Performance of num_head=1?

[FacePoluke]

Thank you for your excellent work. I noticed that your codebook head number is 4, which means downstream generation tasks need to output 4 tokens at once, potentially making training more challenging. I would like to know how much the performance would differ if the head number is 1.
Looking forward to your response.

[zbr17]

Hello!

Thank you for your interest in our OptVQ project. In the early stages of the project, I experimented with the setting num-head=1, and generally, we found that the performance was inferior compared to when num-head=4. However, due to limited laboratory resources, I am currently using an 8-GPU 4090 server, and in the latest version of the training code, I only have results for num-head=4 and num-head=8. Please bear with me for a moment, as I plan to include checkpoints and evaluation results for num-head=1 in the project in the future.

Best regards,
Borui Zhang

[FacePoluke]

Hello!

Thank you for your interest in our OptVQ project. In the early stages of the project, I experimented with the setting num-head=1, and generally, we found that the performance was inferior compared to when num-head=4. However, due to limited laboratory resources, I am currently using an 8-GPU 4090 server, and in the latest version of the training code, I only have results for num-head=4 and num-head=8. Please bear with me for a moment, as I plan to include checkpoints and evaluation results for num-head=1 in the project in the future.

Best regards, Borui Zhang

Thanks for your reply. Looking forward to it.