def valid(self):
self.model.eval()
is_, _ = inception_score(model=self.model,
N=10000,
splits=10,
cuda=self.cuda_mode,
resize=True,
SNGAN=True)
with torch.no_grad():
if self.cuda_mode:
z_ = self.fixed_noise.cuda()
else:
z_ = self.fixed_noise
x_gen = self.model.forward(z_)
logits = self.fid_model.forward(
x_gen.cpu(), downsample_=False).detach().numpy()
m = logits.mean(0)
C = np.cov(logits, rowvar=False)
fid = ((self.m - m)**
2).sum() + np.matrix.trace(C + self.C -
2 * sla.sqrtm(np.matmul(C, self.C)))
return fid, is_
print('Min FID:', np.min(history['FID-c']))
print('Epoch with min FID:', np.argmin(history['FID-c']))
if not args.no_plots:
plot_learningcurves(history, 'gen_loss')
plot_learningcurves(history, 'disc_loss')
plot_learningcurves(history, 'gen_loss_minibatch')
plot_learningcurves(history, 'disc_loss_minibatch')
plot_learningcurves(history, 'FID-c')
test_model(model=model,
n_tests=args.n_tests,
cuda_mode=args.cuda,
SNGAN=True)
save_samples(prefix='CIFAR10_SNGAN',
generator=model,
cp_name=args.cp_path.split('/')[-1].split('.')[0],
cuda_mode=args.cuda,
im_size=32,
SNGAN=True)
if args.inception:
print(
inception_score(model,
batch_size=4,
N=args.n_inception,
cuda=args.cuda,
resize=True,
splits=10,
SNGAN=True))
if args.cuda:
model = model.cuda()
print('Cuda Mode is: {}'.format(args.cuda))
history = ckpt['history']
print('Min FID:', np.min(history['FID-c']))
print('Epoch with min FID:', np.argmin(history['FID-c']))
if not args.no_plots:
plot_learningcurves(history, 'gen_loss')
plot_learningcurves(history, 'disc_loss')
plot_learningcurves(history, 'gen_loss_minibatch')
plot_learningcurves(history, 'disc_loss_minibatch')
plot_learningcurves(history, 'FID-c')
test_model(model=model, n_tests=args.n_tests, cuda_mode=args.cuda)
save_samples(prefix='CIFAR10_DCGAN',
generator=model,
cp_name=args.cp_path.split('/')[-1].split('.')[0],
cuda_mode=args.cuda)
if args.inception:
print(
inception_score(model,
N=args.n_inception,
cuda=args.cuda,
resize=True,
splits=10))