Raising RAVE's discriminator update period fixes phase-2 instability when the discriminator is too strong

During phase 2, the generator and discriminator can fall out of balance. If the discriminator struggles to differentiate real from synthetic data (or dominates the generator), adjusting the update periodicity of the discriminator can help. RAVE updates the discriminator every N generator updates (default N=2, controlled by update_discriminator_every). Increasing N (e.g., to 4) gives the generator more updates per discriminator step, weakening the discriminator’s advantage. This is the recommended fix when phase 2 shows instability; restart training from a phase-2 checkpoint with the override applied.

Examples

Restart training with --override "model.RAVE.update_discriminator_every = 4" (the default value is 2).

Assessment

You observe that your RAVE phase 2 run shows pred_real and pred_fake converging to the same value (both ~0.5). Is this good or bad? What does it indicate about discriminator/generator balance?