Harry Fan

Results 4 comments of Harry Fan

Training SegFormer model not working (goes through notebook, but model loss becomes nan) on dataset I created (stuck for a week or so)

@sayakpaul

Training SegFormer model not working (goes through notebook, but model loss becomes nan) on dataset I created (stuck for a week or so)

The model's loss still becomes nan even with 10x (1000x?) lower learning rate than what was originally in the notebook. A few errors also occur; a screenshot and some text...

Training SegFormer model not working (goes through notebook, but model loss becomes nan) on dataset I created (stuck for a week or so)

I think I figured it out; the labels file I used had pixel value 255 as contrails, pixel value 1 as another ("filler") class, and pixel value 0 as unlabeled....

Training SegFormer model not working (goes through notebook, but model loss becomes nan) on dataset I created (stuck for a week or so)

@sayakpaul