initial commit

Author: kexinzhao

paddle/fluid/operators/dropout_op.cu

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <thrust/random.h>
 #include <thrust/transform.h>
 #include "paddle/fluid/operators/dropout_op.h"
+#include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace operators {
@@ -51,7 +52,7 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
     auto* x = context.Input<Tensor>("X");
     auto* y = context.Output<Tensor>("Out");
     y->mutable_data<T>(context.GetPlace());
-    AttrType dropout_prob = context.Attr<AttrType>("dropout_prob");
+    AttrType dropout_prob = context.Attr<AttrType>("dropout_prob"));
 
     auto X = EigenMatrix<T>::Reshape(*x, 1);
     auto Y = EigenMatrix<T>::Reshape(*y, 1);
@@ -74,7 +75,7 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
                                      context.cuda_device_context().stream()>>>(
           size, seed, dropout_prob, x_data, mask_data, y_data);
     } else {
-      Y.device(place) = X * (1.0f - dropout_prob);
+      Y.device(place) = X * static_cast<T>(1.0f - dropout_prob);
     }
   }
 };
@@ -83,9 +84,9 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
+namespace plat = paddle::platform;
 REGISTER_OP_CUDA_KERNEL(
-    dropout,
-    ops::GPUDropoutKernel<paddle::platform::CUDADeviceContext, float, float>);
-REGISTER_OP_CUDA_KERNEL(
-    dropout_grad,
-    ops::DropoutGradKernel<paddle::platform::CUDADeviceContext, float>);
+    dropout, ops::GPUDropoutKernel<plat::CUDADeviceContext, float, float>,
+    ops::GPUDropoutKernel<plat::CUDADeviceContext, plat::float16, float>);
+REGISTER_OP_CUDA_KERNEL(dropout_grad,
+                        ops::DropoutGradKernel<plat::CUDADeviceContext, float>);

python/paddle/fluid/tests/unittests/test_dropout_op.py

@@ -82,5 +82,31 @@ def test_check_output(self):
         self.check_output()
 
 
+class TestFP16DropoutOp1(OpTest):
+    def setUp(self):
+        x = np.random.random((32, 64)).astype("float16")
+        self.op_type = "dropout"
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {'dropout_prob': 0.35, 'fix_seed': True, 'is_test': True}
+        self.outputs = {'Out': x * (1.0 - self.attrs['dropout_prob'])}
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda() and core.op_support_gpu("dropout"):
+            self.check_output_with_place(core.CUDAPlace(0), atol=1e-3)
+
+
+class TestFP16DropoutOp2(OpTest):
+    def setUp(self):
+        x = np.random.random((32, 64, 3)).astype("float16")
+        self.op_type = "dropout"
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.attrs = {'dropout_prob': 0.75, 'is_test': True}
+        self.outputs = {'Out': x * (1.0 - self.attrs['dropout_prob'])}
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda() and core.op_support_gpu("dropout"):
+            self.check_output_with_place(core.CUDAPlace(0), atol=1e-3)
+
+
 if __name__ == '__main__':
     unittest.main()