RandWeightedCrop binary weight map

[babbu3682]

Even though I have a weight map with a value of [0, 1], sometime RandWeightedCrop cut the area with only zero.

[wyli]

I couldn't reproduce this issue with this toy example, could you help change it according to your test case?

import matplotlib.pyplot as plt
import numpy as np

from monai.transforms import RandWeightedCropd

label = np.zeros((1, 512, 512))
label[0, 130:170, 60:180] = 1.0
label[0, 160:180, 50:195] = 0.5

num_samples = 5
crop = RandWeightedCropd("label", "w", spatial_size=(256, 256), num_samples=num_samples)
d = {"label": label, "w": (label == label.max()).astype("uint8")}
result = crop(d)

plt.imshow(label[0])
plt.show()

# fig, axs = plt.subplots(1, num_samples)
# for idx in range(num_samples):
#     axs[idx].imshow(result[idx]['label'][0])
# plt.show()

label map:

samples:

[babbu3682]

Here
@wyli

[wyli]

that's expected, with label==label.max(), the only positive weights will be at the lower right corner, which is not feasible for sampling because the window would be outside of the image. so, in this case, the effective weight map is actually all zeros, it'll use a uniform sampling instead. to verify this you can try label[0, 450:480, 450:480] = 3-> label[0, 447:480, 447:480] = 3

[babbu3682]

Hi @wyli !

If I want to crop the images which have the target class, how can I revise your code? because I want to balance the mask class in cropped images.

[babbu3682]

that's expected, with label==label.max(), the only positive weights will be at the lower right corner, which is not feasible for sampling because the window would be outside of the image. so, in this case, the effective weight map is actually all zeros, it'll use a uniform sampling instead. to verify this you can try label[0, 450:480, 450:480] = 3-> label[0, 447:480, 447:480] = 3

Hi @wyli !

If I want to crop the images which have the target class, how can I revise your code? because I want to balance the mask class in cropped images.

[wyli]

Hi @babbu3682 you can add a tiny value to all the locations in the weight map, something like w = (label == label.max()) + 1e-2. ( because it'll not sample from the zero weight regions)