Gaurav Parmar
Gaurav Parmar
Yeah batchsize of 8 is likely too high
Could you share the input and output images? -Gaurav
Yeah these images look very out of distribution for my pre-trained night-to-day model. The pretrained model is trained on BDD images. I suspect you will achieve better results if you...
Can you share you accelerate config file you are using for multi-gpu training?
Looking at your results I can spot a couple issues with your setting: 1. Since you are doing medical image segmentation, you can probably remove the CLIP_SIM loss. The text-image...
From this error, it looks like you need to cast the input also to fp16.
Hi, The package does not directly support this. However, you can wrap your generator function to resize the input noise based on the requirements of the model. -Gaurav
I think typically having a large enough "real" and "fake" folders should be sufficient.
Can you print out the magnitude of difference between the two outputs?
Hi, Could you expand a bit more on the issue you are facing? -Gaurav