|
Algorithm 1: Teacher-student training overview |
-
1:
Train model on supervised set in burn-in step for 10,000 iterations
-
2:
After burn-in duplicate model into teacher and student
-
3:
for each training iteration on a set of unsupervised images do
-
4:
Teacher generates pseudo-labels on images with weak augmentation
-
5:
Student uses pseudo-labels to update network with strong augmentation
-
6:
Teacher network refined using EMA from update student network
-
7:
end for
|