前言

莫名其妙发现了这个warning，很奇怪。

发现了，原来是有个一个label 超过了n_classes。

正文

1. Assertion t < n_classes failed

发现这样的问题：

/lib/python3.12/site-packages/torch/nn/modules/conv.py:456: UserWarning: Using padding='same' with even kernel lengths and odd dilation may require a zero-padded copy of the input be created (Triggered internally at ../aten/src/ATen/native/Convolution.cpp:1031.)
  return F.conv2d(input, weight, bias, self.stride,
../aten/src/ATen/native/cuda/Loss.cu:250: nll_loss_forward_reduce_cuda_kernel_2d: block: [0,0,0], thread: [9,0,0] Assertion `t < n_classes` failed.
../aten/src/ATen/native/cuda/Loss.cu:250: nll_loss_forward_reduce_cuda_kernel_2d: block: [0,0,0], thread: [10,0,0] Assertion `t < n_classes` failed.
../aten/src/ATen/native/cuda/Loss.cu:250: nll_loss_forward_reduce_cuda_kernel_2d: block: [0,0,0], thread: [22,0,0] Assertion `t < n_classes` failed.
../aten/src/ATen/native/cuda/Loss.cu:250: nll_loss_forward_reduce_cuda_kernel_2d: block: [0,0,0], thread: [24,0,0] Assertion `t < n_classes` failed.
../aten/src/ATen/native/cuda/Loss.cu:250: nll_loss_forward_reduce_cuda_kernel_2d: block: [0,0,0], thread: [25,0,0] Assertion `t < n_classes` failed.
Training:   0%|          | 0/79 [00:00<?, ?it/s]

仔细检查了之后，发现有个255混在了里面，很神奇。

不同类别数: 10
类别标签: [  0   1   2   3   4   5   6   7   8 255]
每个类别的数量: [500 500 500 500 500 500 500 500 500 500]

2. 解决办法

发现了之前是这样的code：

# Adjust labels to be zero-based
y_train = y_train - 1  # Now labels are from 0 to 9
y_test = y_test - 1

可能是超过int范围了，从0 变成255了，把这个代码注释掉，就变回这样了

不同类别数: 10
类别标签: [0 1 2 3 4 5 6 7 8 9]
每个类别的数量: [500 500 500 500 500 500 500 500 500 500]

成功，完事儿！问题消失！

总结

以后遇见新的dataset，一定要好好看看label是啥，不然就会发现这样的事情。

参考

[1] ChatGPT
[2] 自己

Q.E.D.