RAVE training is monitored via TensorBoard distance, fidelity, and adversarial-loss logs

RAVE logs training progress to TensorBoard under the run directory. To track reconstruction error, watch the distance and validation logs; they should decrease steadily in phase 1. When phase 2 starts, these values increase — this is expected as the GAN shifts the optimization landscape. The fidelity logs (80–99%) show how many latent dimensions are needed to explain each fraction of dataset variance. The loss_dis, loss_gen, pred_true, pred_fake logs appear only in phase 2 and are harder to read, as most GAN losses are.

Examples

Connect TensorBoard: tensorboard --logdir /model/out. Look at the distance and validation logs for reconstruction error; expect them to rise when phase 2 begins.

Assessment

During phase 2, a user’s loss_dis drops to near zero and stays there. What does this indicate about the discriminator/generator balance and what should they adjust?