def accumulate_inception_activations(sample, net, num_inception_images=50000):
pool, logits, labels = [], [], []
while (torch.cat(logits, 0).shape[0] if len(logits) else 0) < num_inception_images:
with torch.no_grad():
images, labels_val = sample()
pool_val, logits_val = net(images.astype("float32"))
pool += [pool_val]
logits += [F.softmax(logits_val, 1)]
labels += [labels_val]
return torch.cat(pool, 0), torch.cat(logits, 0), torch.cat(labels, 0)
def forward(self, x, y=None):
# Apply convs
theta = self.theta(x)
phi = F.max_pool2d(self.phi(x), [2, 2])
g = F.max_pool2d(self.g(x), [2, 2])
# Perform reshapes
theta = theta.view(-1, self.ch // 8, x.shape[2] * x.shape[3])
phi = phi.view(-1, self.ch // 8, x.shape[2] * x.shape[3] // 4)
g = g.view(-1, self.ch // 2, x.shape[2] * x.shape[3] // 4)
# Matmul and softmax to get attention maps
beta = F.softmax(torch.bmm(theta.transpose(1, 2), phi), -1)
# Attention map times g path
o = self.o(
torch.bmm(g, beta.transpose(1, 2)).view(-1, self.ch // 2,
x.shape[2], x.shape[3]))
return self.gamma * o + x
def run(config):
# Get loader
config['drop_last'] = False
loaders = utils.get_data_loaders(**config)
# Load inception net
net = inception_utils.load_inception_net(parallel=config['parallel'])
pool, logits, labels = [], [], []
device = 'cuda'
for i, (x, y) in enumerate(tqdm(loaders[0])):
x = x.to(device)
with torch.no_grad():
pool_val, logits_val = net(x)
pool += [np.asarray(pool_val)]
logits += [np.asarray(F.softmax(logits_val, 1))]
labels += [np.asarray(y)]
pool, logits, labels = [
np.concatenate(item, 0) for item in [pool, logits, labels]
]
# uncomment to save pool, logits, and labels to disk
# print('Saving pool, logits, and labels to disk...')
# np.savez(config['dataset']+'_inception_activations.npz',
# {'pool': pool, 'logits': logits, 'labels': labels})
# Calculate inception metrics and report them
print('Calculating inception metrics...')
IS_mean, IS_std = inception_utils.calculate_inception_score(logits)
print('Training data from dataset %s has IS of %5.5f +/- %5.5f' %
(config['dataset'], IS_mean, IS_std))
# Prepare mu and sigma, save to disk. Remove "hdf5" by default
# (the FID code also knows to strip "hdf5")
print('Calculating means and covariances...')
mu, sigma = np.mean(pool, axis=0), np.cov(pool, rowvar=False)
print('Saving calculated means and covariances to disk...')
np.savez(config['dataset'].strip('_hdf5') + '_inception_moments.npz', **{
'mu': mu,
'sigma': sigma
})