def get_inception_score():
all_samples = []
samples = torch.randn(N_SAMPLES, N_LATENT)
for i in xrange(0, N_SAMPLES, 100):
batch_samples = samples[i:i+100].cuda(0)
all_samples.append(gen(batch_samples).cpu().data.numpy())
all_samples = np.concatenate(all_samples, axis=0)
return inception_score(torch.from_numpy(all_samples), resize=True, cuda=True)
def is_ch1():
class IgnoreLabelDataset(torch.utils.data.Dataset):
def __init__(self, orig):
self.orig = orig
def __getitem__(self, index):
return self.orig[index]
def __len__(self):
return len(self.orig)
'''
import torchvision.datasets as dset
import torchvision.transforms as transforms
cifar = dset.CIFAR10(root='data/', download=True,
transform=transforms.Compose([
transforms.Scale(32),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
)
IgnoreLabelDataset(cifar)
'''
p = 0.5
bernoulli = torch.distributions.Bernoulli(torch.tensor([p]))
netG = GeneratorDCGAN_cifar(z_dim=100, model_dim=64, num_classes=10).cuda()
netG.load_state_dict(
torch.load('results/cifar_10/main/d500_i40000_10/netGS_80000.pth'))
sample_list = []
for class_id in range(10):
noise = bernoulli.sample((5000, 100)).view(5000, 100).cuda()
label = torch.full((5000, ), class_id, dtype=torch.long).cuda()
sample = netG(noise, label)
sample = sample.view(5000, 32, 32)
sample = sample.cpu().data.numpy()
sample_list.append(sample)
samples = np.transpose(np.array(sample_list), [1, 0, 2, 3])
samples = np.reshape(samples, [50000, 1, 32, 32])
samples = np.repeat(samples, 3, axis=1)
print("Calculating Inception Score...")
print(
inception_score(IgnoreLabelDataset(samples),
cuda=True,
batch_size=32,
resize=True,
splits=10))
