Compute (and backpropagate) fast through FID. You can start by using the minimal working example below or this notebook which fine-tunes BigGAN to minimize FID.
Please let us know if you find errors, I only had 90 min to prepare the code.
The code below uses a function fastfid(imagesA, imagesB) which supports gradients wrt imagesA.
import torch
from fid.fastfid import fastfid
# Compute FID between CIFAR10 mini-batch and noise
from cifar10 import images
batch_size = 64
noise = torch.rand(images[:batch_size].shape)
fid = fastfid(images[:batch_size], noise, gradient_checkpointing=True)
# Construct adversarial examples for FID.
noise = noise.requires_grad_(True)
adam = torch.optim.Adam([noise], 0.001)
steps = images.shape[0]//batch_size
for i in range(steps):
adam.zero_grad()
fid = fastfid(noise, images[i*batch_size:(i+1)*batch_size], gradient_checkpointing=True)
fid.backward()
adam.step()
print("\r[%4i / %4i] fid=%4f"%(i+1, steps, fid), end="")
Note 1. You can disable gradient checkpointing by using gradient_checkpointing=False. This increases memory consumption but may speed up fastfid.
Note 2. You can get further gains by using gradient accumulation and float16 instead of float32. If you choose to do so, you can consult the BigGAN colab example which does both.
Note 3. You can make fastfid faster by precomputing the mean and covariance matrix of the Inception activations for imagesB. If you want to do this you need to modify fid/fastfid.py as done in the BigGAN colab example. I chose to share the above slower code because it works for any dataset out of the box.
> python test_fid.py
| Batch Size | Prev Time | New Time | Prev FID | New FID |
| 8 | 13.9932 | 0.0560 | 263.7587 | 263.7572 |
| 16 | 14.1004 | 0.0632 | 267.7613 | 267.7590 |
| 32 | 13.9612 | 0.1125 | 217.7122 | 217.7117 |
> python test_sqrtm.py
| Batch Size | SciPy Time | New Time | Real Tr | SciPy Tr | New Tr | SciPy Error | New Error |
| 8 | 4.2649 | 0.3275 | 14337.4844 | 14513.4677 | 14337.4854 | -175.9833 | -0.0010 |
| 16 | 4.4771 | 0.0022 | 30551.6914 | 30777.3852 | 30551.6895 | -225.6938 | 0.0020 |
| 32 | 4.4943 | 0.0029 | 64191.7812 | 64496.6673 | 64191.7812 | -304.8860 | 0.0000 |
| 64 | 5.0165 | 0.0040 | 129579.5312 | 129989.2554 | 129579.5156 | -409.7242 | 0.0156 |
| 128 | 4.3869 | 0.0072 | 259934.5469 | 260500.5697 | 259934.5625 | -566.0228 | -0.0156 |
@article{mathiasen2020backpropagating,
title={Backpropagating through Fr$\backslash$'echet Inception Distance},
author={Mathiasen, Alexander and Hvilsh{\o}j, Frederik},
journal={arXiv preprint arXiv:2009.14075},
year={2020}
}