Generate target latent state using the right encoder

[antoine-galataud]

In the forward prediction objective and when observations vector is used, it looks like the wrong encoder is used.

smart/src/models/components/mingpt.py

Line 411 in a75de0a

next_state_embeddings = self.state_encoder(states).detach()

Instead it should be the momentum encoder self.target_state_encoder to correctly generate target latent state embeddings.

[ycsun2017]

@antoine-galataud Thank you for the comment! This is indeed a typo. But don't worry, since this line is under the "vector_obs" condition, it will not affect the current experiments with pixel obs. We apologize for the confusion and have removed the lines related to this unused condition.