as transforms transform = transforms.Compose([ transforms.PILToTensor() ]) equals...

That is specified [here](https://github.com/EdoardoBotta/RQ-VAE-Recommender/blob/main/data/preprocessing.py#L42). I used sentence-transformers/sentence-t5-xl from Huggingface in this implementation.

The model testing is done during training in the train_decoder.py. This repo does not currently achieve the same results as the original paper.

Could you provide some examples?

Thanks for flagging this. I'll take a look at it.

Yes. I have seen codebook collapse a lot in my experiments. This is a very common issue with VQ-VAE. I would recommend keeping track of the entropy of the codebook...

Hi @Bazinga-ai , thank you for reaching out. This behavior is, to some degree, expected. There are two parts to the question you asked: * **Why does vl start low?**...

This was caused by some experiment I was performing with the CategoricalReconstructionLoss, trying to integrate the movie genres features. Ultimately, I decided not to use the genre features at all...

Hi @Levia-Mobius , thank you for pointing this our. I have pushed a fix to correctly propagate the commitment weight: https://github.com/EdoardoBotta/RQ-VAE-Recommender/commit/a9e1d97529e87a04d4f3d5b063657b0c4e53729a. Feel free to pull the latest commit.

Can you clarify what you mean by 10/128 have duplicate ids? How are these distributed? For example, is it 10 items with the same id or they are only pairwise...