Added more regression test for groupconv w/o bias. (#18519)
Summary:
Follow-up of https://github.com/pytorch/pytorch/issues/18218, which was fixed by https://github.com/pytorch/pytorch/pull/18463 with mkl-dnn upgraded to v0.18.1.
Covering special case when group > 1, input-channel / group < 16 and output-channel is multiple of 16.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18519
Differential Revision: D14643071
Pulled By: soumith
fbshipit-source-id: d0ebed59326c67089e042b50583b87ed2c3ccc2f
diff --git a/test/test_nn.py b/test/test_nn.py
index bece298..e0b5c5b 100644
--- a/test/test_nn.py
+++ b/test/test_nn.py
@@ -4774,6 +4774,41 @@
torch.cat([m1.weight.grad.data, m2.weight.grad.data], 0),
1e-1 if dtype == torch.half else dtype2prec[dtype])
+ # Almost identical to the above `test_Conv2d_naive_groups`
+ # Covering special case when group > 1, input-channel / group < 16 and output-channel is multiple of 16
+ # See also https://github.com/pytorch/pytorch/pull/18463#issuecomment-476563686
+ # and https://github.com/pytorch/pytorch/pull/18463#issuecomment-477001024
+ def test_Conv2d_groups_nobias_v2(self):
+ dev_dtypes = [("cpu", torch.float)]
+ if TEST_CUDA:
+ dev_dtypes += [("cuda", torch.float), ("cuda", torch.half)]
+ for device, dtype in dev_dtypes:
+ m = nn.Conv2d(4, 16, kernel_size=3, groups=2, bias=False).to(device, dtype)
+ i = torch.randn(2, 4, 6, 6, device=device, dtype=dtype, requires_grad=True)
+ output = m(i)
+ grad_output = torch.randn(2, 16, 4, 4, device=device, dtype=dtype)
+ output.backward(grad_output)
+
+ m1 = nn.Conv2d(2, 8, kernel_size=3, bias=False).to(device, dtype)
+ m1.weight.data.copy_(m.weight.data[:8])
+ i1 = Variable(i.data[:, :2].contiguous(), requires_grad=True)
+ output1 = m1(i1)
+ output1.backward(grad_output[:, :8].contiguous())
+
+ m2 = nn.Conv2d(2, 8, kernel_size=3, bias=False).to(device, dtype)
+ m2.weight.data.copy_(m.weight.data[8:])
+ i2 = Variable(i.data[:, 2:].contiguous(), requires_grad=True)
+ output2 = m2(i2)
+ output2.backward(grad_output[:, 8:].contiguous())
+
+ self.assertEqual(output, torch.cat([output1, output2], 1))
+ self.assertEqual(i.grad.data,
+ torch.cat([i1.grad.data, i2.grad.data], 1),
+ dtype2prec[dtype])
+ self.assertEqual(m.weight.grad.data,
+ torch.cat([m1.weight.grad.data, m2.weight.grad.data], 0),
+ 1e-1 if dtype == torch.half else dtype2prec[dtype])
+
# Very similar to test_Conv2d_naive_groups but with special care to handle
# the number of groups == number of input channels
@unittest.skipIf(not TEST_CUDA, 'CUDA not available')
