def gen_dataset(args, train=True):
pyro.clear_param_store()
images, labels, _ = load_morphomnist_like(args.data_dir, train=train)
if args.digit_class is not None:
mask = (labels == args.digit_class)
images = images[mask]
labels = labels[mask]
n_samples = len(images)
with torch.no_grad():
thickness, width = model(n_samples, scale=args.scale)
metrics = pd.DataFrame(data={'thickness': thickness, 'width': width})
for n, (thickness, width) in enumerate(tqdm(zip(thickness, width), total=n_samples)):
morph = ImageMorphology(images[n], scale=16)
tmp_img = morph.downscale(np.float32(SetWidth(width)(morph)))
tmp_morph = ImageMorphology(tmp_img, scale=16)
tmp_img = np.float32(SetThickness(thickness)(tmp_morph))
images[n] = morph.downscale(tmp_img)
# TODO: do we want to save the sampled or the measured metrics?
save_morphomnist_like(images, labels, metrics, args.out_dir, train=train)
def gen_dataset(args, train=True):
pyro.clear_param_store()
images, labels, _ = load_morphomnist_like(args.data_dir, train=train)
mask = (labels == args.digit_class)
images = images[mask]
labels = labels[mask]
n_samples = len(images)
with torch.no_grad():
slant, thickness = model(n_samples)
metrics = pd.DataFrame(data={'thickness': thickness, 'slant': slant})
for n, (slant, thickness) in enumerate(
tqdm(zip(slant, thickness), total=n_samples)):
morph = ImageMorphology(images[n], scale=16)
tmp_img = np.float32(SetThickness(thickness)(morph))
tmp_morph = ImageMorphology(tmp_img, scale=1)
tmp_img = np.float32(SetSlant(np.deg2rad(slant))(tmp_morph))
images[n] = morph.downscale(tmp_img)
save_morphomnist_like(images, labels, metrics, args.out_dir, train=train)
def gen_dataset(args, train=True):
pyro.clear_param_store()
images_, labels, _ = load_morphomnist_like(args.data_dir, train=train)
if args.digit_class is not None:
mask = (labels == args.digit_class)
images_ = images_[mask]
labels = labels[mask]
images = np.zeros_like(images_)
n_samples = len(images)
with torch.no_grad():
thickness, intensity = model(n_samples,
scale=args.scale,
invert=args.invert)
metrics = pd.DataFrame(data={
'thickness': thickness,
'intensity': intensity
})
for n, (thickness, intensity) in enumerate(
tqdm(zip(thickness, intensity), total=n_samples)):
morph = ImageMorphology(images_[n], scale=16)
tmp_img = morph.downscale(np.float32(SetThickness(thickness)(morph)))
avg_intensity = get_intensity(tmp_img)
mult = intensity.numpy() / avg_intensity
tmp_img = np.clip(tmp_img * mult, 0, 255)
images[n] = tmp_img
# TODO: do we want to save the sampled or the measured metrics?
save_morphomnist_like(images, labels, metrics, args.out_dir, train=train)
def subsample(filter_fn, source_dir, train: bool, target_dir):
images, labels, metrics = load_morphomnist_like(source_dir, train,
source_dir)
idx = filter_fn(metrics)
save_morphomnist_like(images[idx], labels[idx], metrics[idx], train,
target_dir)
# target_thickness = 6 if label % 2 == 0 else 1.5
# trf = transforms.SetThickness(target_thickness)
target_thickness = max(1, 2.5 * np.exp(2 * metrics.slant))
trf = transforms.SetThickness(target_thickness)
trf_image = morph.downscale(trf(morph))
trf_metrics = measure_image(trf_image, scale=4, verbose=False)
return trf_image, label, trf_metrics
if __name__ == '__main__':
import os
root_dir = "/vol/biomedic/users/dc315/mnist"
source_dir = os.path.join(root_dir, "original")
target_dir = os.path.join(root_dir, "sub_th3_sl0")
images, labels, metrics = load_morphomnist_like(source_dir, train=False)
# with multiprocessing.Pool() as pool:
pool = None
nrow, ncol = 9, 12
trf_images, trf_labels, trf_metrics = apply_conditional_transformation(
example_fn, images[:nrow * ncol], labels, metrics, pool)
import matplotlib.pyplot as plt
fig, axs = plt.subplots(nrow, ncol)
for i, ax in enumerate(axs.flat):
if i >= len(trf_images):
break
ax.imshow(trf_images[i], cmap='gray_r')
# ax.set_title(f"label: {trf_labels[i]}")
ax.axis('off')