Request for DCVC-RT Training Code for Pruning and Quantization Research

[vicse7en]

Dear DCVC-RT Authors,
Thank you for releasing DCVC-RT and making the inference codebase publicly available. The work is impressive, and the architectural innovations — particularly the implicit temporal modeling and single-scale latent representation — make it a compelling target for compression and deployment research.

I have successfully set up the inference pipeline using the provided pretrained checkpoints (cvpr2025_image.pth.tar and cvpr2025_video.pth.tar) and am now looking to implement the training loop from scratch based on the inference code and the paper description. While this is feasible in principle, several implementation details are difficult to recover from the paper alone, particularly:

The multi-stage training schedule and how the Module Bank (rate-control vector banks q_e, q_d, q_f, q_r) is handled during training across different QP values
Whether model integerization is applied or disabled during training
Any warm-up or stabilization strategies used for the rate-distortion loss in the early training stages

I fully understand if the training code is not in a state suitable for public release or official maintenance. Even a rough internal version, a configuration file, or brief guidance on the above points would be enormously helpful and would allow me to avoid building on incorrect assumptions.
If you are able to share any materials, please feel free to reach out at: 2595490024@qq.com
Thank you again for your outstanding contribution. I look forward to hearing from you.