add more tests

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -282,7 +282,8 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
     platform::CUDAPlace gpu = boost::get<platform::CUDAPlace>(ctx.GetPlace());
     cudnn_workspace = paddle::memory::Alloc(gpu, workspace_size_in_bytes);
     // ------------------- cudnn conv backward data ---------------------
-    T alpha = 1.0f, beta = 0.0f;
+    typename platform::CudnnDataType<T>::ScalingParamType alpha = 1.0f,
+                                                          beta = 0.0f;
     if (input_grad) {
       T* input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
       // Because beta is zero, it is unnecessary to reset input_grad.

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

@@ -470,29 +470,26 @@ def _numpy_to_lod_tensor(np_value, lod, place):
         return tensor
 
     @staticmethod
-    def create_view(input):
-        """Create a view of the input numpy array
+    def np_dtype_to_fluid_dtype(input):
+        """Change the dtype of float16 numpy array
 
         numpy float16 is binded to paddle::platform::float16 
-        in tensor_py.h via the help of numpy uint16 because
+        in tensor_py.h via the help of uint16 data type since
         the internal memory representation of float16 is 
-        uint16_t in paddle or np.uint16 in numpy, which are
-        themselves binded together.
+        uint16_t in paddle and np.uint16 in numpy, which are
+        themselves binded together by pybind.
 
         Args:
             input: input numpy array
 
         Returns:
-            input_view: if the dtype of input is np.float16, input_view 
-                will reinterpret input as with dtype np.uint16. 
-                Otherwise, input_view will be input itself.
+            input: if the dtype of input is np.float16, its dtype will be
+                changed to np.uint16 so that the internal memory will be 
+                reinterpreted input as of dtype np.uint16. 
         """
         if input.dtype == np.float16:
-            # view will only reinterpret memory without copying
-            input_view = input.view(np.uint16)
-        else:
-            input_view = input
-        return input_view
+            input.dtype = np.uint16
+        return input
 
     def _get_gradient(self, input_to_check, place, output_names, no_grad_set):
         prog = Program()

python/paddle/fluid/tests/unittests/test_conv2d_op.py

@@ -82,18 +82,9 @@ def setUp(self):
         output = conv2d_forward_naive(input, filter, self.groups,
                                       conv2d_param).astype(self.dtype)
 
-        # numpy float16 is binded to paddle::platform::float16 
-        # in tensor_py.h via the help of numpy uint16 because
-        # the internal memory representation of float16 is 
-        # uint16_t in paddle or np.uint16 in numpy, which are
-        # themselves binded together.        
         self.inputs = {
-            #'Input': (input.view(np.uint16)
-            #          if self.dtype == np.float16 else input),
-            #'Filter': (filter.view(np.uint16)
-            #          if self.dtype == np.float16 else filter)
-            'Input': OpTest.create_view(input),
-            'Filter': OpTest.create_view(filter)
+            'Input': OpTest.np_dtype_to_fluid_dtype(input),
+            'Filter': OpTest.np_dtype_to_fluid_dtype(filter)
         }
         self.attrs = {
             'strides': self.stride,
@@ -113,6 +104,8 @@ def test_check_output(self):
             self.check_output()
 
     def test_check_grad(self):
+        if self.dtype == np.float16:
+            return
         if self.use_cudnn:
             place = core.CUDAPlace(0)
             self.check_grad_with_place(
@@ -125,6 +118,8 @@ def test_check_grad(self):
                 set(['Input', 'Filter']), 'Output', max_relative_error=0.02)
 
     def test_check_grad_no_filter(self):
+        if self.dtype == np.float16:
+            return
         if self.use_cudnn:
             place = core.CUDAPlace(0)
             self.check_grad_with_place(
@@ -140,6 +135,8 @@ def test_check_grad_no_filter(self):
                 no_grad_set=set(['Filter']))
 
     def test_check_grad_no_input(self):
+        if self.dtype == np.float16:
+            return
         if self.use_cudnn:
             place = core.CUDAPlace(0)
             self.check_grad_with_place(
@@ -259,46 +256,92 @@ def test_check_output(self):
             if core.is_float16_supported(place):
                 self.check_output_with_place(place, atol=2e-2)
 
-    def test_check_grad(self):
-        pass
-
-    def test_check_grad_no_filter(self):
-        pass
-
-    def test_check_grad_no_input(self):
-        pass
-
 
 class TestCUDNNWithPad(TestWithPad):
     def init_op_type(self):
         self.use_cudnn = True
         self.op_type = "conv2d"
 
 
+class TestFP16CUDNNWithPad(TestCUDNNWithPad):
+    def init_data_type(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                self.check_output_with_place(place, atol=2e-2)
+
+
 class TestCUDNNWithStride(TestWithStride):
     def init_op_type(self):
         self.use_cudnn = True
         self.op_type = "conv2d"
 
 
+class TestFP16CUDNNWithStride(TestCUDNNWithStride):
+    def init_data_type(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                self.check_output_with_place(place, atol=2e-2)
+
+
 class TestCUDNNWithGroup(TestWithGroup):
     def init_op_type(self):
         self.use_cudnn = True
         self.op_type = "conv2d"
 
 
+class TestFP16CUDNNWithGroup(TestCUDNNWithGroup):
+    def init_data_type(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                self.check_output_with_place(place, atol=2e-2)
+
+
 class TestCUDNNWith1x1(TestWith1x1):
     def init_op_type(self):
         self.use_cudnn = True
         self.op_type = "conv2d"
 
 
+class TestFP16CUDNNWith1x1(TestCUDNNWith1x1):
+    def init_data_type(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                self.check_output_with_place(place, atol=2e-2)
+
+
 class TestCUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1):
     def init_op_type(self):
         self.use_cudnn = True
         self.op_type = "conv2d"
 
 
+class TestFP16CUDNNWithInput1x1Filter1x1(TestCUDNNWithInput1x1Filter1x1):
+    def init_data_type(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        if core.is_compiled_with_cuda():
+            place = core.CUDAPlace(0)
+            if core.is_float16_supported(place):
+                self.check_output_with_place(place, atol=2e-2)
+
+
 class TestDepthwiseConv(TestConv2dOp):
     def init_test_case(self):
         self.pad = [1, 1]