Can I finetune the model with Pytorch for "horizon_len" lower than 128?

[sebassaras02]

I've been trying to finetune the model with my custom dataset. My dataset has many time-series, I wrangled the data to have a context_len of 32 values (train with the previous 32) and horizon_len of 4 (predict the next 4).

I am getting this error:
"The size of tensor a (128) must match the size of tensor b (4) at non-singleton dimension 1"

I understand this is a mismatch in the loss function dimensions, but I didn't figure it out how to fine-tune for a lower horizon_len.

This is all the error obtained.

{
	"name": "RuntimeError",
	"message": "The size of tensor a (128) must match the size of tensor b (4) at non-singleton dimension 1",
	"stack": "---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[14], line 2
      1 print(\"\
Starting finetuning...\")
----> 2 results = finetuner.finetune(train_dataset=train_dataset,
      3                             val_dataset=val_dataset)
      4 print(\"\
Finetuning completed!\")
 
File ~/anaconda3/envs/timefm/lib/python3.11/site-packages/finetuning/finetuning_torch.py:367, in TimesFMFinetuner.finetune(self, train_dataset, val_dataset)
    365 try:
    366   for epoch in range(self.config.num_epochs):
--> 367     train_loss = self._train_epoch(train_loader, optimizer)
    368     val_loss = self._validate(val_loader)
    369     current_lr = optimizer.param_groups[0][\"lr\"]
 
File ~/anaconda3/envs/timefm/lib/python3.11/site-packages/finetuning/finetuning_torch.py:295, in TimesFMFinetuner._train_epoch(self, train_loader, optimizer)
    292 num_batches = len(train_loader)
    294 for batch in train_loader:
--> 295   loss, _ = self._process_batch(batch)
    297   optimizer.zero_grad()
    298   loss.backward()
 
File ~/anaconda3/envs/timefm/lib/python3.11/site-packages/finetuning/finetuning_torch.py:268, in TimesFMFinetuner._process_batch(self, batch)
    265 predictions_mean = predictions[..., 0]
    266 last_patch_pred = predictions_mean[:, -1, :]
--> 268 loss = self.loss_fn(last_patch_pred, x_future.squeeze(-1))
    269 if self.config.use_quantile_loss:
    270   quantiles = self.config.quantiles or create_quantiles()
 
File ~/anaconda3/envs/timefm/lib/python3.11/site-packages/finetuning/finetuning_torch.py:187, in TimesFMFinetuner.__init__.<locals>.<lambda>(x, y)
    184 self.logger = logger or logging.getLogger(__name__)
    185 self.device = torch.device(
    186     f\"cuda:{rank}\" if torch.cuda.is_available() else \"cpu\")
--> 187 self.loss_fn = loss_fn or (lambda x, y: torch.mean((x - y.squeeze(-1))**2))
    189 if config.use_wandb:
    190   self.metrics_logger = WandBLogger(config.wandb_project, config.__dict__,
    191                                     rank)
 
RuntimeError: The size of tensor a (128) must match the size of tensor b (4) at non-singleton dimension 1"
}