- Add argument
--save-stats
allowing to compute dataset statistics and save them as an.npz
file (#80). The.npz
file can be used in subsequent FID computations instead of recomputing the dataset statistics. This option can be used in the following way:python -m pytorch_fid --save-stats path/to/dataset path/to/outputfile
.
- Do not use
os.sched_getaffinity
to get number of available CPUs on Windows, as it is not available there (232b3b14, #84). - Do not use Inception model argument
pretrained
, as it was deprecated in torchvision 0.13 (#88).
- Add argument
--num-workers
to select number of dataloader processes (#66). Defaults to 8 or the number of available CPUs if less than 8 CPUs are available.
- Load images using a Pytorch dataloader, which should result in a speed-up. (#47)
- Support more image extensions (#53)
- Improve tooling by setting up Nox, add linting and test support (#52)
- Add some unit tests
- Fixed software license string in
setup.py
Initial release as a pypi package. Use pip install pytorch-fid
to install.