RuntimeError in compute_loss function: Tensor with 2 elements cannot be converted to Scalar

[lenijwp]

When executing the code, an error occurs in the compute_loss function in hfutil/hf_trainers.py. The error message is as follows:

Error executing job with overrides: ['data=harry', 'data.dataset.split=None', 'model=llama-2-7b', 'model_mode=uld', 'model_mode.num_layer=8', 'unlearn_loss=remember+uniform', 'trainer.strategy=ddp', 'OUTPUTMODELDIR=../Model/unlearned_model/uld']
Traceback (most recent call last):
  File "/home/lenijwp/codes/testUnlearning/unlearning_methods/ULD/scripts/hf_forget_train.py", line 193, in main
    trainer.train()
  File "/home/lenijwp/.conda/envs/uld/lib/python3.10/site-packages/transformers/trainer.py", line 1624, in train
    return inner_training_loop(
  File "/home/lenijwp/.conda/envs/uld/lib/python3.10/site-packages/transformers/trainer.py", line 1961, in innertraining_loop
    tr_loss_step = self.training_step(model, inputs)
  File "/home/lenijwp/.conda/envs/uld/lib/python3.10/site-packages/transformers/trainer.py", line 2902, in training_step
    loss = self.compute_loss(model, inputs)
  File "/home/lenijwp/codes/testUnlearning/unlearning_methods/ULD/uld/hfutil/hf_trainers.py", line 93, in compute_loss
    'trainloss/loss': loss.item(),
RuntimeError: a Tensor with 2 elements cannot be converted to Scalar

The problematic code seems:

logitems = {
    'trainloss/loss': loss.item(),
    'trainloss/forgetloss': forgetloss.item(),
    'trainloss/retainloss': retainloss.item()
}

To resolve the issue, I needed to replace it with:

logitems = {
    'trainloss/loss': loss.sum().item(),
    'trainloss/forgetloss': forgetloss.sum().item(),
    'trainloss/retainloss': retainloss.sum().item()
}