dsgoficial / pytorch_segmentation_models_trainer Goto Github PK

There is bug in class_freq computation, more specifically, reading value of class_freq from generated csv dataset.
In here, the value of class freq for segmentation mask is read from df.columns, but this value in csv is somehow computed from image array as follows. This lead to NaN or negative loss during training, because here sum of the class_freq was larger than 1.
My temporary solution was to ignore class_freq column from csv file here

Hi,
thanks for great framework! However there is huge fundamental bug in FrameField approach.
First here
crossfield is treated as mask during loading, thus, in here all anglefield are vanished and set to value one.
Moreover, any affine augmentation from albumentation is not usable, because anglefield has to be treated in different way for affine transformations, as in the original FrameField repo f.e. rotate_anglefield

Am I right?