Some doubts about quantization loss

[xiaoShen110141]

Thanks for your great work
I recently found that when trying to replicate VQVAE (which I think is RQVAE), due to the convolutional layers (\Phi) added before loss calculation, the quantization loss showed a continuous upward trend during my training process. I tested it on both 128 * 128 FFHQ and 256 * 256 FFHQ datasets, and the quantization loss of 256 FFHQ reached around 1.03 after 120 rounds of training. I want to know if this is a normal phenomenon in VAR?
I will share the training logs of FFHQ 256x, please open it using TensorBoard.https://drive.google.com/file/d/1euW5LI6KYBO3UZyYWBsSN_BFV5lG1ajy/view?usp=sharing