How can i use torchmetrics.image.fid to evaluate the quality of my generated images?

[yuanxiqd]

Based on DCGAN, I write my gan code as follows:

import basic_libraries
from torchmetrics.image.fid import FrechetInceptionDistance

transform = transforms.Compose([
    transforms.Grayscale(num_output_channels=3),
    transforms.Resize(args.image_size),
    transforms.CenterCrop(args.image_size),
    transforms.ToTensor(),
    transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5))
])
trainset = torchvision.datasets.MNIST(root=args.dataroot, train=True, download=True, transform=transform)
testset = torchvision.datasets.MNIST(root=args.dataroot, train=False, download=True, transform=transform)

trainloader = torch.utils.data.DataLoader(
    trainset,
    batch_size=args.batch_size,
    shuffle=True,
    num_workers=args.n_workers
)

testloader = torch.utils.data.DataLoader(
    testset,
    batch_size=args.batch_size,
    shuffle=False,
    num_workers=args.n_workers
)

...

netD = NetD() # discriminator
netG = NetG() # generator
fid = FrechetInceptionDistance(feature=64)
netG.fid=fid
...

fid_scores = []
for epoch in range(args.n_epochs):
    # Step 1: train gan model using trainloader
    ...
    # Step 2: evaluate gan model using testloader
    for image, _ in tqdm(testloader):
        noise = torch.randn(args.batch_size, args.nz, 1, 1, device=device)
        with torch.no_grad():
            fake = netG(noise)
        fake.data = fake.data.mul(0.5).add(0.5)
        image.data = image.data.mul(0.5).add(0.5)
        netG.fid.update(convert_image_dtype(fake, torch.uint8), real=False)
        netG.fid.update(convert_image_dtype(image.to(device),torch.uint8),real=True)
  
    current_fid = netG.fid.compute().item()
    netG.fid.reset()
    fid_scores.append(current_fid)
    print(f'FID = {current_fid:.4f}')

I want to ask whether this code is correct for my purpose: use torchmetrics to evaluate my gan model on MNIST.
Thanks you very much.

[Borda]

Your code for evaluating FID with torchmetrics.image.fid.FrechetInceptionDistance broadly follows the right steps: applying appropriate image transformations, updating the metric with both real and generated images during evaluation, then computing and resetting the metric scores each epoch. Just ensure that your image tensors passed to update() are correctly scaled and converted to uint8, as you are doing. Also, typically FID expects RGB images, so your grayscale-to-3-channel transform is appropriate for MNIST.
One detail: the feature parameter feature=64 in FrechetInceptionDistance(feature=64) is unusual, as the default uses a pre-trained Inception model and the feature extractor is internally handled. You might want to instantiate FrechetInceptionDistance() without the feature argument unless you have a specific custom feature extractor implemented.
Overall, your approach looks sound for evaluating GAN-generated images on MNIST with torchmetrics FID. If you experience any unexpected results, check that your transformed images and device placement are consistent.