Prevent metric dtype conversion when metric is part of LightningModule

CHANGELOG.md

@@ -103,6 +103,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Fixed evaluation of `R2Score` with near constant target ([#1576](https://github.com/Lightning-AI/metrics/pull/1576))
 
 
+- Fixed dtype conversion when metric is submodule ([#1583](https://github.com/Lightning-AI/metrics/pull/1583))
+
+
 - Fixed corner case for `PearsonCorrCoef` when running in ddp mode but only on single device ([#1587](https://github.com/Lightning-AI/metrics/pull/1587))
 
 

src/torchmetrics/collections.py

@@ -472,3 +472,13 @@ def __repr__(self) -> str:
         if self.postfix:
             repr_str += f"{',' if not self.prefix else ''}\n  postfix={self.postfix}"
         return repr_str + "\n)"
+
+    def set_dtype(self, dst_type: Union[str, torch.dtype]) -> "MetricCollection":
+        """Transfer all metric state to specific dtype. Special version of standard `type` method.
+
+        Arguments:
+            dst_type (type or string): the desired type.
+        """
+        for _, m in self.items(keep_base=True, copy_state=False):
+            m.set_dtype(dst_type)
+        return self

src/torchmetrics/metric.py

@@ -145,6 +145,7 @@ def __init__(
         self._to_sync = self.sync_on_compute
         self._should_unsync = True
         self._enable_grad = False
+        self._dtype_convert = False
 
         # initialize state
         self._defaults: Dict[str, Union[List, Tensor]] = {}
@@ -692,14 +693,23 @@ def set_dtype(self, dst_type: Union[str, torch.dtype]) -> "Metric":
         Arguments:
             dst_type (type or string): the desired type.
         """
-        return super().type(dst_type)
+        self._dtype_convert = True
+        out = super().type(dst_type)
+        out._dtype_convert = False
+        return out
 
     def _apply(self, fn: Callable) -> Module:
         """Overwrite _apply function such that we can also move metric states to the correct device.
 
-        This method is called by the base ``nn.Module`` class whenever `.to`, `.cuda`, etc. methods are called.
+        This method is called by the base ``nn.Module`` class whenever `.to`, `.cuda`, `.float`, `.half` etc. methods
+        are called. Dtype conversion is garded and will only happen through the special `set_dtype` method.
         """
         this = super()._apply(fn)
+        fs = str(fn)
+        cond = any(f in fs for f in ["Module.type", "Module.half", "Module.float", "Module.double", "Module.bfloat16"])
+        if not self._dtype_convert and cond:
+            return this
+
         # Also apply fn to metric states and defaults
         for key, value in this._defaults.items():
             if isinstance(value, Tensor):

tests/integrations/test_lightning.py

@@ -16,6 +16,7 @@
 import torch
 from pytorch_lightning import LightningModule, Trainer
 from torch import tensor
+from torch.nn import Linear
 from torch.utils.data import DataLoader
 
 from integrations.helpers import no_warning_call
@@ -304,3 +305,48 @@ def training_step(self, batch, batch_idx):
     output = model(rand_input)
     script_output = script_model(rand_input)
     assert torch.allclose(output, script_output)
+
+
+def test_dtype_in_pl_module_transfer(tmpdir):
+    """Test that metric states don't change dtype when .half() or .float() is called on the LightningModule."""
+
+    class BoringModel(LightningModule):
+        def __init__(self, metric_dtype=torch.float32):
+            super().__init__()
+            self.layer = Linear(32, 32)
+            self.metric = SumMetric()
+            self.metric.set_dtype(metric_dtype)
+
+        def forward(self, x):
+            return self.layer(x)
+
+        def training_step(self, batch, batch_idx):
+            pred = self.forward(batch)
+            loss = self(batch).sum()
+            self.metric.update(torch.flatten(pred), torch.flatten(batch))
+
+            return {"loss": loss}
+
+        def configure_optimizers(self):
+            return torch.optim.SGD(self.layer.parameters(), lr=0.1)
+
+    model = BoringModel()
+    assert model.metric.value.dtype == torch.float32
+    model = model.half()
+    assert model.metric.value.dtype == torch.float32
+
+    model = BoringModel()
+    assert model.metric.value.dtype == torch.float32
+    model = model.double()
+    assert model.metric.value.dtype == torch.float32
+
+    model = BoringModel(metric_dtype=torch.float16)
+    assert model.metric.value.dtype == torch.float16
+    model = model.float()
+    assert model.metric.value.dtype == torch.float16
+
+    model = BoringModel()
+    assert model.metric.value.dtype == torch.float32
+
+    model = model.type(torch.half)
+    assert model.metric.value.dtype == torch.float32

tests/unittests/bases/test_collections.py

@@ -93,11 +93,11 @@ def test_device_and_dtype_transfer_metriccollection(tmpdir):
     for _, metric in metric_collection.items():
         assert metric.x.is_cuda
 
-    metric_collection = metric_collection.double()
+    metric_collection = metric_collection.set_dtype(torch.double)
     for _, metric in metric_collection.items():
         assert metric.x.dtype == torch.float64
 
-    metric_collection = metric_collection.half()
+    metric_collection = metric_collection.set_dtype(torch.half)
     for _, metric in metric_collection.items():
         assert metric.x.dtype == torch.float16
 

tests/unittests/bases/test_metric.py

@@ -308,6 +308,24 @@ def test_device_and_dtype_transfer(tmpdir):
     assert metric.x.dtype == torch.float16
     metric.reset()
     assert metric.x.dtype == torch.float16
+    import pdb
+
+    pdb.set_trace()
+
+
+def test_disable_of_normal_dtype_methods():
+    """Check that the default dtype changing methods does nothing."""
+    metric = DummyMetricSum()
+    assert metric.x.dtype == torch.float32
+
+    metric = metric.half()
+    assert metric.x.dtype == torch.float32
+
+    metric = metric.double()
+    assert metric.x.dtype == torch.float32
+
+    metric = metric.type(torch.half)
+    assert metric.x.dtype == torch.float32
 
 
 def test_warning_on_compute_before_update():