Skip to content

Fix RNG for VAE, add main.py, add noised encoder, regression problem.

Kevin Haningerrequested to merge
67-further-validation-anomaly-detection-kevdev into 67-further-validation-anomaly-detection

Hi @nik20652 I took a look at the code this weekend.

  • The variatoinal autoencoders were not variational - there was no noise in the latent sampling (sampling with the mean output is used for eval, not during training).
  • Noise: Additionally, the PRNGKey was not being advanced in the training loop. Unfortunately, it seems the only way to add noise in the training process is to add the PRNGKey to the call function signature on the network. I set up VAE so when the PRNGKey is given, it samples, when not (e.g. during eval) it uses the mean
  • I added a main.py to switch train/eval/visualize
  • I added a few network ideas to try reducing the force information in the input: adding noise to the force (VAENoised), masking out some of the force (VAEMasked), or leaving it out entirely (AEStiff). They all train but haven't been thoroughly eval'd.

My questions:

  • I didn't really see a quantitative and quick eval method. I don't want to look at 6 graphs to try and read out whats happening each time (those are useful for debugging though). I'm thinking like max anomaly score on a success bag and anomaly bag, that should give a quick way to see if the network is (1) fitting nominal case, (2) not fitting anomaly case.

Merge request reports

Datenschutz und Impressum