Training from A - B

[SasiKiranK]

I can observe that the tool is training data for A to B and B to A both.But we would like to use it only for A to B. so it can train fast.

[torzdf]

It doesn't work that way.

As simply as possible:

The model is made up of a shared encoder (which encodes faces it sees into an algorithm) and 2 decoders (one for B and one for A) which take the algorithm from the encoder and try to construct a face from it.

The encoder learns the identity of BOTH faces A + B then the respective decoders learn how to decode the encoded face into A or B If the encoder only ever sees face A, it will not learn encodings to decode to face B.