SMAPE() Expected all tensors to be on the same device

[jacktang]

• PyTorch-Forecasting version: 1.0.0
• PyTorch version: 1.12.1
• Python version: 3.10.6
• Operating System: Ubuntu 20.4

Expected behavior

The code from N-beats tutorial:

actuals = torch.cat([y[0] for x, y in iter(val_dataloader)])
baseline_predictions = Baseline().predict(val_dataloader)
SMAPE()(baseline_predictions, actuals)

expected to get result without error

Actual behavior

However, errors came up:

GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
File [~/miniconda3/envs/pf/lib/python3.10/site-packages/torchmetrics/metric.py:446](https://vscode-remote+ssh-002dremote-002bd1-002eys.vscode-resource.vscode-cdn.net/home/dev/Codes/caebigdata-jack/torch%2Bjax/forecasting/~/miniconda3/envs/pf/lib/python3.10/site-packages/torchmetrics/metric.py:446), in Metric._wrap_update..wrapped_func(*args, **kwargs)
    445 try:
--> 446     update(*args, **kwargs)
    447 except RuntimeError as err:

File [~/miniconda3/envs/pf/lib/python3.10/site-packages/pytorch_forecasting/metrics/base_metrics.py:784](https://vscode-remote+ssh-002dremote-002bd1-002eys.vscode-resource.vscode-cdn.net/home/dev/Codes/caebigdata-jack/torch%2Bjax/forecasting/~/miniconda3/envs/pf/lib/python3.10/site-packages/pytorch_forecasting/metrics/base_metrics.py:784), in MultiHorizonMetric.update(self, y_pred, target)
    782     lengths = torch.full((target.size(0),), fill_value=target.size(1), dtype=torch.long, device=target.device)
--> 784 losses = self.loss(y_pred, target)
    785 # weight samples

File [~/miniconda3/envs/pf/lib/python3.10/site-packages/pytorch_forecasting/metrics/point.py:69](https://vscode-remote+ssh-002dremote-002bd1-002eys.vscode-resource.vscode-cdn.net/home/dev/Codes/caebigdata-jack/torch%2Bjax/forecasting/~/miniconda3/envs/pf/lib/python3.10/site-packages/pytorch_forecasting/metrics/point.py:69), in SMAPE.loss(self, y_pred, target)
     68 y_pred = self.to_prediction(y_pred)
---> 69 loss = 2 * (y_pred - target).abs() [/](https://vscode-remote+ssh-002dremote-002bd1-002eys.vscode-resource.vscode-cdn.net/) (y_pred.abs() + target.abs() + 1e-8)
     70 return loss

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

The above exception was the direct cause of the following exception:

RuntimeError                              Traceback (most recent call last)
Cell In[10], line 5
      3 actuals = torch.cat([y[0] for x, y in iter(val_dataloader)])
      4 baseline_predictions = Baseline().predict(val_dataloader)
----> 5 SMAPE()(baseline_predictions, actuals)

File [~/miniconda3/envs/pf/lib/python3.10/site-packages/torch/nn/modules/module.py:1501](https://vscode-remote+ssh-002dremote-002bd1-002eys.vscode-resource.vscode-cdn.net/home/dev/Codes/caebigdata-jack/torch%2Bjax/forecasting/~/miniconda3/envs/pf/lib/python3.10/site-packages/torch/nn/modules/module.py:1501), in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File [~/miniconda3/envs/pf/lib/python3.10/site-packages/torchmetrics/metric.py:290](https://vscode-remote+ssh-002dremote-002bd1-002eys.vscode-resource.vscode-cdn.net/home/dev/Codes/caebigdata-jack/torch%2Bjax/forecasting/~/miniconda3/envs/pf/lib/python3.10/site-packages/torchmetrics/metric.py:290), in Metric.forward(self, *args, **kwargs)
    288     self._forward_cache = self._forward_full_state_update(*args, **kwargs)
    289 else:
--> 290     self._forward_cache = self._forward_reduce_state_update(*args, **kwargs)
    292 return self._forward_cache

File [~/miniconda3/envs/pf/lib/python3.10/site-packages/torchmetrics/metric.py:357](https://vscode-remote+ssh-002dremote-002bd1-002eys.vscode-resource.vscode-cdn.net/home/dev/Codes/caebigdata-jack/torch%2Bjax/forecasting/~/miniconda3/envs/pf/lib/python3.10/site-packages/torchmetrics/metric.py:357), in Metric._forward_reduce_state_update(self, *args, **kwargs)
    354 self._enable_grad = True  # allow grads for batch computation
    356 # calculate batch state and compute batch value
--> 357 self.update(*args, **kwargs)
    358 batch_val = self.compute()
    360 # reduce batch and global state

File [~/miniconda3/envs/pf/lib/python3.10/site-packages/torchmetrics/metric.py:449](https://vscode-remote+ssh-002dremote-002bd1-002eys.vscode-resource.vscode-cdn.net/home/dev/Codes/caebigdata-jack/torch%2Bjax/forecasting/~/miniconda3/envs/pf/lib/python3.10/site-packages/torchmetrics/metric.py:449), in Metric._wrap_update..wrapped_func(*args, **kwargs)
    447     except RuntimeError as err:
    448         if "Expected all tensors to be on" in str(err):
--> 449             raise RuntimeError(
    450                 "Encountered different devices in metric calculation (see stacktrace for details)."
    451                 " This could be due to the metric class not being on the same device as input."
    452                 f" Instead of `metric={self.__class__.__name__}(...)` try to do"
    453                 f" `metric={self.__class__.__name__}(...).to(device)` where"
    454                 " device corresponds to the device of the input."
    455             ) from err
    456         raise err
    458 if self.compute_on_cpu:

RuntimeError: Encountered different devices in metric calculation (see stacktrace for details). This could be due to the metric class not being on the same device as input. Instead of `metric=SMAPE(...)` try to do `metric=SMAPE(...).to(device)` where device corresponds to the device of the input.

[jacktang]

The code should be fixed

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
actuals = torch.cat([y[0] for x, y in iter(val_dataloader)]).to(device)