Metrics computation on pytorch MPS backend fails, needs float32 instead of float64

[blackplane]

🐛 Bug description

Running metrics via evaluator.run(dataloader) on MacOS fails, because the pytorch MPS backend doesn't support the float64 type that the result is cast into. source code link

Suggestion: Cast to float32 instead. Although it is correct that int64 cannot cast to float32 without a loss of precision and float64 is technically the correct choice, pragmatically I'd think that this loss won't matter in practice as y_pred represents class label and 2^32 are already more than anyone will need.

File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/engine/engine.py", line 889, in run
return self._internal_run()
^^^^^^^^^^^^^^^^^^^^
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/engine/engine.py", line 932, in _internal_run
return next(self._internal_run_generator)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/engine/engine.py", line 990, in _internal_run_as_gen
self._handle_exception(e)
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/engine/engine.py", line 644, in _handle_exception
raise e
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/engine/engine.py", line 956, in _internal_run_as_gen
epoch_time_taken += yield from self._run_once_on_dataset_as_gen()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/engine/engine.py", line 1096, in _run_once_on_dataset_as_gen
self._handle_exception(e)
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/engine/engine.py", line 644, in _handle_exception
raise e
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/engine/engine.py", line 1078, in _run_once_on_dataset_as_gen
self._fire_event(Events.ITERATION_COMPLETED)
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/engine/engine.py", line 431, in _fire_event
func(*first, *(event_args + others), **kwargs)
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/metrics/metric.py", line 469, in iteration_completed
self.update(output)
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/metrics/metric.py", line 864, in wrapper
func(self, *args, **kwargs)
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/metrics/recall.py", line 227, in update
_, y, correct = self._prepare_output(output)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/user/Library/Caches/pypoetry/virtualenvs/columbo-HtoL9iYn-py3.12/lib/python3.12/site-packages/ignite/metrics/precision.py", line 84, in _prepare_output
y_pred = y_pred.to(dtype=torch.float64, device=self._device)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.

Environment

• PyTorch Version (e.g., 1.4): 2.6.0
• Ignite Version (e.g., 0.3.0): 0.5.1
• OS (e.g., Linux): MacOS
• How you installed Ignite (conda, pip, source): pip
• Python version: 3.12.8
• Any other relevant information:

[vfdev-5]

@blackplane thanks for the report!

A question, do you explicitly specify mps device in the metric definition?

precision = Precision(device="mps")

It is possible to add a route and use f32 for MPS device in the code where we set f64 dtype for computations and probably increase the tolerance in the tests.