python 计算生成图像数据集的inception score 代码

下面是使用 PyTorch 计算 Inception Score 的示例代码：

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

from torchvision import models, transforms, datasets

import numpy as np
from scipy.stats import entropy

def inception_score(imgs, batch_size=32, resize=False):
    """
    Computes the Inception Score of the generated images imgs
    """
    assert (type(imgs) == np.ndarray)
    assert (imgs.shape[1] == 3)
    assert (np.min(imgs[0]) >= 0 and np.max(imgs[0]) > 10), 'Image values should be in the range [0, 255]'

    N = len(imgs)

    # Set up the Inception model
    inception_model = models.inception_v3(pretrained=True)
    inception_model.eval()

    if resize:
        # Resize the images to 299x299
        transform = transforms.Compose([
            transforms.ToPILImage(),
            transforms.Resize((299, 299)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
        ])
    else:
        # Crop the images to 299x299
        transform = transforms.Compose([
            transforms.ToPILImage(),
            transforms.RandomCrop((299, 299)),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
        ])

    dataset = ImageDataset(imgs, transform)
    dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=False)

    # Compute the activations of the Inception model for all batches of images
    activations = []
    for batch in dataloader:
        pred = inception_model(batch)[0]
        activations.append(pred.detach().cpu().numpy())

    activations = np.concatenate(activations, axis=0)

    # Compute the mean and covariance of the activations
    mu = np.mean(activations, axis=0)
    sigma = np.cov(activations, rowvar=False)

    # Compute the Inception Score
    scores = []
    for i in range(N // batch_size):
        batch = torch.from_numpy(imgs[i * batch_size:(i + 1) * batch_size]).float().cuda()
        pred = inception_model(batch)[0].detach().cpu().numpy()
        p_yx = np.exp(-0.5 * np.sum((pred - mu) ** 2 * np.linalg.inv(sigma), axis=1)) / np.sqrt(
            np.linalg.det(sigma) * (2 * np.pi) ** pred.shape[1])
        scores.append(p_yx)

    scores = np.concatenate(scores, axis=0)
    scores = np.mean(scores.reshape((-1, 1)), axis=0)
    scores = np.exp(entropy(scores))
    return scores

class ImageDataset(torch.utils.data.Dataset):
    def __init__(self, imgs, transform=None):
        self.imgs = imgs
        self.transform = transform

    def __getitem__(self, index):
        img = self.imgs[index]
        if self.transform is not None:
            img = self.transform(img)
        return img

    def __len__(self):
        return len(self.imgs)

这里需要注意的一点是，这个代码中用到了 PyTorch 的 `models.inception_v3` 模型，需要安装 torchvision 库才能使用。另外，在计算 Inception Score 时，建议对生成的图片进行大小调整或裁剪，以便与 Inception v3 的输入要求相匹配。