sndnyang / inception_score_fid

Tensorflow and PyTorch of evaluating FID, Inception Score only has Tensorflow Version

Compute the Inception Score and FID using TF and PyTorch

I give the demo in jupyter notebook files to show the usage and difference. PyTorch code is faster. FID values are also different.

The difference on cifar10:

dataset	TF	PyTorch	gap
Train	0.08768	0.20264
Test	2.20365	5.12468
Fake samples	16.0426	9.87148

Tensorflow: please check demo.py, or eval_is_fid_tf.ipynb, using inception.py and fid.py

# [n, 32, 32, 3]
fid = get_fid_score(test_images, test_ims)
score, std = get_inception_score(test_images, splits=splits)

PyTorch: please check eval_is_fid_pytorch.ipynb, using fid_score.py

# [n, 3, 32, 32]
calculate_fid_score(feed_imgs2, test_imgs)
No Inception Score

A set of generated images (10000 images) npy file

Tensorflow Code

The code is based on the repo released by IGEBM

It works fine, but

1. sometimes TF is hard to run on GPU (CPU is so slow). Suggest to use conda install tensorflow in conda
2. or occupies all memory of GPU so that you can't trace your training process. I have added the config in the code [inception score](https://github.com/sndnyang/inception_score_fid/blob/master/inception.py#L25], FID but I can't guarantee.

As a refer: Inception Score Row 231 and FID Row 39

PyTorch Code

mainly based on https://github.com/mseitzer/pytorch-fid this one computes based on the files stored in disk (two directorys)

As a refer: FID pytorch