Evaluating GANs

Deep Learning per Immagini con PyTorch

Michal Oleszak

Machine Learning Engineer

Generating images

num_images_to_generate = 9
noise = torch.randn(num_images_to_generate, 16)

with torch.no_grad():
    fake = gen(noise)

print(f"Generated shape: {fake.shape}")

Generated shape: torch.Size([9, 3, 96, 96])

for i in range(num_images_to_generate):

    image_tensor = fake[i, :, :, :]

    image_permuted = image_tensor.permute(1, 2, 0)

    plt.imshow(image_permuted)
    plt.show()

Create random noise tensor
Pass noise to generator
Iterate over number of images
Slice fake to select i-th image
Rearrange the image dimensions
Plot the image

GAN generations

Sample of Pokemon images generated by a GAN.

Fréchet Inception Distance

Inception: Image classification model
Fréchet distance: Distance measure between two probability distributions
Fréchet Inception Distance:
1. Use Inception to extract features from both real and fake images samples
2. Calculate means and covariances of the features for real and fake images
3. Calculate Fréchet distance between the real and the fake normal distributions
Low FID = fakes similar to training data and diverse
FID < 10 = good

FID in PyTorch

from torchmetrics.image.fid import \
FrechetInceptionDistance


fid = FrechetInceptionDistance(feature=64)


fid.update(
  (fake * 255).to(torch.uint8), real=False)

fid.update(
  (real * 255).to(torch.uint8), real=True)


fid.compute()

tensor(7.5159)

Import FrechetInceptionDistance
Instantiate the FID metric
Update the metric with fake images:
- Multiply by 255
- Parse to torch.uint8
Similarly, update the metric with real images
Compute metric value

Let's practice!

Deep Learning per Immagini con PyTorch