training bpp_y

[wulayu]

Dear author,

I am a student researching video coding, and I must say your work has been truly outstanding and inspiring to me.

I have previously reproduced and fine-tuned the DCVC-HEM model, which has granted me a deeper understanding of the video coding field. However, recently, I encountered difficulties while trying to reproduce the DCVC-FM code. I consistently faced a situation where the bpp training easily becomes divergent. I have scrutinized the training process and found that it is particularly prone to training above 4.0, whether in training bpp_y or bpp_mv_y.

I attempted to import pre-trained parameters and train the model using only bpp as the loss, but unfortunately, I still could not achieve training below 0.6. The difference between the results and the pre-trained model, which achieved 0.01, is quite significant. Could you kindly provide me with some valuable advice regarding training the model? I am eagerly looking forward to your response.