Closed mobassir94 closed 4 years ago
@mobassir94 It depends on the dataset type, model size, etc.
In my experience, usually alpha=1
works well for many datasets.
But when we use small datasets or small models, a small alpha (e.g., ~0.2) works better than a larger one.
So I would like to recommend that if you have computational resources to try the same experiments several times, try two or three alphas (e.g., 1.0, 0.5, 0.2) and find the best one.
for my dataset how do i choose value of alpha for cutmix to get best performance?