torch.set_grad_enabled(False)

lpips_fn = lpips.LPIPS(net='vgg').to(args.device)

lpips_fn.eval()

transform = transforms.Compose([transforms.Resize((256, 256)),

transforms.ToTensor(),

transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5])])

real_dataset = MultiResolutionDataset(path=args.real_path, transform=transform, resolution=256)

real_loader = DataLoader(real_dataset, batch_size=10, shuffle=False)

data_size = len(real_dataset)

print(data_size)

real_tensors = None

for real_tensor in real_loader:

if real_tensors is None:

real_tensors = real_tensor.to(args.device)

else:

real_tensors = torch.cat((real_tensors, real_tensor.to(args.device)), dim=0)

fake_images = os.listdir(args.fake_path)[:1000]

cluster = {}

for i in range(data_size):

cluster[i] = []

# clustering

pbar = tqdm(fake_images, desc='Clustering...')

for fake_image_path in pbar:

image_path = os.path.join(args.fake_path, fake_image_path)

fake_image = Image.open(image_path)

fake_tensor = transform(fake_image).to(args.device)

dists = np.zeros(data_size)

for i in range(data_size):

out = lpips_fn(fake_tensor, real_tensors[i]).view(-1).item()

dists[i] = out

closest_cluster = np.argmin(dists)

cluster[closest_cluster].append(os.path.join(args.fake_path, fake_image_path))

# compute average pairwise distance

print('Clustered as :', [len(cluster[c]) for c in cluster])

cluster_length_list = [len(cluster[c]) for c in cluster]

p = torch.tensor(cluster_length_list) / len(fake_images) + 1e-5

entropy = -1 * (p * p.log10())

print(f'Entropy : {entropy.sum().item():.4f}')

dists = []

cluster_size = args.cluster_size

cluster = {c: cluster[c][:cluster_size] for c in cluster}

total_length = sum([len(cluster[c]) * (len(cluster[c]) - 1) for c in cluster]) // 2

with tqdm(range(total_length), desc='Computing...') as pbar:

for c in cluster:

temp = []

cluster_length = len(cluster[c])

for i in range(cluster_length):

img1 = Image.open(cluster[c][i])

img1 = transform(img1).cuda()

for j in range(i + 1, cluster_length):

img2 = Image.open(cluster[c][j])

img2 = transform(img2).cuda()

pairwise_dist = lpips_fn(img1, img2)

temp.append(pairwise_dist.item())

pbar.update(1)

dists.append(np.mean(temp))

dists = np.array(dists)

print(f'LPIPS of each cluster:')

for i in range(data_size):

print(f'Cluster {i} : {dists[i]:.4f}')

entropy = entropy.numpy()

new_metric = 0

for i in range(data_size):

if np.isnan(dists[i]):

continue

else:

new_metric += entropy[i] * dists[i]

print(f'Intra-Cluster LPIPS : {dists[~np.isnan(dists)].mean():.4f}')