disent
disent copied to clipboard
nmichlo
🧶 Modular VAE disentanglement framework for python built with PyTorch Lightning ▸ Including metrics and datasets ▸ With strongly supervised, weakly supervised and unsupervised methods ▸ Easily config...
Hi Nathan, Just curious would it be feasible to add MNNIST to the data options and add DMS (from Beta-VAE paper https://openreview.net/forum?id=Sy2fzU9gl) to metrics. Just started learning PyTorch Lightning, curious...
**Is your feature request related to a problem? Please describe.** Model saving and checkpointing is currently disabled for `experiment/run.py` This was due to old pickling errors and the extensive use...
**Is your feature request related to a problem? Please describe.** The current examples are very limited and only show how to use `disent`. **Describe the solution you'd like** Add examples...
**Is your feature request related to a problem? Please describe.** Current metrics require that you provide a representation function. This is inconvenient and always repeated. Metrics also always require that...
Documentation is missing key framework features - augmentations - schedules - creating your own framework - creating your own models - creating your own datasets - visualisations
Tests are currently lacking across disent - data - datasets - schedule - metrics - transformations
The betatcvae implementation is definitely not correct. - loss scaling is not implemented - sane defaults for config
The InfoVAE implementation is probably not correct. - loss scaling might not be correctly implemented - sane defaults for config - irq kernel was removed
The DipVAE implementation is probably not correct. - loss scaling is not be correctly implemented - sane defaults for config
**Describe the bug** The downloads for MPI3D and dSprites do not work automatically **To Reproduce** ``` from disent.dataset.data import Mpi3dData data = Mpi3dData(in_memory=True) ``` Leads to ``` FileNotFoundError: [Errno 2]...
