When train in lower bitrates ( < 0.14)?

[JXH-SHU]

Hi, thank you for your contributions. Following https://github.com/Justin-Tan/high-fidelity-generative-compression/issues/12, I train the model on my own datasets (very small, contain 10,000 images with size 256*256). I change the values in the target_rate_map dict in default_config.py to [0.05,0.07,0.1]. However, the model shows higher bpp than the pretrained low model (0.14). Specially, the model trained with target rate 0.05 shows poor performance. And the q_bpp of test is 0.325, while that of train is 0.078. Could you give me some suggestions.

[JXH-SHU]

Fortunately, I found in the next GAN fine-tuning training (low model) that the value of bpp drops and meets expectations.

[yifeipet]

What do you mean by "next GAN fine-tuning"? Have you changed the channel number of output of the encoder?

@JXH-SHU Hello Shu, What do you mean by "the next GAN fine-tuning"? Have you changed the channel number of output layer of the encoder? Thank you!

Sincerely, Yifei

[lzhzju]

Fortunately, I found in the next GAN fine-tuning training (low model) that the value of bpp drops and meets expectations.

May I ask how you manage to do that?