def disrupt_stimuli(save_folder, inputs_and_targets, encoder, roi_mask,
target_roi, loss_method):
roi_mask = torch.from_numpy(roi_mask.astype(np.uint8))
towards_target = True if loss_method == 'towards' else False
if not target_roi:
loss_func = roi_loss_func(None, towards_target)
else:
loss_func = roi_loss_func(roi_mask, towards_target)
for input_and_target in tqdm(inputs_and_targets):
orig_image = input_and_target['stimulus_path']
orig_image = utils.image_to_tensor(orig_image)
target = input_and_target['target_voxels']
target = torch.from_numpy(target)
if not target_roi:
target = target[roi_mask]
else:
with torch.no_grad():
orig_voxels = encoder(orig_image.unsqueeze(0)).squeeze(0)
target[1 - roi_mask] = orig_voxels[1 - roi_mask]
disrupted_image = deepdream(orig_image, target, encoder, loss_func)
metrics = loss_metrics(orig_image, disrupted_image, target, encoder,
roi_mask if target_roi else None)
disrupted_image = utils.tensor_to_image(disrupted_image)
save_disrupted_image(save_folder, input_and_target, disrupted_image,
metrics)
def disrupt_stimulus(stimulus, target, encoder, roi_mask, towards_target,
random):
roi_mask = torch.from_numpy(roi_mask.astype(np.uint8))
loss_func = roi_loss_func(roi_mask, towards_target)
with torch.no_grad():
if random:
encoder = encoder.random_weights()
orig_voxels = encoder(stimulus.unsqueeze(0)).squeeze(0)
if random and not towards_target:
target = orig_voxels
else:
target[1 - roi_mask] = orig_voxels[1 - roi_mask]
noise = (torch.rand_like(target) - 0.5) * 1e-5
disrupted = deepdream(stimulus, target + noise, encoder, loss_func)
metrics = loss_metrics(stimulus, disrupted, target, encoder, roi_mask)
return disrupted, metrics
def generate_stimulus(target, encoder, towards_target):
loss_func = roi_loss_func(roi_mask=None, towards_target=towards_target)
noise = utils.sample_imagenet_noise(resolution=256)
generated = deepvisualize(noise, target, encoder, loss_func)
metrics = loss_metrics(noise, generated, target, encoder, roi_mask=None)
return generated, metrics
encoder.cuda()
generator.cuda()
shutil.rmtree(save_folder, ignore_errors=True)
os.mkdir(save_folder)
shutil.copyfile(image_path, os.path.join(save_folder, 'original.jpg'))
image = image_to_tensor(image_path, resolution=256)
with torch.no_grad():
if torch.cuda.is_available():
target = encoder(image.unsqueeze(0).cuda()).squeeze(0).cpu()
else:
target = encoder(image.unsqueeze(0)).squeeze(0)
loss_func = roi_loss_func(roi_mask=None, towards_target=True)
gen_images = []
fig, axs = plt.subplots(len(alphas),
len(decays),
squeeze=False,
figsize=(len(decays) * 10, len(alphas) * 5))
for i, alpha in tqdm(enumerate(alphas)):
for j, decay in enumerate(decays):
gen_image, _, loss, losses = optimize(generator,
encoder,
target,
loss_func,
alpha=alpha,
decay=decay)
gen_images.append(to_pil_image(gen_image))