search

xinntao / Real-ESRGAN

Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration.
BSD 3-Clause "New" or "Revised" License
28.41k stars 3.57k forks source link

Inference net4000_g_latest fine tuning, results are black images #574

Open mindFlayer123 opened 1 year ago

mindFlayer123 commented 1 year ago

Hi, I finetuned the model using [finetune_realesrgan_x4plus_pairdata.yml] for about 5000 iterations and some other settings like:

  batch_size_per_gpu: 1

setting learning rate from 1e-4 to 1e-3 and changing this:

 scheduler:
    milestones: [4000]

  total_iter: 4000

Rest as it is. Total epochs count was 3 only, as I was using colab I done these settings for faster training. But, after obtaining the new weights when I tested with inputs, the results I got were complete black images. What could be the possible issue ? Training eta : 22 mins. @xinntao

yulongblue commented 1 year ago

I am also experiencing the same problem. Have you solved it?