SimonKohl/probabilistic_unet

Model does not train at all

hajtos opened this issue · 0 comments

I have successfully ran the training on the Cityscapes dataset while only changing the paths in the config files and fixing issue #2 and executed the evaluation scripts.

However, the results were completely different than those presented in the readme. In fact, after investigating, it seems that the trained model only outputs a single number across all images - 3, which seems to be the id for the class 'wall'. This shows that the model failed to learn any task, even though all the learning parameters were preserved from your code.

Do you have any pointers you could give as to what might be causing such a failure and how could I go about making the model actually learn something?