【PaddlePaddle Hackathon 4】：为maxout算子支持 float16 数据类型 (#50976)

* support fp16 for maxout op

* format code

* change api

* add test for static float16

* format code

* formatting code

* atol alignment

* experiment—1

* experiment-2

* experiment-3

* format code

paddle/phi/kernels/funcs/maxouting.cu

@@ -175,9 +175,11 @@ void MaxOutGradFunctor<DeviceContext, T>::operator()(
 }
 
 template class MaxOutGradFunctor<phi::GPUContext, float>;
+template class MaxOutGradFunctor<phi::GPUContext, phi::dtype::float16>;
 template class MaxOutGradFunctor<phi::GPUContext, double>;
 
 template class MaxOutFunctor<phi::GPUContext, float>;
+template class MaxOutFunctor<phi::GPUContext, phi::dtype::float16>;
 template class MaxOutFunctor<phi::GPUContext, double>;
 
 }  // namespace funcs

paddle/phi/kernels/gpu/maxout_grad_kernel.cu

@@ -15,5 +15,10 @@
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/impl/maxout_grad_kernel_impl.h"
 
-PD_REGISTER_KERNEL(
-    maxout_grad, GPU, ALL_LAYOUT, phi::MaxOutGradKernel, float, double) {}
+PD_REGISTER_KERNEL(maxout_grad,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::MaxOutGradKernel,
+                   float,
+                   phi::dtype::float16,
+                   double) {}

paddle/phi/kernels/gpu/maxout_kernel.cu

@@ -15,4 +15,10 @@
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/impl/maxout_kernel_impl.h"
 
-PD_REGISTER_KERNEL(maxout, GPU, ALL_LAYOUT, phi::MaxOutKernel, float, double) {}
+PD_REGISTER_KERNEL(maxout,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::MaxOutKernel,
+                   float,
+                   phi::dtype::float16,
+                   double) {}

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

@@ -136,5 +136,40 @@ class TestMaxoutAPI(unittest.TestCase):
             self.assertRaises(ValueError, F.maxout, x_float32, 2, 2)
 
 
+class TestMaxOutOpFP16(TestMaxOutOp):
+    def set_attrs(self):
+        self.dtype = 'float16'
+
+
+class TestMaxoutFP16Case1(TestMaxOutOpFP16):
+    def set_attrs(self):
+        self.axis = -1
+
+
+class TestMaxoutFP16Case2(TestMaxOutOpFP16):
+    def set_attrs(self):
+        self.axis = 3
+
+
+@unittest.skipIf(
+    not core.is_compiled_with_cuda(), "core is not compiled with CUDA"
+)
+class TestMaxoutStaticAPIFP16(unittest.TestCase):
+    def setUp(self):
+        self.x_np = np.random.uniform(-1, 1, [2, 6, 5, 4]).astype(np.float16)
+        self.groups = 2
+        self.axis = 1
+        self.place = paddle.CUDAPlace(0)
+
+    def test_static_api(self):
+        with paddle.static.program_guard(paddle.static.Program()):
+            x = paddle.static.data('X', self.x_np.shape, self.x_np.dtype)
+            out = F.maxout(x, self.groups, self.axis)
+            exe = paddle.static.Executor(self.place)
+            res = exe.run(feed={'X': self.x_np}, fetch_list=[out])
+        out_ref = maxout_forward_naive(self.x_np, self.groups, self.axis)
+        np.testing.assert_allclose(out_ref, res[0], rtol=1e-05)
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/nn/functional/activation.py

@@ -787,7 +787,7 @@ def maxout(x, groups, axis=1, name=None):
 
     Parameters:
         x (Tensor): The input is 4-D Tensor with shape [N, C, H, W] or [N, H, W, C], the data type
-            of input is float32 or float64.
+            of input is float16, float32 or float64.
         groups (int): The groups number of maxout. `groups` specifies the
             index of channel dimension where maxout will be performed. This must be
             a factor of number of features.
@@ -822,7 +822,9 @@ def maxout(x, groups, axis=1, name=None):
     if in_dygraph_mode():
         return _C_ops.maxout(x, groups, axis)
     else:
-        check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'maxout')
+        check_variable_and_dtype(
+            x, 'x', ['float16', 'float32', 'float64'], 'maxout'
+        )
         if axis not in [1, -1, 3]:
             raise ValueError(
                 "Attr(axis) should be 1 when data format is NCHW, -1 or 3 when data format is NHWC. Received "