add kernel for flatten_op, test=develop (#19472)

* add kernel for flatten_op, test=develop

* add kernel for flatten_op, test=develop

* fix the license and remove redundant code, test=develop

paddle/fluid/operators/flatten_op.cu.cc

@@ -0,0 +1,44 @@
+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/fluid/operators/flatten_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_CUDA_KERNEL(
+    flatten, ops::FlattenKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::FlattenKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::FlattenKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::FlattenKernel<paddle::platform::CUDADeviceContext, int8_t>,
+    ops::FlattenKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    flatten_grad,
+    ops::FlattenGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::FlattenGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::FlattenGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::FlattenGradKernel<paddle::platform::CUDADeviceContext, int8_t>,
+    ops::FlattenGradKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    flatten2, ops::Flatten2Kernel<paddle::platform::CUDADeviceContext, float>,
+    ops::Flatten2Kernel<paddle::platform::CUDADeviceContext, double>,
+    ops::Flatten2Kernel<paddle::platform::CUDADeviceContext, int>,
+    ops::Flatten2Kernel<paddle::platform::CUDADeviceContext, int8_t>,
+    ops::Flatten2Kernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    flatten2_grad,
+    ops::Flatten2GradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::Flatten2GradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::Flatten2GradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::Flatten2GradKernel<paddle::platform::CUDADeviceContext, int8_t>,
+    ops::Flatten2GradKernel<paddle::platform::CUDADeviceContext, int64_t>);

paddle/fluid/operators/flatten_op.h

@@ -0,0 +1,116 @@
+/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include <vector>
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/math/blas.h"
+#include "paddle/fluid/operators/math/math_function.h"
+#include "paddle/fluid/operators/math/pooling.h"
+#include "paddle/fluid/platform/device_context.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class FlattenKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto *in = context.Input<framework::LoDTensor>("X");
+    auto *out = context.Output<framework::LoDTensor>("Out");
+
+    auto &axes = context.Attr<int>("axis");
+    auto x_dims = in->dims();
+    auto out_dims = framework::make_ddim(GetOutputShape(axes, x_dims));
+
+    out->mutable_data(context.GetPlace(), in->type());
+    framework::TensorCopy(
+        *in, context.GetPlace(),
+        context.template device_context<platform::DeviceContext>(), out);
+    out->Resize(out_dims);
+  }
+
+  static std::vector<int32_t> GetOutputShape(const int axis,
+                                             const framework::DDim &in_dims) {
+    int64_t outer = 1, inner = 1;
+    for (int i = 0; i < in_dims.size(); ++i) {
+      if (i < axis) {
+        outer *= in_dims[i];
+      } else {
+        inner *= in_dims[i];
+      }
+    }
+    std::vector<int32_t> out_shape(2);
+    out_shape[0] = outer;
+    out_shape[1] = inner;
+    return out_shape;
+  }
+};
+
+template <typename DeviceContext, typename T>
+class FlattenGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto *d_x = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *d_out =
+        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
+    auto in_dims = ctx.Input<framework::LoDTensor>("X")->dims();
+
+    d_x->mutable_data(ctx.GetPlace(), d_out->type());
+    framework::TensorCopySync(*d_out, ctx.GetPlace(), d_x);
+    d_x->Resize(in_dims);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class Flatten2Kernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {
+    auto &axes = context.Attr<int>("axis");
+
+    auto *in = context.Input<framework::LoDTensor>("X");
+    auto x_dims = in->dims();
+
+    auto *out = context.Output<framework::LoDTensor>("Out");
+
+    auto out_dims = framework::make_ddim(
+        FlattenKernel<DeviceContext, T>::GetOutputShape(axes, x_dims));
+
+    out->mutable_data(context.GetPlace(), in->type());
+    framework::TensorCopy(
+        *in, context.GetPlace(),
+        context.template device_context<platform::DeviceContext>(), out);
+    out->Resize(out_dims);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class Flatten2GradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto *d_x = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *d_out =
+        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"));
+
+    auto xshape_dims = ctx.Input<framework::LoDTensor>("XShape")->dims();
+    auto x_dims = framework::slice_ddim(xshape_dims, 1, xshape_dims.size());
+
+    d_x->mutable_data(ctx.GetPlace(), d_out->type());
+    framework::TensorCopySync(*d_out, ctx.GetPlace(), d_x);
+    d_x->Resize(x_dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

python/paddle/fluid/tests/unittests/test_flatten2_op.py

@@ -0,0 +1,73 @@
+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+
+from op_test import OpTest
+
+
+class TestFlattenOp(OpTest):
+    def setUp(self):
+        self.op_type = "flatten2"
+        self.init_test_case()
+        self.inputs = {"X": np.random.random(self.in_shape).astype("float32")}
+        self.init_attrs()
+        self.outputs = {
+            "Out": self.inputs["X"].reshape(self.new_shape),
+            "XShape": np.random.random(self.in_shape).astype("float32")
+        }
+
+    def test_check_output(self):
+        self.check_output(no_check_set=["XShape"])
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+    def init_test_case(self):
+        self.in_shape = (3, 2, 2, 5)
+        self.axis = 1
+        self.new_shape = (3, 20)
+
+    def init_attrs(self):
+        self.attrs = {"axis": self.axis}
+
+
+class TestFlattenOp(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 2, 3)
+        self.axis = 0
+        self.new_shape = (1, 36)
+
+
+class TestFlattenOpWithDefaultAxis(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 2, 3)
+        self.new_shape = (3, 12)
+
+    def init_attrs(self):
+        self.attrs = {}
+
+
+class TestFlattenOpSixDims(TestFlattenOp):
+    def init_test_case(self):
+        self.in_shape = (3, 2, 3, 2, 4, 4)
+        self.axis = 4
+        self.new_shape = (36, 16)
+
+
+if __name__ == "__main__":
+    unittest.main()

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

@@ -1,4 +1,4 @@
-#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -22,17 +22,14 @@ from op_test import OpTest
 
 class TestFlattenOp(OpTest):
     def setUp(self):
-        self.op_type = "flatten2"
+        self.op_type = "flatten"
         self.init_test_case()
         self.inputs = {"X": np.random.random(self.in_shape).astype("float32")}
         self.init_attrs()
-        self.outputs = {
+        self.outputs = {"Out": self.inputs["X"].reshape(self.new_shape)}
-            "Out": self.inputs["X"].reshape(self.new_shape),
-            "XShape": np.random.random(self.in_shape).astype("float32")
-        }
 
     def test_check_output(self):
-        self.check_output(no_check_set=["XShape"])
+        self.check_output()
 
     def test_check_grad(self):
         self.check_grad(["X"], "Out")