Small 0.4.0 fixes

docs/source/quickstart.rst

@@ -98,9 +98,10 @@ And update functions of the trainer and evaluator are simply:
 
     def _inference(engine, batch):
         model.eval()
-        x, y = _prepare_batch(batch, device=device)
-        y_pred = model(x)
-        return y_pred, y
+        with torch.no_grad():
+            x, y = _prepare_batch(batch, device=device)
+            y_pred = model(x)
+            return y_pred, y
 
 Note that the helper function :meth:`create_supervised_evaluator` to create an evaluator accepts an
 argument `metrics`:

ignite/metrics/loss.py

@@ -1,7 +1,5 @@
 from __future__ import division
 
-from torch.autograd import Variable
-
 from ignite.exceptions import NotComputableError
 from ignite.metrics.metric import Metric
 

ignite/metrics/mean_absolute_error.py

@@ -19,7 +19,7 @@ def reset(self):
     def update(self, output):
         y_pred, y = output
         absolute_errors = torch.abs(y_pred - y.view_as(y_pred))
-        self._sum_of_absolute_errors += torch.sum(absolute_errors)
+        self._sum_of_absolute_errors += torch.sum(absolute_errors).item()
         self._num_examples += y.shape[0]
 
     def compute(self):

tests/ignite/metrics/test_mean_absolute_error.py

@@ -16,10 +16,12 @@ def test_compute():
     y_pred = torch.Tensor([[2.0], [-2.0]])
     y = torch.zeros(2)
     mae.update((y_pred, y))
+    assert isinstance(mae.compute(), float)
     assert mae.compute() == 2.0
 
     mae.reset()
     y_pred = torch.Tensor([[3.0], [-3.0]])
     y = torch.zeros(2)
     mae.update((y_pred, y))
+    assert isinstance(mae.compute(), float)
     assert mae.compute() == 3.0

tests/ignite/test__utils.py

@@ -1,6 +1,5 @@
 import pytest
 import torch
-from torch.autograd import Variable
 from ignite._utils import convert_tensor, to_onehot