search

swz30 / MPRNet

[CVPR 2021] Multi-Stage Progressive Image Restoration. SOTA results for Image deblurring, deraining, and denoising.
Other
1.16k stars 188 forks source link

When i train Rain_Heavy #66

Closed szpxmu closed 3 years ago

szpxmu commented 3 years ago

the dataset has1800 training data, 200 test data, Why the loss is so high, I try to change the learning rate to 1.5e-4. but it is no helpful

image

swz30 commented 3 years ago

Hi @hellogry

Try this: https://github.com/swz30/MPRNet/issues/51#issuecomment-843802843

szpxmu commented 3 years ago

I've used it and it doesn't work

swz30 commented 3 years ago

Can you share the training log for few epochs? It will be fine as long as the loss is gradually decreasing (and not exploding).

szpxmu commented 3 years ago

All losses remained almost at around 1442. The minimum is 1441. How do I display the training log.

adityac8 commented 3 years ago

Hi @hellogry

To save the logs, you can use python train.py | tee deraining_logs.txt

Thanks

FanChiMao commented 3 years ago

Isn't this normal? I'm not sure about that but I think that it's because the loss is epoch loss, the sum of every iteration loss. Divide by the training images, the iteration loss is about 0.8. It's acceptable and correct I think.