Closed tianyu-z closed 3 years ago
Now, seg advent and masker advent can call different kinds of GAN_type by dis.m.gan_type and dis.s.gan_type
Training exps that have been running for 25 hrs: WGAN with mask only: https://www.comet.ml/tianyu-z/omnigan/01f839bc294a4e569e7f528a1a7017b0?experiment-tab=chart&showOutliers=true&smoothing=0&transformY=smoothing&xAxis=step
WGAN_gp with mask only: https://www.comet.ml/tianyu-z/omnigan/efacf54c359d4cd1ba21e8df52065b73?experiment-tab=chart&showOutliers=true&smoothing=0&transformY=smoothing&xAxis=step
WGAN_gp with m+s: https://www.comet.ml/tianyu-z/omnigan/bfa8d1b4ff444e7b93a7ddddb7fd756b?experiment-tab=chart&showOutliers=true&smoothing=0&transformY=smoothing&xAxis=step
WGAN_norm with mask only: https://www.comet.ml/tianyu-z/omnigan/1c034e4d64f94cba912e028aa16d0c0e?experiment-tab=chart&showOutliers=true&smoothing=0&transformY=smoothing&xAxis=step
WGAN_norm with m+s: https://www.comet.ml/tianyu-z/omnigan/c81c59dd659343deac7f30de6f7384b9?experiment-tab=chart&showOutliers=true&smoothing=0&transformY=smoothing&xAxis=step
I will run eval script before the weekly meeting and update links here.
@vict0rsch Fixed
Eval exps:
WGAN_norm_masker (I think it's pretty well): https://www.comet.ml/tianyu-z/omnigan/cb25e99931a84d0ca92eb32c50e43705?experiment-tab=images&imageId=f6e1b3245e164a4e98de069077e9fa60
WGAN_gp_masker https://www.comet.ml/tianyu-z/omnigan/126265d60b514119a34edcee3461354e?experiment-tab=images
WGAN_masker: https://www.comet.ml/tianyu-z/omnigan/2b2cd5fdac4e49b38288b5f8f6b09c64?experiment-tab=images
WGAN_norm_m+s: https://www.comet.ml/tianyu-z/omnigan/e5de0daf511848b7822ec560b82bc7a9?experiment-tab=images
WGAN_gp_m+s: https://www.comet.ml/tianyu-z/omnigan/2367f1d4b7944ff480a6852a606fc27d?experiment-tab=images
WGAN_m+s: https://www.comet.ml/tianyu-z/omnigan/d187b94593b64652bb6762ae6b9b1945?experiment-tab=images
Comet.ml - Supercharging Machine LearningComet lets you track code, experiments, and results on ML projects. It’s fast, simple, and free for open source projects.
Comet.ml - Supercharging Machine LearningComet lets you track code, experiments, and results on ML projects. It’s fast, simple, and free for open source projects.
Comet.ml - Supercharging Machine LearningComet lets you track code, experiments, and results on ML projects. It’s fast, simple, and free for open source projects.
Comet.ml - Supercharging Machine LearningComet lets you track code, experiments, and results on ML projects. It’s fast, simple, and free for open source projects.
Comet.ml - Supercharging Machine LearningComet lets you track code, experiments, and results on ML projects. It’s fast, simple, and free for open source projects.
Comparing WGAN_norm_masker with WGAN_norm_m+s: https://www.comet.ml/tianyu-z/omnigan/e5de0daf511848b7822ec560b82bc7a9/cb25e99931a84d0ca92eb32c50e43705/compare?experiment-tab=images
Comet.ml - Supercharging Machine LearningComet lets you track code, experiments, and results on ML projects. It’s fast, simple, and free for open source projects.
I agree with you masking alone with WGAN+spectral norm is the best. We need to make the segmentation work! :)
I added an option for ADVENT discriminator:
dis.m.gan_type now can be "GAN", "WGAN", "WGAN_gp", "WGAN_norm" and I added a "rmsprop" option for our optimizer which works for "WGAN" (Only "WGAN", not "WGAN_gp" and "WGAN_norm").