passes gradient check for offest @ DOUBLE precision

peggiezhu · Dec 5, 2018 · 41ee18d · 41ee18d
1 parent 9b00cff
commit 41ee18d
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 8 deletions.
diff --git a/src/dcn_v2_cuda.c b/src/dcn_v2_cuda.c
@@ -142,7 +142,7 @@ void dcn_v2_cuda_backward(THCudaTensor *input, THCudaTensor *weight,
     {
         // Resize plane and fill with ones...
         THCudaTensor_resize2d(state, ones, height_out, width_out);
-        THCudaTensor_fill(state, ones, 1);
+        THCudaTensor_fill(state, ones, 1.0f);
     }
 
     THCudaTensor_resize4d(state, grad_input, batch, channels, height, width);

diff --git a/src/dcn_v2_cuda_double.c b/src/dcn_v2_cuda_double.c
@@ -158,7 +158,7 @@ void dcn_v2_cuda_backward(THCudaDoubleTensor *input, THCudaDoubleTensor *weight,
     {
         // Resize plane and fill with ones...
         THCudaDoubleTensor_resize2d(state, ones, height_out, width_out);
-        THCudaDoubleTensor_fill(state, ones, 1);
+        THCudaDoubleTensor_fill(state, ones, 1.0);
     }
 
     // THCudaDoubleTensor_resize4d(state, grad_input, batch, channels, height, width);
@@ -239,9 +239,9 @@ void dcn_v2_cuda_backward(THCudaDoubleTensor *input, THCudaDoubleTensor *weight,
         // long k__ = height_out * width_out;
         THCudaBlas_Dgemv(state,
                          't',
-                         k_, m_, 1.0f,
+                         k_, m_, 1.0,
                          THCudaDoubleTensor_data(state, grad_output_n), k_,
-                         THCudaDoubleTensor_data(state, ones), 1, 1.0f,
+                         THCudaDoubleTensor_data(state, ones), 1, 1.0,
                          THCudaDoubleTensor_data(state, grad_bias), 1);
     }
 

diff --git a/test.py b/test.py
@@ -68,8 +68,8 @@ def check_gradient_double():
     input.requires_grad = True
 
     offset = torch.randn(N, deformable_groups * 2 * kW * kH, inH, inW, dtype=torch.float64).cuda()
-    offset.data.zero_()
-    offset.data -= 0.5
+    # offset.data.zero_()
+    # offset.data -= 0.00001
     offset.requires_grad = True
 
     mask = torch.rand(N, deformable_groups * 1 * kW * kH, inH, inW, dtype=torch.float64).cuda()
@@ -93,8 +93,8 @@ def check_gradient():
     input.requires_grad = True
 
     offset = torch.randn(N, deformable_groups * 2 * kW * kH, inH, inW).cuda()
-    offset.data.zero_()
-    offset.data -= 0.5
+    # offset.data.zero_()
+    # offset.data -= 0.5
     offset.requires_grad = True
 
     mask = torch.rand(N, deformable_groups * 1 * kW * kH, inH, inW).cuda()