Clarifications on Decoder and Embeddings about 3d-vq-vae-2 HOT 4 OPEN

Hi! Answers to your questions:

1. You cannot sample the decoder directly, you need to train an autoregressive prior (i.e. pixelcnn, pixelsnail, ViT, ..., maybe using a discrete denoising model would be cool...) on the embeddings obtained by putting your dataset through the encoder.
   You then sample your autoregressive model for embeddings, and put those embeddings through the decoder. See the original VQ-VAE paper: https://arxiv.org/pdf/1711.00937.pdf
2. • calc_ssim_from_checkpoint -> I simply had not added SSIM as a metric to tensorboard yet when I wrote this script, so this file can be ignored (or removed) now.
   • decode_embeddings.py -> the db_path are the generated embeddings by your autoregressive model, so you don't have them right now.
   • extract_embeddings.py -> yes, this file in principle takes your model + dataset and created the embeddings which should be used as training input for your autoregressive model.
   • As a general note, these three files are scripts and not intended as library files, and thus should be treated as such (i.e. low quality control, hardcoding a lot of stuff).

Nice to see that you're progressing :)

from 3d-vq-vae-2.

@robogast - Appreciate all of the information! Need to review the paper again :)

I look forward to trying the other scripts and posting how things go!

from 3d-vq-vae-2.

Hi @robogast

Your comments make much more sense now after reviewing the literature further :)

This is a nice overview from AI Epiphany!

https://www.youtube.com/watch?v=VZFVUrYcig0&t=1736s

from 3d-vq-vae-2.

Hi @robogast

I was trying to better understand encoding_idx. My understanding is that this is the last item in each of the 3 bottle neck layers? Curious why we throw the rest of the information away?

Thanks in advance!
-Akshay

def extract_samples(model, dataloader):
    model.eval()
    model.to(GPU)

    with torch.no_grad():
        for sample, _ in dataloader:
            sample = sample.to(GPU)
            *_, encoding_idx = zip(*model.encode(sample))
            yield encoding_idx

from 3d-vq-vae-2.