[GitHub] szha closed pull request #11231: Fix a bug in sparse embedding operator

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

szha closed pull request #11231: Fix a bug in sparse embedding operator
URL: https://github.com/apache/incubator-mxnet/pull/11231
 
 
   

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.cu b/src/operator/tensor/indexing_op.cu
index 593593ae4e3..b0ee05ea139 100644
--- a/src/operator/tensor/indexing_op.cu
+++ b/src/operator/tensor/indexing_op.cu
@@ -188,7 +188,8 @@ void SparseEmbeddingDeterministicKernelLaunch(const OpContext& ctx,
   // estimate unique temp space
   IType* data_ptr = data.dptr<IType>();
   size_t *null_ptr = nullptr;
-  cub::DeviceSelect::Unique(NULL, unique_workspace_bytes, data_ptr, data_ptr,
+  // unique operations will be applied on sorted data
+  cub::DeviceSelect::Unique(NULL, unique_workspace_bytes, sorted_data, sorted_data,
     null_ptr, data_size, Stream<gpu>::GetStream(s));
   // One more space reserved for unique count
   size_t temp_workspace_bytes = std::max(unique_workspace_bytes,
diff --git a/tests/python/unittest/test_gluon.py b/tests/python/unittest/test_gluon.py
index bf1e0deb200..ced3063448b 100644
--- a/tests/python/unittest/test_gluon.py
+++ b/tests/python/unittest/test_gluon.py
@@ -753,8 +753,22 @@ def check_embedding(sparse_grad):
             y.backward()
         assert (layer.weight.grad().asnumpy()[:5] == 1).all()
         assert (layer.weight.grad().asnumpy()[5:] == 0).all()
+
+    def check_embedding_large_input(sparse_grad):
+        embedding = mx.gluon.nn.Embedding(10, 1, sparse_grad=True)
+        embedding.initialize()
+        embedding.hybridize()
+        shape = (20481,)
+        with mx.autograd.record():
+            emb_in = embedding(mx.nd.ones(shape))
+            loss = emb_in.sum()
+        loss.backward()
+        assert embedding.weight.grad().data.sum().asscalar() == 20481
+
     check_embedding(True)
     check_embedding(False)
+    check_embedding_large_input(True)
+    check_embedding_large_input(False)
 
 @with_seed()
 def test_export():


 

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