I have gone through the baseline that is in Tensorflow and converted it to Fast.ai/ PyTorch. I believe I have almost everything the same except I am doing dynamic augmentations. I am only getting 1/2 the result that is output by the baseline.
I was wondering if anyone could take a look at this notebook and let me know where it might be going wrong. https://www.kaggle.com/code/ltspacemonkey/2021-nasa-harvest-rwanda-field-detection-baseline (No guarantees it will run in Kaggle though.)
I got similar results with BCELoss (didn't work even I tried different learning rates). After that I tried DiceLoss, and my results improved.
Hello, Do you use baseline code from this link: radiantearth/Nasa_harvest_field_boundary_competition (github.com)
I used that code as the basis for my notebook in PyTorch, yes. It is done in TensorFlow, which I don't know.
When I use DiceLoss() my validation loss doesn't change at all and training loss bounces around the same number.
I trained for 200 epochs and the numbers didn't change.