andylolu2 / jax-vqvae-gpt Goto Github PK

Hey, nice and clean implementation!

I think I might have spotted a bug in the gradient of the commitment loss. It's supposed to pull the embeddings to the inputs that mapped to them. But you're computing it after patching the straight-through gradient onto the embeddings, so they actually now have the gradient of the inputs.

In other words, the current code:

quantize = self.codebook[encoding_indices]
quantize = inputs + jax.lax.stop_gradient(quantize - inputs)
encoding_loss = jnp.mean((jax.lax.stop_gradient(quantize) - inputs)**2)
commit_loss = jnp.mean((quantize - jax.lax.stop_gradient(inputs))**2)

Should have the commitment loss computed before applying the straight-through gradient:

quantize = self.codebook[encoding_indices]
encoding_loss = jnp.mean((jax.lax.stop_gradient(quantize) - inputs)**2)
commit_loss = jnp.mean((quantize - jax.lax.stop_gradient(inputs))**2)
quantize = inputs + jax.lax.stop_gradient(quantize - inputs)

There seems to be a discrepancy in the implementation of the codebook loss function for the VQ-VAE model, which does not align with the one described in the original paper "Neural Discrete Representation Learning" by van den Oord et al. (2017).

The current implementation of the loss function is as follows:

# Current implementation
# loss = ||sg[z_e(x)] - e|| + beta||z_e(x) - sg[e]||
encoding_loss = jnp.mean((jax.lax.stop_gradient(quantize) - inputs) ** 2)
commit_loss = jnp.mean((quantize - jax.lax.stop_gradient(inputs)) ** 2)
loss = encoding_loss + self.beta * commit_loss

However, I think you have flipped the variables inputs and quantize. I believe that $z_e(x)$ correponds to input, since it is the output of the encoder. Then $e$ would refer to the output of the codebook: quantize.

Adjusting the loss function as suggested should align the implementation with the theoretical model presented in the original paper, potentially improving the model's performance and stability.

The (simple) fix would be the following:

# Current implementation
# loss = ||sg[z_e(x)] - e|| + beta||z_e(x) - sg[e]||
encoding_loss = jnp.mean((jax.lax.stop_gradient(inputs) - quantize) ** 2)
commit_loss = jnp.mean((inputs - jax.lax.stop_gradient(quantize)) ** 2)
loss = encoding_loss + self.beta * commit_loss

Thank you for taking the time to read this issue!