Target -1 is out of bounds

[126zz6]

return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index, label_smoothing) IndexError: Target -1 is out of bounds. How can I solve this problem

[nickuntitled]

Hello @126zz6 for creating an issue about "Target -1 is out of bounds".

From the issue, the IndexError during using torch._C._nn.cross_entropy_loss is from the target (or label) index is outside the range (which means the head pose labels are outside the range -99 and 99 degrees).

I wrote the code that it caused an error like that below.

test = torch.Tensor([[0, 0.5, 0.2], [0.1, 0.3, 0.4]])
gt = torch.LongTensor([-1, 1])
output = class_criterion(test, gt)

It showed.

IndexError: Target -1 is out of bounds.

Next, I saw the CrossEntropyLoss was available on the HPE training code (train.py). When training the model, you have to follow the instruction inside README.md such as

• How to prepare
• How to train below each dataset name (300W_LP, BIWI, or DAD-3DHeads).

The filename list used for training (argument --filename_list) is available inside the folder

• datasets/300W_LP which has 300wlp_list.txt
• datasets/BIWI which has biwi_list.txt, biwi_train_list.txt, and biwi_test_list.txt.
• or datasets/DAD which has train.json.

However, if you want to create your own filename list, you have to make sure that each head pose values (yaw, pitch, and roll) is between -99 and 99 degrees, otherwise, you will see the error Target -1 is out of bounds.