Fix L1Loss when target.requires_grad is True. (#44471)

gchanan · facebook-github-bot · commit 3de2c0b42f42 · 2020-09-11T09:51:16.000-07:00
Summary: Pull Request resolved: #44471 L1Loss had a completely different (and incorrect, see #43228) path when target.requires_grad was True. This PR does the following: 1) adds derivative support for target via the normal derivatives.yaml route 2) kill the different (and incorrect) path for when target.requires_grad was True 3) modify the L1Loss CriterionTests to verify that the target derivative is checked. Test Plan: Imported from OSS Reviewed By: albanD Differential Revision: D23626008 Pulled By: gchanan fbshipit-source-id: 2828be16b56b8dabe114962223d71b0e9a85f0f5
diff --git a/tools/autograd/derivatives.yaml b/tools/autograd/derivatives.yaml
@@ -1202,6 +1202,7 @@
 
 - name: l1_loss(Tensor self, Tensor target, int reduction=Mean) -> Tensor
   self: l1_loss_backward(grad, self, target, reduction)
+  target: l1_loss_backward(grad, target, self, reduction)
 
 - name: mse_loss(Tensor self, Tensor target, int reduction=Mean) -> Tensor
   self: mse_loss_backward(grad, self, target, reduction)
@@ -1520,6 +1521,7 @@
 - name: l1_loss_backward(Tensor grad_output, Tensor self, Tensor target, int reduction) -> Tensor
   grad_output: l1_loss_double_backward_grad_output(grad, self, target, reduction)
   self: zeros_like(grad, at::MemoryFormat::Preserve)
+  target: zeros_like(grad, at::MemoryFormat::Preserve)
 
 - name: log_sigmoid_backward(Tensor grad_output, Tensor self, Tensor buffer) -> Tensor
   grad_output: log_sigmoid_backward(grad, self, buffer)
diff --git a/torch/nn/functional.py b/torch/nn/functional.py
@@ -2629,15 +2629,10 @@ def l1_loss(input, target, size_average=None, reduce=None, reduction='mean'):
                       stacklevel=2)
     if size_average is not None or reduce is not None:
         reduction = _Reduction.legacy_get_string(size_average, reduce)
-    if target.requires_grad:
-        _Reduction.get_enum(reduction)  # throw an error if reduction is invalid
-        ret = torch.abs(input - target)
-        if reduction != 'none':
-            ret = torch.mean(ret) if reduction == 'mean' else torch.sum(ret)
-    else:
-        expanded_input, expanded_target = torch.broadcast_tensors(input, target)
-        ret = torch._C._nn.l1_loss(expanded_input, expanded_target, _Reduction.get_enum(reduction))
-    return ret
+
+
+    expanded_input, expanded_target = torch.broadcast_tensors(input, target)
+    return torch._C._nn.l1_loss(expanded_input, expanded_target, _Reduction.get_enum(reduction))
 
 
 def mse_loss(input, target, size_average=None, reduce=None, reduction='mean'):
diff --git a/torch/testing/_internal/common_nn.py b/torch/testing/_internal/common_nn.py
@@ -3879,7 +3879,7 @@ def padding3d_circular(input, pad):
     dict(
         module_name='L1Loss',
         input_size=(2, 3, 4),
-        target_size=(2, 3, 4),
+        target_fn=lambda: torch.randn((2, 3, 4), requires_grad=True),
         reference_fn=lambda i, t, _: 1. / i.numel() *
         sum((a - b).abs().sum() for a, b in zip(i, t)),
     ),
@@ -4277,7 +4277,7 @@ def padding3d_circular(input, pad):
     dict(
         module_name='L1Loss',
         input_size=(),
-        target_size=(),
+        target_fn=lambda: torch.randn((), requires_grad=True),
         reference_fn=lambda i, t, _: 1. / i.numel() * (i - t).abs().sum(),
         desc='scalar',
     ),
