yiyixuxu / polyloss-pytorch Goto Github PK

Hello,

While making one-hot encoding of labels. I got these types of errors at scatter PyTorch.

/opt/conda/conda-bld/pytorch_1646756402876/work/aten/src/ATen/native/cuda/ScatterGatherKernel.cu:276: operator(): block: [66,0,0], thread: [5,0,0] Assertion idx_dim >= 0 && idx_dim < index_size && "index out of bounds" failed.

Note : My labels is a mask with classes=2 and ignore_index=255 (label.unique()=[0,1,255]

torch.nn.CrossEntropyLoss expect you to pass raw, unnormalized logit, however, either logit or probability is passing into cross entropy loss. If the probability is used will lead to unintended result.

README example is not reproducible as target_idx does not exist.

import torch
from PolyLoss import to_one_hot, PolyLoss
# Example of target in one-hot encoded format
loss = PolyLoss(softmax=True)
B, C, H, W = 2, 5, 3, 3
input = torch.rand(B, C, H, W, requires_grad=True)
target = torch.randint(low=0, high=C - 1, size=(B, H, W)).long()
target = to_one_hot(target_idx[:, None, ...], num_classes=C)
output = loss(input, target)
output.backward()

...

Traceback (most recent call last):
  File "/Users/glennjocher/PycharmProjects/yolov5-pro/venv/lib/python3.8/site-packages/IPython/core/interactiveshell.py", line 3457, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-4-0534d4642825>", line 8, in <module>
    target = to_one_hot(target_idx[:, None, ...], num_classes=C)
NameError: name 'target_idx' is not defined

Hello, Thank you for your work. I want to use the poly loss for image segmentation. I have input [B,C,H,W] and mask as [B,H,W]. I have ignore_indexes too in masks. how can I use your provided loss function correctly? Thank you