qingshi9974 / ICLR2024-FTIC

[ICLR2024] FTIC: Frequency-aware Transformer for Learned Image Compression
26 stars 1 forks source link

FineTuning #1

Closed wonlee2019 closed 1 month ago

wonlee2019 commented 3 months ago

Thanks for your contributions!

I have one question:

How can I train the low bitrate model by using well-trained high bitrate model?

How can I train the model optimized with MS-SSIM from the well-trained model optimized with MSE?

Thanks again!

qingshi9974 commented 3 months ago

Thanks for your contributions!

I have one question:

How can I train the low bitrate model by using well-trained high bitrate model?

How can I train the model optimized with MS-SSIM from the well-trained model optimized with MSE?

Thanks again!

You can load the checkpoint of the pre-trained high bitrate/MSE-optimized model and use a lower learning rate (i.e., 1e-5) for fine-tuning to obtain a low bitrate/MS-SSIM-optimized model.

wonlee2019 commented 3 months ago

Thanks for your response. I changed the learning rate for training, but it doesn't work.

As same as before, running the code followed by the increase of the GPU-MEM, and it is terminated at 2k MiB (lower than the target 8k MiB).

It's weird. Whatever, thanks again.