My unpool max 2d

paddle/operators/CMakeLists.txt

@@ -191,6 +191,7 @@ set(DEPS_OPS
     sum_op
     pool_op
     maxout_op
+    unpool_op
     pool_with_index_op
     conv_op
     conv_transpose_op
@@ -235,6 +236,7 @@ op_library(adagrad_op DEPS selected_rows_functor)
 op_library(conv_op DEPS vol2col)
 op_library(pool_op DEPS pooling)
 op_library(maxout_op DEPS maxouting)
+op_library(unpool_op DEPS unpooling)
 op_library(pool_with_index_op DEPS pooling)
 op_library(lod_rank_table_op SRCS lod_rank_table_op.cc DEPS lod_rank_table)
 op_library(lod_tensor_to_array_op SRCS lod_tensor_to_array_op.cc DEPS lod_rank_table_op)

paddle/operators/math/CMakeLists.txt

@@ -13,8 +13,9 @@ if(WITH_GPU)
     nv_library(context_project SRCS context_project.cc context_project.cu DEPS device_context math_function)
     nv_library(sequence2batch SRCS sequence2batch.cc sequence2batch.cu DEPS device_context)
     nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions)
-    nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function)
     nv_library(maxouting SRCS maxouting.cc maxouting.cu DEPS device_context)
+    nv_library(unpooling SRCS unpooling.cc unpooling.cu DEPS device_context)
+    nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function)
 else()
     cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context framework_proto)
     cc_library(selected_rows_functor SRCS selected_rows_functor.cc DEPS selected_rows math_function)
@@ -26,8 +27,9 @@ else()
     cc_library(context_project SRCS context_project.cc DEPS device_context math_function)
     cc_library(sequence2batch SRCS sequence2batch.cc DEPS device_context)
     cc_library(lstm_compute SRCS lstm_compute.cc DEPS device_context activation_functions)
-    cc_library(gru_compute SRCS gru_compute.cc DEPS device_context activation_functions math_function)
     cc_library(maxouting SRCS maxouting.cc DEPS device_context)
+    cc_library(unpooling SRCS unpooling.cc DEPS device_context)
+    cc_library(gru_compute SRCS gru_compute.cc DEPS device_context activation_functions math_function)
 endif()
 
 cc_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor)

paddle/operators/math/unpooling.cc

@@ -0,0 +1,91 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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/operators/math/unpooling.h"
+namespace paddle {
+namespace operators {
+namespace math {
+template <typename T>
+class Unpool2dMaxFunctor<platform::CPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& indices, framework::Tensor* output) {
+    const int batch_size = input.dims()[0];
+    const int input_height = input.dims()[2];
+    const int input_width = input.dims()[3];
+    const int output_channels = output->dims()[1];
+    const int output_height = output->dims()[2];
+    const int output_width = output->dims()[3];
+    int input_feasize = input_height * input_width;
+    int output_feasize = output_height * output_width;
+    const T* input_data = input.data<T>();
+    const int* indices_data = indices.data<int>();
+    T* output_data = output->mutable_data<T>(context.GetPlace());
+    for (int b = 0; b < batch_size; ++b) {
+      for (int c = 0; c < output_channels; ++c) {
+        for (int i = 0; i < input_feasize; ++i) {
+          int index = indices_data[i];
+          PADDLE_ENFORCE(index < output_feasize, "err index in unpooling!");
+          output_data[index] = input_data[i];
+        }
+        input_data += input_feasize;
+        indices_data += input_feasize;
+        output_data += output_feasize;
+      }
+    }
+  }
+};
+template <class T>
+class Unpool2dMaxGradFunctor<platform::CPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& indices,
+                  const framework::Tensor& output,
+                  const framework::Tensor& output_grad,
+                  framework::Tensor* input_grad) {
+    const int batch_size = input.dims()[0];
+    const int input_height = input.dims()[2];
+    const int input_width = input.dims()[3];
+    const int output_channels = output.dims()[1];
+    const int output_height = output.dims()[2];
+    const int output_width = output.dims()[3];
+    int input_feasize = input_height * input_width;
+    int output_feasize = output_height * output_width;
+    const int* indices_data = indices.data<int>();
+    const T* output_grad_data = output_grad.data<T>();
+    T* input_grad_data = input_grad->mutable_data<T>(context.GetPlace());
+
+    for (int b = 0; b < batch_size; ++b) {
+      for (int c = 0; c < output_channels; ++c) {
+        for (int i = 0; i < input_feasize; ++i) {
+          int index = indices_data[i];
+          PADDLE_ENFORCE(index < output_feasize, "err index in unpooling!");
+          input_grad_data[i] = output_grad_data[index];
+        }
+        input_grad_data += input_feasize;
+        indices_data += input_feasize;
+        output_grad_data += output_feasize;
+      }
+    }
+  }
+};
+template class Unpool2dMaxGradFunctor<platform::CPUPlace, float>;
+template class Unpool2dMaxGradFunctor<platform::CPUPlace, double>;
+template class Unpool2dMaxFunctor<platform::CPUPlace, float>;
+template class Unpool2dMaxFunctor<platform::CPUPlace, double>;
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/unpooling.cu

@@ -0,0 +1,134 @@
+/* Copyright (c) 2016 paddlepaddle Authors. All Rights Reserve.
+
+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/operators/math/unpooling.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+template <typename T>
+__global__ void KernelUnpool2dMax(const int nthreads, const T* input_data,
+                                  const int* indices_data,
+                                  const int input_height, const int input_width,
+                                  const int channels, T* output_data,
+                                  const int output_height,
+                                  const int output_width) {
+  int in_n_stride = input_height * input_width * channels;
+  int in_c_stride = input_height * input_width;
+  int out_n_stride = output_height * output_width * channels;
+  int out_c_stride = output_height * output_width;
+  int index = blockIdx.x * blockDim.x + threadIdx.x;
+  int offset = blockDim.x * gridDim.x;
+  for (int i = index; i < nthreads; i += offset) {
+    int bidx = i / in_n_stride;
+    int boffset = i % in_n_stride;
+    int cidx = boffset / in_c_stride;
+    int out_offset = bidx * out_n_stride + cidx * out_c_stride;
+    int out_index = indices_data[i];
+    PADDLE_ASSERT(out_index < out_c_stride);
+    output_data[out_offset + out_index] = input_data[i];
+  }
+}
+template <typename T>
+__global__ void KernelUnpool2dMaxGrad(
+    const int nthreads, const T* input_data, const int* indices_data,
+    const int input_height, const int input_width, const int channels,
+    const T* output_data, const T* output_grad, const int output_height,
+    const int output_width, T* input_grad) {
+  int in_n_stride = input_height * input_width * channels;
+  int in_c_stride = input_height * input_width;
+  int out_n_stride = output_height * output_width * channels;
+  int out_c_stride = output_height * output_width;
+  int index = blockIdx.x * blockDim.x + threadIdx.x;
+  int offset = blockDim.x * gridDim.x;
+  for (int i = index; i < nthreads; i += offset) {
+    int bidx = i / in_n_stride;
+    int boffset = i % in_n_stride;
+    int cidx = boffset / in_c_stride;
+    int out_offset = bidx * out_n_stride + cidx * out_c_stride;
+    int out_index = indices_data[i];
+    PADDLE_ASSERT(out_index < out_c_stride);
+    input_grad[i] = output_grad[out_offset + out_index];
+  }
+}
+/*
+ * All tensors are in NCHW format.
+ */
+template <typename T>
+class Unpool2dMaxFunctor<platform::GPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& indices, framework::Tensor* output) {
+    const int batch_size = input.dims()[0];
+    const int input_height = input.dims()[2];
+    const int input_width = input.dims()[3];
+    const int output_channels = output->dims()[1];
+    const int output_height = output->dims()[2];
+    const int output_width = output->dims()[3];
+    const T* input_data = input.data<T>();
+    const int* indices_data = indices.data<int>();
+    T* output_data = output->mutable_data<T>(context.GetPlace());
+    int threads = 1024;
+    int grid = (input.numel() + threads - 1) / threads;
+    KernelUnpool2dMax<
+        T><<<grid, threads, 0,
+             reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                 .stream()>>>(input.numel(), input_data, indices_data,
+                              input_height, input_width, output_channels,
+                              output_data, output_height, output_width);
+  }
+};
+/*
+ * All tensors are in NCHW format.
+ */
+template <typename T>
+class Unpool2dMaxGradFunctor<platform::GPUPlace, T> {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& indices,
+                  const framework::Tensor& output,
+                  const framework::Tensor& output_grad,
+                  framework::Tensor* input_grad) {
+    const int batch_size = input.dims()[0];
+    const int input_height = input.dims()[2];
+    const int input_width = input.dims()[3];
+    const int output_channels = output.dims()[1];
+    const int output_height = output.dims()[2];
+    const int output_width = output.dims()[3];
+    const T* input_data = input.data<T>();
+    const int* indices_data = indices.data<int>();
+    const T* output_data = output.data<T>();
+    const T* output_grad_data = output_grad.data<T>();
+    T* input_grad_data = input_grad->mutable_data<T>(context.GetPlace());
+    int threads = 1024;
+    int grid = (input.numel() + threads - 1) / threads;
+    KernelUnpool2dMaxGrad<
+        T><<<grid, threads, 0,
+             reinterpret_cast<const platform::CUDADeviceContext&>(context)
+                 .stream()>>>(input.numel(), input_data, indices_data,
+                              input_height, input_width, output_channels,
+                              output_data, output_grad_data, output_height,
+                              output_width, input_grad_data);
+  }
+};
+template class Unpool2dMaxGradFunctor<platform::GPUPlace, float>;
+template class Unpool2dMaxGradFunctor<platform::GPUPlace, double>;
+template class Unpool2dMaxFunctor<platform::GPUPlace, float>;
+template class Unpool2dMaxFunctor<platform::GPUPlace, double>;
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/unpooling.h

@@ -0,0 +1,40 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 "paddle/framework/tensor.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+template <typename Place, typename T>
+class Unpool2dMaxFunctor {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& indices, framework::Tensor* output);
+};
+template <typename Place, class T>
+class Unpool2dMaxGradFunctor {
+ public:
+  void operator()(const platform::DeviceContext& context,
+                  const framework::Tensor& input,
+                  const framework::Tensor& indices,
+                  const framework::Tensor& output,
+                  const framework::Tensor& output_grad,
+                  framework::Tensor* input_grad);
+};
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle