Convolution operator

paddle/framework/tensor_impl.h

@@ -130,15 +130,19 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   PADDLE_ENFORCE_LE(end_idx, dims_[0], "Slice end index is out of bound.");
   PADDLE_ENFORCE_LT(begin_idx, end_idx,
                     "Begin index must be less than end index.");
-  PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1.");
-  size_t base = numel() / dims_[0];
-  Tensor dst;
-  dst.holder_ = holder_;
-  DDim dst_dims = dims_;
-  dst_dims[0] = end_idx - begin_idx;
-  dst.Resize(dst_dims);
-  dst.offset_ = offset_ + begin_idx * base * sizeof(T);
-  return dst;
+
+  if (dims_[0] == 1) {
+    return *this;
+  } else {
+    size_t base = numel() / dims_[0];
+    Tensor dst;
+    dst.holder_ = holder_;
+    DDim dst_dims = dims_;
+    dst_dims[0] = end_idx - begin_idx;
+    dst.Resize(dst_dims);
+    dst.offset_ = offset_ + begin_idx * base * sizeof(T);
+    return dst;
+  }
 }
 
 inline Tensor& Tensor::Resize(const DDim& dims) {

paddle/operators/conv2d_op.cc

@@ -0,0 +1,124 @@
+/* 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/gemm_conv2d_op.h"
+
+namespace paddle {
+namespace operators {
+
+int outputSize(int input_size, int filter_size, int padding, int stride) {
+  int output_size = (input_size - filter_size + 2 * padding) / stride + 1;
+  return output_size;
+}
+
+class Conv2DOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto in = ctx.Input<Tensor>("Input");
+    auto filter = ctx.Input<Tensor>("Filter");
+    auto out = ctx.Output<framework::LoDTensor>("Output");
+    std::vector<int> strides = Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = Attr<std::vector<int>>("paddings");
+    int groups = Attr<int>("groups");
+    int input_channels = in->dims()[1];
+    int output_channels = filter->dims()[0];
+
+    PADDLE_ENFORCE_EQ(in->dims().size(), 4, "Conv2DOp intput should be 4-D.");
+    PADDLE_ENFORCE_EQ(filter->dims().size(), 4,
+                      "Conv2DOp filter should be 4-D.");
+    PADDLE_ENFORCE_EQ(input_channels, filter->dims()[1] * groups,
+                      "The number of input channels should be equal to filter "
+                      "channels * groups.");
+    PADDLE_ENFORCE_EQ(
+        output_channels % groups, 0,
+        "The number of output channels should be divided by groups.");
+
+    auto output_height =
+        outputSize(in->dims()[2], filter->dims()[2], paddings[0], strides[0]);
+    auto output_width =
+        outputSize(in->dims()[3], filter->dims()[3], paddings[1], strides[1]);
+    out->Resize(
+        {in->dims()[0], filter->dims()[0], output_height, output_width});
+  }
+};
+
+class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  Conv2DOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput(
+        "Input",
+        "The input tensor of convolution operator. "
+        "The format of input tensor is NCHW. Where N is batch size, C is the "
+        "number of channels, H and W is the height and width of image.");
+    AddInput(
+        "Filter",
+        "The filter tensor of convolution operator."
+        "The format of the filter tensor is MCHW, where M is the number of "
+        "output image channels, C is the number of input image channels, "
+        "H and W is height and width of filter. "
+        "If the groups attribute is greater than 1, C equal the number of "
+        "input image channels divided by the groups.");
+    AddOutput("Output",
+              "The output tensor of convolution operator."
+              "The format of output tensor is also NCHW.");
+    AddComment(R"DOC(
+The convolution operation calculates the output based on the input, filter
+and strides, paddings, groups parameters. The size of each dimension of the
+parameters is checked in the infer-shape.
+)DOC");
+    AddAttr<std::vector<int>>("strides", "strides of convolution operator.");
+    AddAttr<std::vector<int>>("paddings", "paddings of convolution operator.");
+    AddAttr<int>(
+        "groups",
+        "group size of convolution operator. "
+        "Refer to grouped convolution in Alex Krizhevsky's paper: "
+        "when group=2, the first half of the filters are only connected to the "
+        "first half of the input channels, and the second half only connected "
+        "to the second half.")
+        .SetDefault(1);
+  }
+};
+
+class Conv2DOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto in = ctx.Input<Tensor>("Input");
+    auto filter = ctx.Input<Tensor>("Filter");
+    auto d_in =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Input"));
+    auto d_filter =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Filter"));
+    d_in->Resize(in->dims());
+    d_filter->Resize(filter->dims());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(conv2d, ops::Conv2DOp, ops::Conv2DOpMaker, conv2d_grad,
+            ops::Conv2DOpGrad);
+
+REGISTER_OP_CPU_KERNEL(
+    conv2d, ops::GemmConv2DKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    conv2d_grad, ops::GemmConvGrad2DKernel<paddle::platform::CPUPlace, float>);

paddle/operators/conv2d_op.cu

@@ -0,0 +1,22 @@
+/* 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/gemm_conv2d_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_GPU_KERNEL(
+    conv2d, ops::GemmConv2DKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    conv2d_grad, ops::GemmConvGrad2DKernel<paddle::platform::GPUPlace, float>);

paddle/operators/gemm_conv2d_op.h

@@ -0,0 +1,214 @@
+/* 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/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/im2col.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename Place, typename T>
+class GemmConv2DKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* input = context.Input<Tensor>("Input");
+    // The filter will be reshaped in the calculations,
+    // so here use an assignment operation,
+    // that avoids modifying the variable in the Scope.
+    Tensor filter = *context.Input<Tensor>("Filter");
+    Tensor* output = context.Output<Tensor>("Output");
+    output->mutable_data<T>(context.GetPlace());
+
+    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
+    int groups = context.Attr<int>("groups");
+
+    int batch_size = input->dims()[0];
+    int input_channels = input->dims()[1];
+    int filter_height = filter.dims()[filter.dims().size() - 2];
+    int filter_width = filter.dims()[filter.dims().size() - 1];
+    int output_channels = output->dims()[1];
+    int output_height = output->dims()[2];
+    int output_width = output->dims()[3];
+
+    paddle::operators::math::Im2ColFunctor<
+        paddle::operators::math::ColFormat::kCFO, Place, T>
+        im2col;
+    // use col_shape in the im2col calculation
+    framework::DDim col_shape = {input_channels / groups, filter_height,
+                                 filter_width, output_height, output_width};
+    // use col_matrix_shape in the gemm calculation
+    framework::DDim col_matrix_shape = {
+        input_channels / groups * filter_height * filter_width,
+        output_height * output_width};
+    Tensor col;
+    col.mutable_data<T>(col_shape, context.GetPlace());
+    // col_matrix shares the same piece of data with col,
+    // but will be reshaped into a two-dimensional matrix shape
+    // to call the matrix multiplication interface.
+    Tensor col_matrix = col;
+    col_matrix.Resize(col_matrix_shape);
+
+    framework::DDim input_shape = {input->dims()[1], input->dims()[2],
+                                   input->dims()[3]};
+    framework::DDim filter_matrix_shape = {filter.dims()[0],
+                                           filter.numel() / filter.dims()[0]};
+    filter.Resize(filter_matrix_shape);
+
+    framework::DDim output_matrix_shape = {output_channels,
+                                           output_height * output_width};
+
+    auto* device_context =
+        const_cast<platform::DeviceContext*>(context.device_context_);
+
+    // convolution operator: im2col + gemm
+    int in_step = input_channels / groups;
+    int out_step = output_channels / groups;
+    for (int i = 0; i < batch_size; i++) {
+      Tensor in_batch = input->Slice<T>(i, i + 1).Resize(input_shape);
+      Tensor out_batch = output->Slice<T>(i, i + 1).Resize(output_matrix_shape);
+      for (int g = 0; g < groups; g++) {
+        // im2col
+        Tensor in_slice = in_batch.Slice<T>(g * in_step, (g + 1) * in_step);
+        im2col(in_slice, col, strides[0], strides[1], paddings[0], paddings[1],
+               device_context);
+
+        // gemm
+        Tensor out_slice = out_batch.Slice<T>(g * out_step, (g + 1) * out_step);
+        Tensor filter_slice = filter.Slice<T>(g * out_step, (g + 1) * out_step);
+        math::matmul<Place, T>(filter_slice, false, col_matrix, false, T(1.0),
+                               &out_slice, T(0.0), device_context);
+      }
+    }
+  }
+};
+
+template <typename Place, typename T>
+class GemmConvGrad2DKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    const Tensor* input = context.Input<Tensor>("Input");
+    const Tensor* output_grad =
+        context.Input<Tensor>(framework::GradVarName("Output"));
+    Tensor* input_grad =
+        context.Output<Tensor>(framework::GradVarName("Input"));
+    Tensor* filter_grad_ =
+        context.Output<Tensor>(framework::GradVarName("Filter"));
+    input_grad->mutable_data<T>(context.GetPlace());
+    filter_grad_->mutable_data<T>(context.GetPlace());
+
+    // The filter and filter_grad will be reshaped in the calculations,
+    // so here use an assignment operation,
+    // that avoids modifying the variable in the Scope.
+    Tensor filter = *context.Input<Tensor>("Filter");
+    Tensor filter_grad = *filter_grad_;
+
+    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
+    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
+    int groups = context.Attr<int>("groups");
+
+    int batch_size = input->dims()[0];
+    int input_channels = input->dims()[1];
+    int filter_height = filter.dims()[filter.dims().size() - 2];
+    int filter_width = filter.dims()[filter.dims().size() - 1];
+    int output_channels = output_grad->dims()[1];
+    int output_height = output_grad->dims()[2];
+    int output_width = output_grad->dims()[3];
+
+    paddle::operators::math::Col2ImFunctor<
+        paddle::operators::math::ColFormat::kCFO, Place, T>
+        col2im;
+    paddle::operators::math::Im2ColFunctor<
+        paddle::operators::math::ColFormat::kCFO, Place, T>
+        im2col;
+    // use col_shape in the im2col and col2im calculation
+    framework::DDim col_shape = {input_channels / groups, filter_height,
+                                 filter_width, output_height, output_width};
+    // use col_matrix_shape in the gemm calculation
+    framework::DDim col_matrix_shape = {
+        input_channels / groups * filter_height * filter_width,
+        output_height * output_width};
+    Tensor col;
+    col.mutable_data<T>(col_shape, context.GetPlace());
+    // col_matrix shares the same piece of data with col,
+    // but will be reshaped into a two-dimensional matrix shape
+    // to call the matrix multiplication interface.
+    Tensor col_matrix = col;
+    col_matrix.Resize(col_matrix_shape);
+
+    framework::DDim input_shape = {input->dims()[1], input->dims()[2],
+                                   input->dims()[3]};
+    framework::DDim output_matrix_shape = {
+        output_grad->dims()[1],
+        output_grad->dims()[2] * output_grad->dims()[3]};
+
+    framework::DDim filter_matrix_shape = {filter.dims()[0],
+                                           filter.numel() / filter.dims()[0]};
+    filter.Resize(filter_matrix_shape);
+    filter_grad.Resize(filter_matrix_shape);
+
+    auto t1 = framework::EigenVector<T>::Flatten(filter_grad);
+    t1.device(context.GetEigenDevice<Place>()) = t1.constant(static_cast<T>(0));
+    auto t2 = framework::EigenVector<T>::Flatten(*input_grad);
+    t2.device(context.GetEigenDevice<Place>()) = t2.constant(static_cast<T>(0));
+
+    auto* device_context =
+        const_cast<platform::DeviceContext*>(context.device_context_);
+
+    // convolution backward input operator:  gemm + col2im
+    // convolution backward weight operator: im2col + gemm
+    int in_step = input_channels / groups;
+    int out_step = output_channels / groups;
+    for (int i = 0; i < batch_size; i++) {
+      Tensor out_grad_batch =
+          output_grad->Slice<T>(i, i + 1).Resize(output_matrix_shape);
+      Tensor in_grad_batch = input_grad->Slice<T>(i, i + 1).Resize(input_shape);
+      Tensor in_batch = input->Slice<T>(i, i + 1).Resize(input_shape);
+      for (int g = 0; g < groups; g++) {
+        // gemm
+        Tensor out_grad_slice =
+            out_grad_batch.Slice<T>(g * out_step, (g + 1) * out_step);
+        Tensor filter_slice = filter.Slice<T>(g * out_step, (g + 1) * out_step);
+        math::matmul<Place, T>(filter_slice, true, out_grad_slice, false,
+                               T(1.0), &col_matrix, T(0.0), device_context);
+
+        // col2im
+        Tensor in_grad_slice =
+            in_grad_batch.Slice<T>(g * in_step, (g + 1) * in_step);
+        col2im(in_grad_slice, col, strides[0], strides[1], paddings[0],
+               paddings[1], device_context);
+
+        // im2col
+        Tensor in_slice = in_batch.Slice<T>(g * in_step, (g + 1) * in_step);
+        im2col(in_slice, col, strides[0], strides[1], paddings[0], paddings[1],
+               device_context);
+
+        // gemm
+        Tensor filter_grad_slice =
+            filter_grad.Slice<T>(g * out_step, (g + 1) * out_step);
+        math::matmul<Place, T>(out_grad_slice, false, col_matrix, true, T(1.0),
+                               &filter_grad_slice, T(1.0), device_context);
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle