Updated inception and fid scores code.
requires:
- python >= 3.6
- pytorch >= 0.4.0
- tensorflow >= 1.10
features:
- fixed open-ai tensorflow inception v2 code with bs=100 for fast computation (change bs=100 -> 1 is slow, see https://github.com/openai/improved-gan/commit/0b7ed92e47ff7047701be3e10a3bd6363999f5e7)
- tests for both fid and inception v2 and v3 scores
- invoking tensorflow inception v2 score from within pytorch session
notes:
- be careful when computing inception scores, verify (1) nhwc, ncwh, rgb order, (2) normalization, (3) sample size
- see general inception score issues: https://github.com/sbarratt/inception-score-pytorch/issues?q=is%3Aissue+is%3Aclosed
- the official, widely used open-ai tensorflow v2 code appears to have a bug: sbarratt/inception-score-pytorch#1
- read about general issues regarding the inception score: https://arxiv.org/pdf/1801.01973.pdf
- do NOT use the inception score for datasets other than imagenet (even though in the code we used it for cifar), use fid score instead