def test_optimiser_output(tmpdir, simple_net):
"""Save image per iteration of the optimiser."""
model, objective, image = setup_optimiser(simple_net)
n_iterations = 10
opt = sumie.Optimiser(image, model, objective)
opt.run(iterations=10, output=Path(tmpdir))
search = str(tmpdir.join('*.jpg'))
assert len(glob(search)) == 11
def test_optimiser_output_str(tmpdir, simple_net):
"""Save image to new folder of type string."""
model, objective, image = setup_optimiser(simple_net)
n_iterations = 10
new_folder = 'tmp'
opt = sumie.Optimiser(image, model, objective)
opt.run(iterations=10, output=new_folder)
search = new_folder + '/*.jpg'
assert len(glob(search)) == 11
def test_optimiser(simple_net):
"""Optimiser class optimises an image given an objective."""
model, objective, image = setup_optimiser(simple_net)
original_image, start_value = get_state(model, image, objective)
opt = sumie.Optimiser(image, model, objective)
opt.run(iterations=10)
new_image, end_value = get_state(model, image, objective)
assert torch.any(new_image != original_image)
assert end_value > start_value
def test_optimiser_history(simple_net):
"""Optimiser stores objective history."""
model, objective, image = setup_optimiser(simple_net)
n_iterations = 10
opt = sumie.Optimiser(image, model, objective)
opt.run(iterations=10)
new_image, end_value = get_state(model, image, objective)
# +1 as we get the initial and final losses
assert len(opt.history) == n_iterations + 1
assert opt.history[-1] == end_value
def run(target, selected_module, average=True):
device = 'cuda'
im = sumie.Image(imsize, param='fft', transforms=[
sumie.transforms.PositionJitter(8),
sumie.transforms.ScaleJitter(1.01),
sumie.transforms.RotationJitter(0.1),
sumie.transforms.PositionJitter(8),
sumie.transforms.Interpolate(0.1),
sumie.transforms.Normalise(),
])
im.to(device)
content = Direction(selected_module, target.detach())
opt = sumie.Optimiser(im, model, content)
opt.add_callback(change_scale)
opt.run(iterations=512, lr=0.05, progress=True)
return im.get_image()
all_target = monitor.values
values = all_target.detach().squeeze(0).cpu().numpy()
values = values.reshape((values.shape[0], -1)).T
n_clusters = 5
kmeans = KMeans(n_clusters=n_clusters, random_state=0).fit(values)
print(kmeans.cluster_centers_.shape)
for i in range(n_clusters):
cluster = torch.Tensor(kmeans.cluster_centers_[i, :])
cluster = cluster[None, :, None, None]
im = sumie.Image((1000, 300),
param='fft',
transforms=[
sumie.transforms.PositionJitter(8),
sumie.transforms.ScaleJitter(1.01),
sumie.transforms.RotationJitter(0.1),
sumie.transforms.PositionJitter(8),
sumie.transforms.Interpolate(0.1),
sumie.transforms.Normalise(),
])
im.to(device)
content = Direction(selected_module, cluster.to(device))
opt = sumie.Optimiser(im, model, content)
opt.add_callback(change_scale)
opt.run(iterations=512, lr=0.05, progress=True)
sumie.io.save(im.get_image(), f'tmp/output_{i:03}.png')