Have you ever tried to generate pseudo-labels online?

[Chen-Song]

In STAC, the pseudo-labels are predicted in an offline manner, i.e., after training a network in labeled data and then using it to predict pseudo-labels. It is a multi-stage training manner. However, in FixMatch, also your work, the pseudo-labels are predicted in an online manner, i.e., in a mini-batch, the pseudo-labels are generated. It is a one-stage training manner.

[kihyuks]

We haven't tried using online pseudo-label generation approach in the current work but opted for offline generation as it was simpler and faster to implement. It is definitely a viable idea to generate pseudo-labels online as in FixMatch and would be interested to see how it performs.

[Chen-Song]

Thanks!