PearsonCorrcoef breaks when a single sample is passed at a time

[raimis]

🐛 Bug

When passing a single sample to the PearsonCorrcoef metric, it will crash. Similar to #227 and partially fixed by #229. Originally reported in #227 (comment)

To Reproduce

import torch
import torchmetrics

print(torch.__version__)
print(torchmetrics.__version__)

correlation = torchmetrics.PearsonCorrcoef()

correlation.update(torch.tensor([0.3]), torch.tensor([0.4]))
correlation.update(torch.tensor([0.7]), torch.tensor([0.5]))
correlation.update(torch.tensor([0.9]), torch.tensor([0.4]))
correlation.compute()

Output:

1.8.0
0.3.2
/shared/raimis/opt/miniconda/envs/tmp/lib/python3.8/site-packages/torchmetrics/utilities/prints.py:36: UserWarning: Metric `PearsonCorrcoef` will save all targets and predictions in buffer. For large datasets this may lead to large memory footprint.
  warnings.warn(*args, **kwargs)
Traceback (most recent call last):
  File "bug.py", line 12, in <module>
    correlation.compute()
  File "/shared/raimis/opt/miniconda/envs/tmp/lib/python3.8/site-packages/torchmetrics/metric.py", line 251, in wrapped_func
    self._computed = compute(*args, **kwargs)
  File "/shared/raimis/opt/miniconda/envs/tmp/lib/python3.8/site-packages/torchmetrics/regression/pearson.py", line 95, in compute
    preds = torch.cat(self.preds, dim=0)
RuntimeError: zero-dimensional tensor (at position 0) cannot be concatenated

Expected behavior

The metric is computed without throwing the RuntimeError.

[github-actions]

Hi! thanks for your contribution!, great first issue!

[edgarriba]

@raimis please try to update torchmetrics>0.3.2 - in master this issue is fixed.

[raimis]

Thanks @edgarriba. Do you know when the next version will be released?

[edgarriba]

I guess at some point soon since last release was a month ago. Any insights here @Borda @edenlightning ?

[Borda]

Do you know when the next version will be released?

do you need anything specific from master?

[raimis]

do you need anything specific from master?

Just the fix of this issue.

[Borda]

@raimis it was released in 0.3.2, see https://github.com/PyTorchLightning/metrics/releases/tag/v0.3.2

[raimis]

I tried again:

• Install torchmetrics with conda:

  $ conda create -n tmp_1 -c conda-forge torchmetrics pytorch-gpu

• The version of torchmetrics is 0.3.2, as expected:

  $ conda activate tmp_1
  $ conda list | grep torchmetrics
  torchmetrics              0.3.2              pyhd8ed1ab_0    conda-forge

• Run the code above and get the same error:

  $ python
  Python 3.9.4 | packaged by conda-forge | (default, May 10 2021, 22:13:33) 
  [GCC 9.3.0] on linux
  Type "help", "copyright", "credits" or "license" for more information.
  >>> import torch
  >>> import torchmetrics
  >>> print(torch.__version__)
  1.8.0
  >>> print(torchmetrics.__version__)
  0.3.2
  >>> correlation = torchmetrics.PearsonCorrcoef()
  /shared/raimis/opt/miniconda/envs/tmp_1/lib/python3.9/site-packages/torchmetrics/utilities/prints.py:36: UserWarning: Metric `PearsonCorrcoef` will save all targets and predictions in buffer. For large datasets this may lead to large memory footprint.
  warnings.warn(*args, **kwargs)
  >>> correlation.update(torch.tensor([0.3]), torch.tensor([0.4]))
  >>> correlation.update(torch.tensor([0.7]), torch.tensor([0.5]))
  >>> correlation.update(torch.tensor([0.9]), torch.tensor([0.4]))
  >>> correlation.compute()
  Traceback (most recent call last):
    File "<stdin>", line 1, in <module>
    File "/shared/raimis/opt/miniconda/envs/tmp_1/lib/python3.9/site-packages/torchmetrics/metric.py", line 251, in wrapped_func
      self._computed = compute(*args, **kwargs)
    File "/shared/raimis/opt/miniconda/envs/tmp_1/lib/python3.9/site-packages/torchmetrics/regression/pearson.py", line 95, in compute
      preds = torch.cat(self.preds, dim=0)
  RuntimeError: zero-dimensional tensor (at position 0) cannot be concatenated
  >>> 

So, the issues is not fixed in 0.3.2.

[SkafteNicki]

Fix included in v0.4.0 just released (https://github.com/PyTorchLightning/metrics/releases/tag/v0.4.0). Closing.

[raimis]

Thanks!