def visualize_generator(self,
epoch,
to_mlflow=False,
is_remote=False,
vmin=0,
vmax=1,
*args,
**kwargs):
# Check how the generator is doing by saving G's output on fixed_noise
path = TMP_IMAGES_DIR
# Evaluation mode
self.generator.eval()
with torch.no_grad():
fake = self.generator(self.fixed_noise)[0].detach().cpu()
img = vutils.make_grid(fake,
padding=20,
normalize=False,
range=(vmin, vmax))
path = f'{path}/epoch{epoch}.png'
vutils.save_image(img, path)
if to_mlflow:
log_artifact(path, 'images', is_remote=is_remote)
os.remove(path)
def forward_and_save_one_image(self, inp_image, label, epoch, path=TMP_IMAGES_DIR, to_mlflow=False,
is_remote=False, vmin=0, vmax=1):
"""
Reconstructs one image and writes two images (original and reconstructed) in one figure to :param path.
:param inp_image: Image for evaluation
:param label: Label of image
:param epoch: Epoch
:param path: Path to save image to
"""
# Evaluation mode
self.generator.eval()
self.encoder.eval()
with torch.no_grad():
# Format input batch
inp = inp_image.to(self.device)
# Forward pass
z_mean, z_logvar, _, _ = self.encoder(inp)
# z_hat = z_mean + z_logvar * torch.randn(z_mean.size()).to(self.device)
if len(z_mean.size()) == 1:
z_mean = z_mean.view(1, z_mean.size(0))
x_rec, _, _ = self.generator(z_mean)
inp_image = inp_image.to('cpu')
x_rec = x_rec.to('cpu')
fig, ax = plt.subplots(nrows=1, ncols=2, figsize=(10, 5))
ax[0].imshow(inp_image.numpy()[0, 0, :, :], cmap='gray', vmin=vmin, vmax=vmax)
ax[1].imshow(x_rec.numpy()[0, 0, :, :], cmap='gray', vmin=vmin, vmax=vmax)
path = f'{path}/epoch{epoch}_label{int(label)}.png'
plt.savefig(path)
plt.close(fig)
if to_mlflow:
log_artifact(path, 'images', is_remote=is_remote)
os.remove(path)
def forward_and_save_one_image(self,
inp_image,
label,
epoch,
path=TMP_IMAGES_DIR,
to_mlflow=False,
is_remote=False):
"""
Reconstructs one image and writes two images (original and reconstructed) in one figure to :param path.
:param is_remote:
:param to_mlflow:
:param inp_image: Image for evaluation
:param label: Label of image
:param epoch: Epoch
:param path: Path to save image to
"""
# Evaluation mode
self.eval()
with torch.no_grad():
# Format input batch
inp = inp_image.to(self.device)
# Forward pass
output = self(inp)
output_img = output.to('cpu')
fig, ax = plt.subplots(nrows=1, ncols=2, figsize=(10, 5))
ax[0].imshow(inp_image.numpy()[0, 0, :, :],
cmap='gray',
vmin=0,
vmax=1)
ax[1].imshow(output_img.numpy()[0, 0, :, :],
cmap='gray',
vmin=0,
vmax=1)
path = f'{path}/epoch{epoch}_label{int(label)}.png'
plt.savefig(path)
plt.close(fig)
if to_mlflow:
log_artifact(path, 'images', is_remote=is_remote)
os.remove(path)
def forward_and_save_one_image(self,
inp_image,
label,
epoch,
path=TMP_IMAGES_DIR,
to_mlflow=False,
is_remote=False):
"""
Save random sample of original and reconstructed image
:param inp_image: input original image
:param label: label of image
:param epoch: number of epoch
:param device: device type: CPU or CUDA GPU
:param path: path to save
"""
self.eval()
with torch.no_grad():
inp = inp_image.to(self.device)
output, _, _ = self(inp)
output_img = output.to('cpu')
fig, ax = plt.subplots(nrows=1, ncols=2, figsize=(10, 5))
ax[0].imshow(inp_image.numpy()[0, 0, :, :],
cmap='gray',
vmin=0,
vmax=1)
ax[1].imshow(output_img.numpy()[0, 0, :, :],
cmap='gray',
vmin=0,
vmax=1)
path = f'{path}/epoch{epoch}_label{int(label)}.png'
plt.savefig(path)
plt.close(fig)
if to_mlflow:
log_artifact(path, 'images', is_remote=is_remote)
os.remove(path)