Search code examples
pythonpytorchevaluationmnistgenerative-adversarial-network

How to use torcheval.metrics.FrechetInceptionDistance in pytorch for mnist dataset?

I defined a GAN model and I want to evaluate it using FID score. I have 1 channel images which are mnist dataset but this method wants 3 channels images. How can I do to solve this problem?

Solution

  • try to split it into 3 channels before evaulating.

    import torch
import torchvision
from torcheval import metrics

# Load the MNIST dataset
mnist_dataset = torchvision.datasets.MNIST(root='./data', train=True, download=True)

# Convert the 1 channel images to 3 channel images
mnist_dataset.data = mnist_dataset.data.unsqueeze(1)
mnist_dataset.data = mnist_dataset.data.repeat(1, 3, 1, 1)

# Calculate the FID score
fid_score = metrics.FrechetInceptionDistance()(mnist_dataset.data)

# Evaluate the FID score
print('FID score:', fid_score)`enter code here`