search

junyanz / pytorch-CycleGAN-and-pix2pix

Image-to-Image Translation in PyTorch
Other
23.09k stars 6.32k forks source link

ValueError:Unknown resampling filter #1489

Open yourbikun opened 2 years ago

yourbikun commented 2 years ago

I need your help! I use training argument "python train.py --dataroot /gemini/data-1/render_data1 --name r_cyclegan --model cycle_gan --preprocess scale_width_and_crop --load_size 800 --crop_size 400", and then get this error.ValueError: Unknown resampling filter (InterpolationMode.BICUBIC). Use Image.NEAREST (0), Image.LANCZOS (1), Image.BILINEAR (2), Image.BICUBIC (3), Image.BOX (4) or Image.HAMMING (5) image

yourbikun commented 2 years ago

and I find if I use the default preprocess, and set --load_size 800 --crop_size 400,this error don't happend.

taesungp commented 2 years ago

Sorry about that. There was some mismatch in typing interpolation methods between torchvision.transforms.InterpolationMode and PIL.Image. It's a new bug introduced as we tried to suppress warnings (PR #1414). I pushed the fix to the master branch (d5e62dd021d51b).

yourbikun commented 2 years ago

thanks

yourbikun commented 2 years ago

I have another question. I used your code in RTX3090 and set the batchsize to 4. In the past, I should increase the learning rate at the same time, but I did not see the tip document that doubled the learning rate. So, do I need to double the learning rate? If so, please ask me how to set it

taesungp commented 2 years ago

You are right to point out the general rule of thumb about determining the learning rate. However, it's still empirical particularly for the GAN training. I'd just try it with the same learning rate, and also with a larger learning rate.

We actually recommend sticking to the batch size of 1, because overfitting is a big issue on Pix2Pix and CycleGAN. Note that the datasets we used in this work are quite small. To prevent overfitting, having a small batch size helps.

yourbikun commented 2 years ago

You are right. After I use a higher learning rate, the test effect is very poor when the loss has reached an ideal level. There are many noises in the pictures, and I perform well when I keep the learning rate unchanged in large batches. Of course, because the "scale_width" option cannot be used normally before, I use the default preprocess. I don't know whether it is related to it.