Class_freq computation about pytorch_segmentation_models_trainer HOT 2 OPEN

How did you build your csv? If you use the build dataset option it will compute properly (I hope so =] ) the class frequency for each class. I've only tested with binary classification and maybe there can be bugs if you are trying with sereval classes.

There is no problem ignoring the csv entry, this is done to save computations on execution time, but if you do not enter this data in the csv, the framework will compute it each time.

from pytorch_segmentation_models_trainer.

I have used this config to generate dataset.
Which also have generated expected output from there. But if you watch closely, class_freq (which is used for mask class_freq) has very odd values, f.e. [0.25065063 0.29929634 0.3453119 ] (last value is vertex frequency, which should be very small number, because mask for verticies is sparse). If you take values from bands_means [63.915911111111114, 76.32056666666666, 88.05453333333334], you can see that [63.915911111111114, 76.32056666666666, 88.05453333333334] / 255 = [0.25065063 0.29929634 0.3453119 ]. Therefore, class_freq is computed from image's band mean.

from pytorch_segmentation_models_trainer.