Tensor Size Mismatch When Train A Technique Recognizer

[Artanisax]

Hi, thanks for the wonderful datasets and downstream task experiments.

I've been trying to train a Technique Recognizer following the instructions in the repository. However, I keep recieving errors like:

Traceback (most recent call last):
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/tasks/run.py", line 19, in <module>
    run_task()
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/tasks/run.py", line 14, in run_task
    task_cls.start()
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/utils/commons/base_task.py", line 230, in start
    trainer.fit(cls)
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/utils/commons/trainer.py", line 122, in fit
    self.run_single_process(self.task)
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/utils/commons/trainer.py", line 186, in run_single_process
    self.train()
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/utils/commons/trainer.py", line 286, in train
    pbar_metrics, tb_metrics = self.run_training_batch(batch_idx, batch)
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/utils/commons/trainer.py", line 332, in run_training_batch
    output = task_ref.training_step(*args)
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/utils/commons/base_task.py", line 106, in training_step
    loss_ret = self._training_step(sample, batch_idx, optimizer_idx)
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/research/singtech/te_task.py", line 215, in _training_step
    loss_output, _ = self.run_model(sample)
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/research/singtech/te_task.py", line 238, in run_model
    self.add_tech_group_loss(output['tech_logits'], techs, tech_ids, losses)
  File "/data1/cse12110524/SSC/methods/GTSinger/Tech-Recognition/research/singtech/te_task.py", line 249, in add_tech_group_loss
    tech_losses_i = F.binary_cross_entropy_with_logits(tech_logits[b_idx, :, tech_ids[b_idx]], techs[b_idx, :, tech_ids[b_idx]], reduction='none')  # [T, len(tech_ids)]
  File "/home/cse12110524/miniconda3/envs/gttr/lib/python3.9/site-packages/torch/nn/functional.py", line 3193, in binary_cross_entropy_with_logits
    raise ValueError(f"Target size ({target.size()}) must be the same as input size ({input.size()})")
ValueError: Target size (torch.Size([113, 2])) must be the same as input size (torch.Size([106, 2]))

It seems like the model outputs are sometimes shorter than the targets. I returned to check and there was no error reported during the preprocessing. So do you have any idea on how to solve the problem? Thank you again!