Lesson 14. Segmentation

Part 2 & Alumni (2018)
21

I have tried TTA for a segmentation problem,so current TTA implementation is not useful?
Is there any alternative way in fastai

22

I used a hacky version. I changed the dataset which contains flipped dataset. We would not be able to use TTA for this.
md.test_dl.dataset = TestFilesDataset(tst_x,tst_x,tfms[1],PATH)

class TestFilesDataset(FilesDataset):
    def __init__(self, fnames, y, transform, path):
        self.y=y
        self.th = 1
        super().__init__(fnames, transform, path)
        
    def get_x(self, i): 
            return np.fliplr(open_image(os.path.join(self.path, self.fnames[i])))
        
    def get_y(self, i):
        
        return np.fliplr(open_image(os.path.join(str(self.path), str(self.y[i]))))
    def get_c(self): return 0
23

Hello,

Has anyone have trouble when loading the model and predicting after closing jupyter notebook (or resetting the kernel). When I try to just predict on my test or validation dataset the predictions are different than when I train + predict them.

Anyone knows?

Thanks.