Training Stuck at Phase Transition (Adversarial Phase Not Triggering)

[LaScienceMusicale]

Hi !

I'm encountering an issue while training a RAVE model: the training gets stuck when transitioning from the autoencoding phase (Phase 1) to the adversarial phase. Even after reaching PHASE_1_DURATION = 1000000, the model does not properly enter Phase 2.

rave train --name lsm --db_path /home//lsmpp --out_path ./
--config /home//miniconda3/envs/RAVE/lib/python3.9/site-packages/rave/configs/v3.gin
--config /home/***/miniconda3/envs/RAVE/lib/python3.9/site-packages/rave/configs/noise.gin
--save_every 5000 --channels 1 --gpu 0
--override "model.RAVE.use_discriminator = True"

However, even after reaching 1,000,000 epochs, the model does not switch to adversarial training. Instead, it seems to remain in Phase 1

Thanks a lot,

Regards,

[domkirke]

Hello! Could you provide some tensorboard screenshots by any chance?