Skip to content

type_as will not move Tensor to the same device #554

Closed
Closed
type_as will not move Tensor to the same device#554
Labels
bug
@leomao

Description

@leomao
leomao
opened on Jul 15, 2019

In some metrics, we use type_as to change the data type of tensors. But type_as does not move the tensor to the same device.

ignite/ignite/metrics/confusion_matrix.py

Lines 73 to 90 in c92838c

def update(self, output):
y_pred, y = self._check_shape(output)
if y_pred.shape != y.shape:
y_ohe = to_onehot(y.reshape(-1), self.num_classes)
y_ohe_t = y_ohe.transpose(0, 1).float()
else:
y_ohe_t = y.transpose(0, 1).reshape(y.shape[1], -1).float()
indices = torch.argmax(y_pred, dim=1)
y_pred_ohe = to_onehot(indices.reshape(-1), self.num_classes)
y_pred_ohe = y_pred_ohe.float()
if self.confusion_matrix.type() != y_ohe_t.type():
self.confusion_matrix = self.confusion_matrix.type_as(y_ohe_t)
self.confusion_matrix += torch.matmul(y_ohe_t, y_pred_ohe).float()
self._num_examples += y_pred.shape[0]

In the code above, if y is on a non-default cuda device, then confusion_matrix will be move to the default cuda device at line 87. This makes line 89 fail because two tensors are not on the same device. I think replacing type_as with to can resolve this issue.

BTW, the .float() in line 89 seems unnecessary.

Metadata

Metadata

Assignees

No one assigned

    Labels

    bug

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions