[GitHub] eric-haibin-lin closed pull request #11983: Add handling for grad req type other than kNullOp for indices

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

eric-haibin-lin closed pull request #11983: Add handling for grad req type other than kNullOp for indices
URL: https://github.com/apache/incubator-mxnet/pull/11983
 
 
   

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/tensor/indexing_op.h b/src/operator/tensor/indexing_op.h
index edaf9397303..1daf0a2cb18 100644
--- a/src/operator/tensor/indexing_op.h
+++ b/src/operator/tensor/indexing_op.h
@@ -1034,8 +1034,8 @@ void TakeOpBackward(const nnvm::NodeAttrs& attrs,
   using namespace mshadow::expr;
   CHECK_EQ(inputs.size(), 2U);
   CHECK_EQ(outputs.size(), 2U);
-  CHECK_EQ(req[take_::kIdx], kNullOp)
-    << "take layer doesn't support gradient into index";
+  CHECK_NE(req[take_::kIdx], kAddTo)
+    << "take layer doesn't support gradient of req type kAddTo to index";
 
   const TakeParam& param = nnvm::get<TakeParam>(attrs.parsed);
 
@@ -1052,6 +1052,11 @@ void TakeOpBackward(const nnvm::NodeAttrs& attrs,
       const TShape& arrshape = outputs[0].shape_;
       const TShape& oshape = inputs[0].shape_;
 
+      if (req[take_::kIdx] != kNullOp) {
+        mxnet_op::Kernel<mxnet_op::set_zero, xpu>::Launch(
+          s, idxshape.Size(), outputs[take_::kIdx].dptr<IType>());
+      }
+
       const int actual_axis = param.axis + ((param.axis < 0) ? arrshape.ndim() : 0);
 
       int idxndim = idxshape.ndim();
diff --git a/tests/python/unittest/test_operator.py b/tests/python/unittest/test_operator.py
index c395199e8ea..aeda36d8b48 100644
--- a/tests/python/unittest/test_operator.py
+++ b/tests/python/unittest/test_operator.py
@@ -3810,6 +3810,31 @@ def check_output_n_grad(data_shape, idx_shape, axis, mode):
         exe.backward([mx.nd.array(grad_out)])
         assert_almost_equal(exe.grad_dict['a'].asnumpy(), grad_in)
 
+    def check_autograd_req():
+        row_len = 2
+        col_len = 8
+        shape = (row_len, col_len)
+        sc = mx.nd.random.uniform(-1.0, 1.0, shape=shape, dtype="float32")
+        sc.attach_grad()
+        i = mx.nd.array([0], dtype="int64")
+        j = mx.nd.array([0], dtype="int64")
+        with mx.autograd.record(train_mode=True):
+            xs = []
+            for _ in range(row_len):
+                x_i = []
+                for _ in range(col_len):
+                    x_ij = sc.take(i).squeeze(axis=0).take(j).squeeze(axis=0)
+                    x_i.append(x_ij)
+                    j = j + 1
+                i = i + 1
+                j = j - col_len  # reset j
+                xs.append(mx.nd.stack(*x_i))
+            x = mx.nd.stack(*xs)
+            x = x.sum()
+
+        x.backward()
+        assert_almost_equal(np.ones(sc.grad.shape), sc.grad.asnumpy())
+
     for mode in ['clip', 'wrap']:
         for data_ndim in range(1, 5):
             for idx_ndim in range(1, 4):
@@ -3822,6 +3847,8 @@ def check_output_n_grad(data_shape, idx_shape, axis, mode):
                         idx_shape += (np.random.randint(low=1, high=5), )
                     check_output_n_grad(data_shape, idx_shape, axis, mode)
 
+    check_autograd_req()
+
 
 @with_seed()
 def test_grid_generator():


 

----------------------------------------------------------------
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