MarcoForte
commented
4 years ago Hi, I'll release the proper code for this soon. But here is the list of augmentations(and their inverses) for now. I loop through each scale then perform each augmentation in that list and average the results. Then I perform a weighted average of the results at the different scales.
augmentations = [(lambda x: x, lambda x: x), (lambda x: scipy.ndimage.rotate(x,90),lambda x:scipy.ndimage.rotate(x, 270)), (lambda x: scipy.ndimage.rotate(x,180),lambda x:scipy.ndimage.rotate(x, 180)), (lambda x: scipy.ndimage.rotate(x,270),lambda x:scipy.ndimage.rotate(x, 90)),
(lambda x: x[:,::-1], lambda x: x[:,::-1]), (lambda x: scipy.ndimage.rotate(x[:,::-1],90),lambda x:scipy.ndimage.rotate(x, 270)[:,::-1]), (lambda x: scipy.ndimage.rotate(x[:,::-1],180),lambda x:scipy.ndimage.rotate(x, 180)[:,::-1]), (lambda x: scipy.ndimage.rotate(x[:,::-1],270),lambda x:scipy.ndimage.rotate(x, 90)[:,::-1]),
(lambda x: x[::-1], lambda x: x[::-1]), (lambda x: scipy.ndimage.rotate(x[::-1],90),lambda x:scipy.ndimage.rotate(x, 270)[::-1]), (lambda x: scipy.ndimage.rotate(x[::-1],180),lambda x:scipy.ndimage.rotate(x, 180)[::-1]), (lambda x: scipy.ndimage.rotate(x[::-1],270),lambda x:scipy.ndimage.rotate(x, 90)[::-1]),
(lambda x: x[::-1,::-1], lambda x: x[::-1,::-1]), (lambda x: scipy.ndimage.rotate(x[::-1,::-1],90),lambda x:scipy.ndimage.rotate(x, 270)[::-1,::-1]), (lambda x: scipy.ndimage.rotate(x[::-1,::-1],180),lambda x:scipy.ndimage.rotate(x, 180)[::-1,::-1]), (lambda x: scipy.ndimage.rotate(x[::-1,::-1],270),lambda x:scipy.ndimage.rotate(x, 90)[::-1,::-1])]
scales = [0.9, 1, 1.1]
scale_weights = np.array([0.15, 0.75, 0.15])
uptodiff
The effect of your proposed model and training method is really amazing. “This, combined with longer training at 45 epochs and TTA, has a bigger impact than the choices of loss functions”. It seems TTA also plays an important role. So I want to konw the detail of TTA, such as the example code or related info. Looking forward to your reply!