[GitHub] eric-haibin-lin closed pull request #13218: Fix a bug in index_copy

[GitBox <gi...@apache.org>]

eric-haibin-lin closed pull request #13218: Fix a bug in index_copy
URL: https://github.com/apache/incubator-mxnet/pull/13218
 
 
   

This is a PR merged from a forked repository.
As GitHub hides the original diff on merge, it is displayed below for
the sake of provenance:

As this is a foreign pull request (from a fork), the diff is supplied
below (as it won't show otherwise due to GitHub magic):

diff --git a/src/operator/contrib/index_copy-inl.h b/src/operator/contrib/index_copy-inl.h
index b97138a88f9..923fb0f4f13 100644
--- a/src/operator/contrib/index_copy-inl.h
+++ b/src/operator/contrib/index_copy-inl.h
@@ -32,6 +32,7 @@
 #include "../elemwise_op_common.h"
 #include "../mshadow_op.h"
 #include "../mxnet_op.h"
+#include "../tensor/init_op.h"
 
 namespace mxnet {
 namespace op {
@@ -83,12 +84,12 @@ void IndexCopyForward(const nnvm::NodeAttrs& attrs,
   });
 }
 
-template<int req>
 struct index_copy_backward {
   template<typename DType, typename IType>
   MSHADOW_XINLINE static void Map(int i,
                                   int dim,
                                   int index_size,
+                                  int req1, int req2,
                                   DType* out_grad,
                                   IType* index,
                                   DType* in_grad_1,
@@ -98,12 +99,12 @@ struct index_copy_backward {
       int idx = static_cast<int>(index[p]);
       if (i >= idx*dim && i < (idx+1)*dim) {
         int offset = i - idx*dim;
-        KERNEL_ASSIGN(in_grad_2[p*dim+offset], req, out_grad[i]);
+        KERNEL_ASSIGN(in_grad_2[p*dim+offset], req2, out_grad[i]);
         return;
       }
     }
     // Copy to in_grad_1
-    KERNEL_ASSIGN(in_grad_1[i], req, out_grad[i]);
+    KERNEL_ASSIGN(in_grad_1[i], req1, out_grad[i]);
   }
 };
 
@@ -122,18 +123,19 @@ void IndexCopyBackward(const nnvm::NodeAttrs& attrs,
   const TBlob& in_grad_2 = outputs[2];
   int dim = inputs[3].Size() / inputs[2].Size();
   int index_size = inputs[2].Size();
+  Fill<false>(s, outputs[0], req[0], 0);
+  Fill<false>(s, outputs[2], req[2], 0);
   // index_copy_backward
   MSHADOW_TYPE_SWITCH(out_grad.type_flag_, DType, {
     MSHADOW_TYPE_SWITCH(index.type_flag_, IType, {
-      MXNET_ASSIGN_REQ_SWITCH(req[0], req_type, {
-        mxnet_op::Kernel<index_copy_backward<req_type>, xpu>::Launch(s,
+      mxnet_op::Kernel<index_copy_backward, xpu>::Launch(s,
                                       out_grad.Size(),
                                       dim, index_size,
+                                      req[0], req[2],
                                       out_grad.dptr<DType>(),
                                       index.dptr<IType>(),
                                       in_grad_1.dptr<DType>(),
                                       in_grad_2.dptr<DType>());
-      });
     });
   });
 }
diff --git a/tests/python/unittest/test_operator.py b/tests/python/unittest/test_operator.py
index 80a83df45da..c59cdfb5389 100644
--- a/tests/python/unittest/test_operator.py
+++ b/tests/python/unittest/test_operator.py
@@ -4768,24 +4768,25 @@ def test_index_copy():
     x = mx.nd.zeros((5,3))
     t = mx.nd.array([[1,2,3],[4,5,6],[7,8,9]])
     index = mx.nd.array([0,4,2], dtype=np.int64)
+    tensor = mx.nd.array([[1,2,3],[0,0,0],[7,8,9],[0,0,0],[4,5,6]])
+    x_grad = mx.nd.array([[0,0,0],[1,1,1],[0,0,0],[1,1,1],[0,0,0]])
+    t_grad = mx.nd.array([[1,1,1],[1,1,1],[1,1,1]])
 
-    x.attach_grad()
     t.attach_grad()
-    index.attach_grad()
-
     with mx.autograd.record():
         out = mx.nd.contrib.index_copy(x, index, t)
     out.backward()
+    assert same(out.asnumpy(), tensor.asnumpy())
+    assert same(t.grad.asnumpy(), t_grad.asnumpy())
 
-    tensor = mx.nd.array([[1,2,3],[0,0,0],[7,8,9],[0,0,0],[4,5,6]])
-    x_grad = mx.nd.array([[0,0,0],[1,1,1],[0,0,0],[1,1,1],[0,0,0]])
-    t_grad = mx.nd.array([[1,1,1],[1,1,1],[1,1,1]])
-    index_grad = mx.nd.array([0,0,0])
-
+    x.attach_grad()
+    t.attach_grad()
+    with mx.autograd.record():
+        out = mx.nd.contrib.index_copy(x, index, t)
+    out.backward()
     assert same(out.asnumpy(), tensor.asnumpy())
     assert same(x.grad.asnumpy(), x_grad.asnumpy())
     assert same(t.grad.asnumpy(), t_grad.asnumpy())
-    assert same(index.grad.asnumpy(), index_grad.asnumpy())
 
 @with_seed()
 def test_div_sqrt_dim():


 

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services