conv_op.h 14.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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

H
hedaoyuan 已提交
17
#include "paddle/framework/eigen.h"
18 19 20
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/im2col.h"
#include "paddle/operators/math/math_function.h"
C
chengduoZH 已提交
21
#include "paddle/operators/math/vol2col.h"
22 23 24 25 26 27

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;

武毅 已提交
28 29
// Base convolution operator definations for other conv
// like operators to reuse the implementation.
C
chengduoZH 已提交
30
inline int OutputSize(int input_size, int filter_size, int dilation,
C
chengduoZH 已提交
31 32 33
                      int padding, int stride) {
  const int dkernel = dilation * (filter_size - 1) + 1;
  const int output_size = (input_size + 2 * padding - dkernel) / stride + 1;
武毅 已提交
34 35
  return output_size;
}
C
chengduoZH 已提交
36 37 38
inline bool IsExpand(std::vector<int64_t>& filter_dim,
                     std::vector<int>& strides, std::vector<int>& paddings,
                     std::vector<int>& dilations) {
C
chengduoZH 已提交
39 40
  bool filter_1 = true, strides_1 = true, padding_0 = true, dilation_1 = true;
  for (size_t j = 0; j < strides.size(); ++j) {
C
chengduoZH 已提交
41
    filter_1 = filter_1 && (static_cast<int>(filter_dim[j + 2]) == 1);
C
chengduoZH 已提交
42 43 44
    strides_1 = strides_1 && (strides[j] == 1);
    padding_0 = padding_0 && (paddings[j] == 0);
    dilation_1 = dilation_1 && (dilations[j] == 1);
C
chengduoZH 已提交
45
  }
C
chengduoZH 已提交
46
  return !(filter_1 && strides_1 && padding_0 && dilation_1);
C
chengduoZH 已提交
47
}
武毅 已提交
48 49 50 51 52 53 54 55 56

// Define Op classes in .h file so that other conv
// operator implementations can reuse the code.
class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  Conv2DOpMaker(framework::OpProto* proto,
                framework::OpAttrChecker* op_checker);
};

C
chengduoZH 已提交
57 58 59 60 61 62 63
class Conv3DOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  Conv3DOpMaker(framework::OpProto* proto,
                framework::OpAttrChecker* op_checker);
};

class ConvOp : public framework::OperatorWithKernel {
武毅 已提交
64 65 66 67 68
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
  void InferShape(framework::InferShapeContext* ctx) const override;
};

C
chengduoZH 已提交
69
class ConvOpGrad : public framework::OperatorWithKernel {
武毅 已提交
70 71 72 73 74
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
  void InferShape(framework::InferShapeContext* ctx) const override;
};

Q
QI JUN 已提交
75
template <typename DeviceContext, typename T>
C
chengduoZH 已提交
76
class GemmConvKernel : public framework::OpKernel<T> {
77 78 79
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    const Tensor* input = context.Input<Tensor>("Input");
H
hedaoyuan 已提交
80 81 82 83
    // 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");
84 85 86
    Tensor* output = context.Output<Tensor>("Output");
    output->mutable_data<T>(context.GetPlace());

C
chengduoZH 已提交
87
    int groups = context.Attr<int>("groups");
88 89
    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
C
chengduoZH 已提交
90
    std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
91

C
chengduoZH 已提交
92 93
    const int batch_size = static_cast<int>(input->dims()[0]);

C
chengduoZH 已提交
94
    // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w}
C
chengduoZH 已提交
95
    std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
C
chengduoZH 已提交
96
    // output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w}
C
chengduoZH 已提交
97
    std::vector<int64_t> output_shape_vec(framework::vectorize(output->dims()));
98

H
hedaoyuan 已提交
99
    // use col_shape in the im2col calculation
C
chengduoZH 已提交
100 101
    // col_shape_vec: {i_c/g, k_h, k_w, o_h, o_w} or {i_c/g, k_d, k_h, k_w, o_d,
    // o_h, o_w}
C
chengduoZH 已提交
102 103 104 105 106 107 108
    size_t data_dim = filter_shape_vec.size() - 2;
    std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
    col_shape_vec[0] = input->dims()[1] / groups;
    for (size_t j = 0; j < data_dim; ++j) {
      col_shape_vec[j + 1] = filter_shape_vec[j + 2];
      col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
    }
C
chengduoZH 已提交
109 110
    framework::DDim col_shape(framework::make_ddim(col_shape_vec));

H
hedaoyuan 已提交
111
    // use col_matrix_shape in the gemm calculation
C
chengduoZH 已提交
112 113 114
    // size: (i_c/g * k_h * k_w, o_h * o_w) or (i_c/g * k_d * k_h * k_w, o_d *
    // o_h * o_w)
    framework::DDim col_matrix_shape =
C
chengduoZH 已提交
115
        framework::flatten_to_2d(col_shape, data_dim + 1);
C
chengduoZH 已提交
116

C
chengduoZH 已提交
117
    bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations);
H
hedaoyuan 已提交
118
    Tensor col;
H
hedaoyuan 已提交
119 120 121
    // 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.
C
chengduoZH 已提交
122
    Tensor col_matrix;
C
chengduoZH 已提交
123
    if (is_expand) {
C
chengduoZH 已提交
124 125 126 127
      col.mutable_data<T>(col_shape, context.GetPlace());
      col_matrix.ShareDataWith(col);
      col_matrix.Resize(col_matrix_shape);
    }
128

C
chengduoZH 已提交
129 130 131
    framework::DDim input_shape = framework::slice_ddim(
        input->dims(), 1, static_cast<int>(input->dims().size()));

H
hedaoyuan 已提交
132 133
    framework::DDim filter_matrix_shape = {filter.dims()[0],
                                           filter.numel() / filter.dims()[0]};
H
hedaoyuan 已提交
134 135
    filter.Resize(filter_matrix_shape);

C
chengduoZH 已提交
136 137 138 139 140 141 142 143
    framework::DDim output_matrix_shape = {
        output->dims()[1],
        output->numel() / (output->dims()[0] * output->dims()[1])};

    // convolution operator: im2col(or vol2col) + gemm
    int in_step = static_cast<int>(input->dims()[1]) / groups;
    int out_step = static_cast<int>(output->dims()[1]) / groups;

Q
QI JUN 已提交
144 145
    math::Vol2ColFunctor<DeviceContext, T> vol2col;
    math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
C
chengduoZH 已提交
146

Q
QI JUN 已提交
147
    auto& dev_ctx = context.template device_context<DeviceContext>();
C
chengduoZH 已提交
148 149 150
    for (int i = 0; i < batch_size; i++) {
      Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
      Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
C
chengduoZH 已提交
151

C
chengduoZH 已提交
152 153
      for (int g = 0; g < groups; g++) {
        Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
H
hedaoyuan 已提交
154

C
chengduoZH 已提交
155
        if (!is_expand) {
C
chengduoZH 已提交
156 157 158
          col.ShareDataWith(in_slice);
          col_matrix.ShareDataWith(col);
          col_matrix.Resize(col_matrix_shape);
C
chengduoZH 已提交
159
        } else if (data_dim == 2U) {
C
chengduoZH 已提交
160
          // im2col
Q
QI JUN 已提交
161
          im2col(dev_ctx, in_slice, dilations, strides,
C
chengduoZH 已提交
162 163 164
                 std::vector<int>{paddings[0], paddings[1], paddings[0],
                                  paddings[1]},
                 &col);
C
chengduoZH 已提交
165
        } else if (data_dim == 3U) {
C
chengduoZH 已提交
166
          // vol2col
Q
QI JUN 已提交
167
          vol2col(dev_ctx, in_slice, dilations, strides, paddings, &col);
C
chengduoZH 已提交
168
        }
C
chengduoZH 已提交
169 170 171 172

        // gemm
        Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
        Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
Q
QI JUN 已提交
173 174
        math::matmul<DeviceContext, T>(dev_ctx, filter_slice, false, col_matrix,
                                       false, T(1.0), &out_slice, T(0.0));
H
hedaoyuan 已提交
175
      }
176 177 178 179
    }
  }
};

Q
QI JUN 已提交
180
template <typename DeviceContext, typename T>
C
chengduoZH 已提交
181
class GemmConvGradKernel : public framework::OpKernel<T> {
182 183
 public:
  void Compute(const framework::ExecutionContext& context) const override {
H
hedaoyuan 已提交
184 185 186 187 188
    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"));
H
hedaoyuan 已提交
189
    Tensor* filter_grad =
H
hedaoyuan 已提交
190
        context.Output<Tensor>(framework::GradVarName("Filter"));
H
hedaoyuan 已提交
191 192 193 194
    // 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");
H
hedaoyuan 已提交
195

C
chengduoZH 已提交
196 197
    if (!input_grad && !filter_grad) return;

C
chengduoZH 已提交
198
    int groups = context.Attr<int>("groups");
H
hedaoyuan 已提交
199 200
    std::vector<int> strides = context.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
C
chengduoZH 已提交
201
    std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
H
hedaoyuan 已提交
202

C
chengduoZH 已提交
203
    const int batch_size = static_cast<int>(input->dims()[0]);
H
hedaoyuan 已提交
204

C
chengduoZH 已提交
205
    // filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w}
C
chengduoZH 已提交
206
    std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
C
chengduoZH 已提交
207
    // output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w}
C
chengduoZH 已提交
208 209
    std::vector<int64_t> output_shape_vec(
        framework::vectorize(output_grad->dims()));
C
chengduoZH 已提交
210

C
chengduoZH 已提交
211 212 213
    // use col_shape in the im2col calculation
    // col_shape_vec: {i_c/g, k_h, k_w, o_h, o_w} or {i_c/g, k_d, k_h, k_w, o_d,
    // o_h, o_w}
C
chengduoZH 已提交
214 215 216 217 218 219 220
    size_t data_dim = filter_shape_vec.size() - 2;
    std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
    col_shape_vec[0] = input->dims()[1] / groups;
    for (size_t j = 0; j < data_dim; ++j) {
      col_shape_vec[j + 1] = filter_shape_vec[j + 2];
      col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
    }
C
chengduoZH 已提交
221
    framework::DDim col_shape(framework::make_ddim(col_shape_vec));
C
chengduoZH 已提交
222 223

    // use col_matrix_shape in the gemm calculation
C
chengduoZH 已提交
224 225 226 227
    // size: (i_c/g * k_h * k_w, o_h * o_w)
    // or
    // (i_c/g * k_d * k_h * k_w, o_d * o_h * o_w)
    framework::DDim col_matrix_shape =
C
chengduoZH 已提交
228
        framework::flatten_to_2d(col_shape, data_dim + 1);
C
chengduoZH 已提交
229 230 231

    framework::DDim input_shape = framework::slice_ddim(
        input->dims(), 1, static_cast<int>(input->dims().size()));
C
chengduoZH 已提交
232

C
chengduoZH 已提交
233 234
    framework::DDim filter_matrix_shape = {filter.dims()[0],
                                           filter.numel() / filter.dims()[0]};
C
chengduoZH 已提交
235 236 237
    filter.Resize(filter_matrix_shape);

    framework::DDim output_matrix_shape = {
C
chengduoZH 已提交
238 239 240
        output_grad->dims()[1],
        output_grad->numel() /
            (output_grad->dims()[0] * output_grad->dims()[1])};
C
chengduoZH 已提交
241

C
chengduoZH 已提交
242 243 244 245
    // convolution backward input operator:  gemm + col2im(or col2vol)
    // convolution backward weight operator: im2col(or vol2col) + gemm
    int in_step = static_cast<int>(input->dims()[1]) / groups;
    int out_step = static_cast<int>(output_grad->dims()[1]) / groups;
C
chengduoZH 已提交
246

C
chengduoZH 已提交
247
    bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations);
C
chengduoZH 已提交
248 249 250 251
    Tensor col;
    // 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.
C
chengduoZH 已提交
252
    Tensor col_matrix;
C
chengduoZH 已提交
253
    if (is_expand) {
C
chengduoZH 已提交
254 255 256 257
      col.mutable_data<T>(col_shape, context.GetPlace());
      col_matrix.ShareDataWith(col);
      col_matrix.Resize(col_matrix_shape);
    }
C
chengduoZH 已提交
258

Q
QI JUN 已提交
259 260
    math::SetConstant<DeviceContext, T> set_zero;
    auto& dev_ctx = context.template device_context<DeviceContext>();
C
chengduoZH 已提交
261 262 263 264

    if (input_grad) {
      input_grad->mutable_data<T>(context.GetPlace());

C
chengduoZH 已提交
265 266 267
      // if is_expand is false, the operation of set_zero is unnecessary,
      // because math::matmul will reset input_grad.
      if (is_expand) {
C
chengduoZH 已提交
268
        set_zero(dev_ctx, input_grad, static_cast<T>(0));
C
chengduoZH 已提交
269
      }
Q
QI JUN 已提交
270 271
      math::Col2VolFunctor<DeviceContext, T> col2vol;
      math::Col2ImFunctor<math::ColFormat::kCFO, DeviceContext, T> col2im;
C
chengduoZH 已提交
272

C
chengduoZH 已提交
273 274 275 276 277 278 279 280 281 282 283 284 285 286
      for (int i = 0; i < batch_size; i++) {
        Tensor out_grad_batch =
            output_grad->Slice(i, i + 1).Resize(output_matrix_shape);
        Tensor in_grad_batch = input_grad->Slice(i, i + 1).Resize(input_shape);
        for (int g = 0; g < groups; g++) {
          // gemm
          Tensor out_grad_slice =
              out_grad_batch.Slice(g * out_step, (g + 1) * out_step);
          Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);

          Tensor in_grad_slice =
              in_grad_batch.Slice(g * in_step, (g + 1) * in_step);

          if (!is_expand) {
C
chengduoZH 已提交
287 288
            col_matrix.ShareDataWith(in_grad_slice);
            col_matrix.Resize(col_matrix_shape);
C
chengduoZH 已提交
289
          }
Q
QI JUN 已提交
290 291 292
          math::matmul<DeviceContext, T>(dev_ctx, filter_slice, true,
                                         out_grad_slice, false, T(1.0),
                                         &col_matrix, T(0.0));
C
chengduoZH 已提交
293

C
chengduoZH 已提交
294
          if (is_expand && data_dim == 2U) {
Q
QI JUN 已提交
295
            col2im(dev_ctx, col, dilations, strides,
C
chengduoZH 已提交
296 297 298
                   std::vector<int>{paddings[0], paddings[1], paddings[0],
                                    paddings[1]},
                   &in_grad_slice);
C
chengduoZH 已提交
299
          } else if (is_expand && data_dim == 3U) {
Q
QI JUN 已提交
300
            col2vol(dev_ctx, col, dilations, strides, paddings, &in_grad_slice);
C
chengduoZH 已提交
301
          }
C
chengduoZH 已提交
302 303 304 305 306 307 308 309
        }
      }
    }

    if (filter_grad) {
      filter_grad->mutable_data<T>(context.GetPlace());
      Tensor filter_grad_ = *filter_grad;
      filter_grad_.Resize(filter_matrix_shape);
Q
QI JUN 已提交
310 311 312
      set_zero(dev_ctx, filter_grad, static_cast<T>(0));
      math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
      math::Vol2ColFunctor<DeviceContext, T> vol2col;
C
chengduoZH 已提交
313 314 315 316 317 318 319 320 321
      for (int i = 0; i < batch_size; i++) {
        Tensor out_grad_batch =
            output_grad->Slice(i, i + 1).Resize(output_matrix_shape);
        Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
        for (int g = 0; g < groups; g++) {
          // im2col
          Tensor out_grad_slice =
              out_grad_batch.Slice(g * out_step, (g + 1) * out_step);
          Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
C
chengduoZH 已提交
322

C
chengduoZH 已提交
323
          if (!is_expand) {
C
chengduoZH 已提交
324 325 326
            col.ShareDataWith(in_slice);
            col_matrix.ShareDataWith(col);
            col_matrix.Resize(col_matrix_shape);
C
chengduoZH 已提交
327
          } else if (data_dim == 2U) {
Q
QI JUN 已提交
328
            im2col(dev_ctx, in_slice, dilations, strides,
C
chengduoZH 已提交
329 330 331
                   std::vector<int>{paddings[0], paddings[1], paddings[0],
                                    paddings[1]},
                   &col);
C
chengduoZH 已提交
332
          } else if (data_dim == 3U) {
Q
QI JUN 已提交
333
            vol2col(dev_ctx, in_slice, dilations, strides, paddings, &col);
C
chengduoZH 已提交
334
          }
C
chengduoZH 已提交
335 336 337 338

          // gemm
          Tensor filter_grad_slice =
              filter_grad_.Slice(g * out_step, (g + 1) * out_step);
Q
QI JUN 已提交
339 340 341
          math::matmul<DeviceContext, T>(dev_ctx, out_grad_slice, false,
                                         col_matrix, true, T(1.0),
                                         &filter_grad_slice, T(1.0));
C
chengduoZH 已提交
342 343 344 345 346
        }
      }
    }
  }
};
347 348
}  // namespace operators
}  // namespace paddle