[GitHub] szha closed pull request #10918: fix a bug in cudnn softmax activation.

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

szha closed pull request #10918: fix a bug in cudnn softmax activation.
URL: https://github.com/apache/incubator-mxnet/pull/10918
 
 
   

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/nn/cudnn/cudnn_softmax_activation-inl.h b/src/operator/nn/cudnn/cudnn_softmax_activation-inl.h
index 239da023668..0845eb79fd6 100644
--- a/src/operator/nn/cudnn/cudnn_softmax_activation-inl.h
+++ b/src/operator/nn/cudnn/cudnn_softmax_activation-inl.h
@@ -48,7 +48,7 @@ class CuDNNSoftmaxActivationOp {
   }
 
   void Forward(const OpContext &ctx, const TBlob &in_data,
-      const OpReqType &req, const TBlob &out_data) {
+               const OpReqType &req, const TBlob &out_data) {
     using namespace mshadow;
     using namespace mshadow::expr;
     Stream<gpu> *s = ctx.get_stream<gpu>();
@@ -102,14 +102,14 @@ class CuDNNSoftmaxActivationOp {
   }
 
   void Backward(const OpContext &ctx, const TBlob &out_grad,
-      const TBlob &out_data, const OpReqType &req, const TBlob &in_grad) {
+                const TBlob &out_data, const OpReqType &req,
+                const TBlob &in_grad) {
     using namespace mshadow;
     using namespace mshadow::expr;
     float alpha = 1.0f;
     float beta = 0.0f;
     Stream<gpu> *s = ctx.get_stream<gpu>();
     Tensor<gpu, 4> grad;
-    Tensor<gpu, 4> data;
     Tensor<gpu, 4> output_data;
     Tensor<gpu, 4> input_grad;
     cudnnSoftmaxMode_t softmax_mode;
@@ -141,6 +141,13 @@ class CuDNNSoftmaxActivationOp {
       softmax_mode = CUDNN_SOFTMAX_MODE_CHANNEL;
     }
     CHECK_EQ(s->dnn_handle_ownership_, mshadow::Stream<gpu>::OwnHandle);
+    CUDNN_CALL(cudnnSetTensor4dDescriptor(shape_desc_,
+                                          CUDNN_TENSOR_NCHW,
+                                          dtype_,
+                                          input_grad.shape_[0],
+                                          input_grad.shape_[1],
+                                          input_grad.shape_[2],
+                                          input_grad.shape_[3]));
     CUDNN_CALL(cudnnSoftmaxBackward(s->dnn_handle_,
                                     CUDNN_SOFTMAX_ACCURATE,
                                     softmax_mode,
diff --git a/tests/python/gpu/test_operator_gpu.py b/tests/python/gpu/test_operator_gpu.py
index c7df2eae9b2..43c62e18845 100644
--- a/tests/python/gpu/test_operator_gpu.py
+++ b/tests/python/gpu/test_operator_gpu.py
@@ -1855,6 +1855,26 @@ def test_create_sparse_ndarray_gpu_to_cpu():
         assert(same(rsp_copy.asnumpy(), rsp_created.asnumpy()))
 
 
+@with_seed()
+def test_softmax_activation():
+    gpu_a = mx.nd.array([[3., 0.5, -0.5, 2., 7.],
+        [2., -.4, 7.,   3., 0.2]], ctx=mx.gpu(0))
+    cpu_a = mx.nd.array([[3., 0.5, -0.5, 2., 7.],
+        [2., -.4, 7.,   3., 0.2]], ctx=mx.cpu())
+
+    cpu_a.attach_grad()
+    gpu_a.attach_grad()
+    with mx.autograd.record():
+        gpu_y = mx.nd.SoftmaxActivation(data = gpu_a)
+        cpu_y = mx.nd.SoftmaxActivation(data = cpu_a)
+        assert_almost_equal(cpu_y.asnumpy(), gpu_y.asnumpy(), atol = 1e-3, rtol = 1e-3)
+
+        gpu_y.backward()
+        cpu_y.backward()
+        assert_almost_equal(cpu_a.grad.asnumpy(), gpu_a.grad.asnumpy(),
+                atol = 1e-3, rtol = 1e-3)
+
+
 if __name__ == '__main__':
     import nose
     nose.runmodule()


 

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