pytorch inception score

Inception Score is a measure of the quality of generated images. It was proposed in the paper "Improved Techniques for Training GANs" by Salimans et al. The score is computed by passing generated images through a pre-trained Inception v3 network and computing the KL divergence between the conditional label distribution (i.e., the predicted class probabilities) and the marginal label distribution (i.e., the class probabilities of the training set). A higher score indicates that the generated images have better quality and diversity.

Here's an implementation of Inception Score in PyTorch:

import torch
import torch.nn as nn
import torchvision.models as models
from torch.utils.data import DataLoader

def inception_score(images, batch_size=32, resize=True):
    """Computes the Inception Score of the generated images."""
    # Load the Inception v3 model
    inception = models.inception_v3(pretrained=True, transform_input=False).cuda()
    inception.eval()

    # Resize and normalize the images
    if resize:
        images = nn.functional.interpolate(images, size=(299, 299), mode='bilinear', align_corners=False)
    images = (images - 0.5) / 0.5

    # Compute the conditional label distribution
    preds = []
    with torch.no_grad():
        for i in range(0, len(images), batch_size):
            batch = images[i:i+batch_size].cuda()
            pred = inception(batch)
            preds.append(pred)
    preds = torch.cat(preds, dim=0)
    py = torch.softmax(preds, dim=1)

    # Compute the marginal label distribution
    py = py.mean(dim=0, keepdim=True)
    kl = (py * (torch.log(py) - torch.log(torch.tensor(1.0/py.shape[1])))).sum().item()

    # Compute the Inception Score
    iscore = torch.exp(kl).item()

    return iscore

The input `images` should be a PyTorch tensor of shape `(N, C, H, W)` representing `N` generated images with `C` channels, height `H` and width `W`. The function returns a scalar value representing the Inception Score.