test_conv_op.cpp 14.4 KB
Newer Older
H
hjchen2 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2018 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. */

H
hjchen2 已提交
15
#include <iostream>
H
hjchen2 已提交
16 17 18 19 20 21
#include "../test_helper.h"
#include "../test_include.h"
#include "operators/conv_op.h"

namespace paddle_mobile {

H
hjchen2 已提交
22
// Reference convolution from Caffe for checking results.
H
hjchen2 已提交
23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132
// accumulate through explicit loops over input, output, and filters.
template <typename Itype, typename Otype>
void conv2d(const framework::Tensor *input, const framework::Tensor *filter,
            const framework::AttributeMap &attrs, framework::Tensor *output) {
  framework::AttrReader attr_reader(attrs);
  std::vector<int> paddings = attr_reader.Get<std::vector<int>>("paddings");
  std::vector<int> strides = attr_reader.Get<std::vector<int>>("strides");
  std::vector<int> dilations = attr_reader.Get<std::vector<int>>("dilations");
  int groups = attr_reader.Get<int>("groups");
  int kernel_h = filter->dims()[2];
  int kernel_w = filter->dims()[3];
  int pad_h = paddings[0];
  int pad_w = paddings[1];
  int stride_h = strides[0];
  int stride_w = strides[1];
  int dilation_h = dilations[0];
  int dilation_w = dilations[1];
  auto in_shape = input->dims();
  auto out_shape = output->dims();

  const bool has_depth = 0;
  int kernel_d, pad_d, stride_d, dilation_d;
  if (has_depth) {
    kernel_d = kernel_h;
    stride_d = stride_h;
    pad_d = pad_h;
    dilation_d = dilation_h;
  } else {
    kernel_d = stride_d = dilation_d = 1;
    pad_d = 0;
  }
  // Groups
  int o_g = out_shape[1] / groups;
  int k_g = in_shape[1] / groups;
  int o_head, k_head;
  // Convolution
  vector<int> weight_offset(4 + has_depth);
  vector<int> in_offset(4 + has_depth);
  vector<int> out_offset(4 + has_depth);
  auto offset = [](const framework::Tensor *input, const vector<int> &indics) {
    framework::DDim shape = input->dims();
    size_t count = 0;
    for (int i = 0; i < indics.size(); ++i) {
      count *= shape[i];
      count += indics[i];
    }
    return count;
  };

  const Itype *in_data = input->data<Itype>();
  const Itype *w_data = filter->data<Itype>();
  Otype *out_data = output->mutable_data<Otype>();
  memset(out_data, 0, output->numel() * sizeof(Otype));
  for (int n = 0; n < out_shape[0]; n++) {
    for (int g = 0; g < groups; g++) {
      o_head = o_g * g;
      k_head = k_g * g;
      for (int o = 0; o < o_g; o++) {
        for (int k = 0; k < k_g; k++) {
          for (int z = 0; z < (has_depth ? out_shape[2] : 1); z++) {
            for (int y = 0; y < out_shape[2 + has_depth]; y++) {
              for (int x = 0; x < out_shape[3 + has_depth]; x++) {
                for (int r = 0; r < kernel_d; r++) {
                  for (int p = 0; p < kernel_h; p++) {
                    for (int q = 0; q < kernel_w; q++) {
                      int in_z = z * stride_d - pad_d + r * dilation_d;
                      int in_y = y * stride_h - pad_h + p * dilation_h;
                      int in_x = x * stride_w - pad_w + q * dilation_w;
                      if (in_z >= 0 && in_z < (has_depth ? in_shape[2] : 1) &&
                          in_y >= 0 && in_y < in_shape[2 + has_depth] &&
                          in_x >= 0 && in_x < in_shape[3 + has_depth]) {
                        weight_offset[0] = o + o_head;
                        weight_offset[1] = k;
                        if (has_depth) {
                          weight_offset[2] = r;
                        }
                        weight_offset[2 + has_depth] = p;
                        weight_offset[3 + has_depth] = q;
                        in_offset[0] = n;
                        in_offset[1] = k + k_head;
                        if (has_depth) {
                          in_offset[2] = in_z;
                        }
                        in_offset[2 + has_depth] = in_y;
                        in_offset[3 + has_depth] = in_x;
                        out_offset[0] = n;
                        out_offset[1] = o + o_head;
                        if (has_depth) {
                          out_offset[2] = z;
                        }
                        out_offset[2 + has_depth] = y;
                        out_offset[3 + has_depth] = x;

                        out_data[offset(output, out_offset)] +=
                            in_data[offset(input, in_offset)] *
                            w_data[offset(filter, weight_offset)];
                      }
                    }
                  }
                }
              }
            }
          }
        }
      }
    }
  }
}

template <typename Itype, typename Otype, int Kernel, int Pad, int Stride>
133 134
int TestConvOp(int in_channels, int in_height, int in_width, int out_channels,
               int groups) {
H
hjchen2 已提交
135 136 137 138 139 140 141 142 143
  int kernel_h = Kernel;
  int kernel_w = Kernel;
  int pad_h = Pad;
  int pad_w = Pad;
  int stride_h = Stride;
  int stride_w = Stride;
  int dilation_h = 1;
  int dilation_w = 1;

H
hjchen2 已提交
144
  int batch_size = 1;
H
hjchen2 已提交
145 146 147 148
  int input_c = in_channels;
  int input_h = in_height;
  int input_w = in_width;
  int output_c = out_channels;
H
hjchen2 已提交
149 150 151
  framework::DDim input_shape =
      framework::make_ddim({batch_size, input_c, input_h, input_w});
  framework::DDim filter_shape =
152
      framework::make_ddim({output_c, input_c / groups, kernel_h, kernel_w});
H
hjchen2 已提交
153 154 155 156 157 158 159 160 161 162

  VariableNameMap inputs;
  VariableNameMap outputs;
  auto scope = std::make_shared<framework::Scope>();
  inputs["Input"] = std::vector<std::string>({"input"});
  inputs["Filter"] = std::vector<std::string>({"filter"});
  outputs["Output"] = std::vector<std::string>({"output"});

  auto input_var = scope.get()->Var("input");
  auto input = input_var->template GetMutable<framework::LoDTensor>();
H
hjchen2 已提交
163
  SetupTensor<Itype>(input, input_shape, -20.0, 20.0);
H
hjchen2 已提交
164 165 166

  auto filter_var = scope.get()->Var("filter");
  auto filter = filter_var->template GetMutable<framework::LoDTensor>();
167
  SetupTensor<Itype>(filter, filter_shape, -20, 20);
H
hjchen2 已提交
168

169 170 171 172 173 174
  //  for (int i = 0; i < input->numel(); ++i) {
  //    DLOG << "input[" << i << "] = " << float(input->data<Itype>()[i]);
  //  }
  //  for (int i = 0; i < filter->numel(); ++i) {
  //    DLOG << "filter[" << i << "] = " << float(filter->data<Itype>()[i]);
  //  }
175

H
hjchen2 已提交
176 177 178 179 180 181
  auto output_var = scope.get()->Var("output");
  framework::AttributeMap attrs;
  attrs["strides"].Set<vector<int>>(std::vector<int>({stride_h, stride_w}));
  attrs["paddings"].Set<vector<int>>(std::vector<int>({pad_h, pad_w}));
  attrs["dilations"].Set<vector<int>>(
      std::vector<int>({dilation_h, dilation_w}));
182
  attrs["groups"].Set<int>(groups);
H
hjchen2 已提交
183 184

  auto *op = new operators::ConvOp<CPU, float>("conv2d", inputs, outputs, attrs,
185
                                               scope.get());
H
hjchen2 已提交
186
  op->InferShape();
H
hjchen2 已提交
187 188
  op->Init();
  //  struct timespec ts_begin, ts_end;
189
  // warmup
190 191 192
  //  op->Run();
  //  clock_gettime(CLOCK_MONOTONIC, &ts_begin);
  //  for (int i = 0; i < 10; ++i) {
H
hjchen2 已提交
193
  op->Run();
194 195 196 197 198 199
  //  }
  //  clock_gettime(CLOCK_MONOTONIC, &ts_end);
  //  uint64_t elapsed = (ts_end.tv_sec - ts_begin.tv_sec) * 1e3 +
  //                     (ts_end.tv_nsec - ts_begin.tv_nsec) / 1e6;
  //  LOG(kLOG_INFO) << "elapsed: " << elapsed / 10.0 << " ms";

200 201
  // compare results
  auto *output = output_var->template Get<framework::LoDTensor>();
202
  framework::Tensor output_cmp;
203
  output_cmp.mutable_data<Otype>(output->dims());
204 205 206 207 208
  conv2d<Itype, Otype>(input, filter, attrs, &output_cmp);

  const Otype *output_data = output->data<Otype>();
  Otype *output_cmp_data = output_cmp.data<Otype>();
  for (int i = 0; i < output->numel(); ++i) {
H
hjchen2 已提交
209
    float gap = output_data[i] - output_cmp_data[i];
210 211 212
    //    PADDLE_MOBILE_ENFORCE(std::abs(gap / (output_data[i] + 1e-5)) < 1e-3,
    //                          "output[%d] = %d, output_cmp[%d] = %d", i,
    //                          output_data[i], i, output_cmp_data[i]);
H
hjchen2 已提交
213 214 215 216
    if (gap > 1e-2 && std::abs(gap / (output_data[i] + 1e-5)) > 1e-3) {
      std::cerr << "output_data[" << i << "] = " << output_data[i]
                << ", output_cmp_data[" << i << "] = " << output_cmp_data[i]
                << std::endl;
217 218
      exit(1);
    }
H
hjchen2 已提交
219 220 221 222 223 224 225
  }
  delete op;
  return 0;
}

}  // namespace paddle_mobile

226 227
int TestAll(const int in_channels, const int in_height, const int in_width,
            const int out_channels, const int groups) {
H
hjchen2 已提交
228 229 230
  std::cerr << "in_channels=" << in_channels << ", in_height=" << in_height
            << ", in_width=" << in_width << ", out_channels=" << out_channels
            << ", groups=" << groups << std::endl;
H
hjchen2 已提交
231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267
  std::cerr << "float, kernel=1, pad=0, stride=1" << std::endl;
  paddle_mobile::TestConvOp<float, float, 1, 0, 1>(
      in_channels, in_height, in_width, out_channels, groups);

  // kernel = 3, pad = 0, stride = 1
  std::cerr << "float, kernel=3, pad=0, stride=1" << std::endl;
  paddle_mobile::TestConvOp<float, float, 3, 0, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 1, stride = 1
  std::cerr << "float, kernel=3, pad=1, stride=1" << std::endl;
  paddle_mobile::TestConvOp<float, float, 3, 1, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 2, stride = 1
  std::cerr << "float, kernel=3, pad=2, stride=1" << std::endl;
  paddle_mobile::TestConvOp<float, float, 3, 2, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 5, stride = 1
  std::cerr << "float, kernel=3, pad=5, stride=1" << std::endl;
  paddle_mobile::TestConvOp<float, float, 3, 5, 1>(
      in_channels, in_height, in_width, out_channels, groups);

  // kernel = 3, pad = 0, stride = 2
  std::cerr << "float, kernel=3, pad=0, stride=2" << std::endl;
  paddle_mobile::TestConvOp<float, float, 3, 0, 2>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 1, stride = 2
  std::cerr << "float, kernel=3, pad=1, stride=2" << std::endl;
  paddle_mobile::TestConvOp<float, float, 3, 1, 2>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 2, stride = 2
  std::cerr << "float, kernel=3, pad=2, stride=2" << std::endl;
  paddle_mobile::TestConvOp<float, float, 3, 2, 2>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 5, stride = 2
  std::cerr << "float, kernel=3, pad=5, stride=2" << std::endl;
  paddle_mobile::TestConvOp<float, float, 3, 5, 2>(
      in_channels, in_height, in_width, out_channels, groups);
H
hjchen2 已提交
268 269 270 271

#ifndef __aarch64__
  // kernel = 3, pad = 0, stride = 1
  std::cerr << "int8, kernel=3, pad=0, stride=1" << std::endl;
272 273
  paddle_mobile::TestConvOp<int8_t, int32_t, 3, 0, 1>(
      in_channels, in_height, in_width, out_channels, groups);
274
  // kernel = 3, pad = 1, stride = 1
H
hjchen2 已提交
275
  std::cerr << "int8, kernel=3, pad=1, stride=1" << std::endl;
276 277 278
  paddle_mobile::TestConvOp<int8_t, int32_t, 3, 1, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 2, stride = 1
H
hjchen2 已提交
279
  std::cerr << "int8, kernel=3, pad=2, stride=1" << std::endl;
280 281 282
  paddle_mobile::TestConvOp<int8_t, int32_t, 3, 2, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 5, stride = 1
H
hjchen2 已提交
283
  std::cerr << "int8, kernel=3, pad=5, stride=1" << std::endl;
284 285 286 287
  paddle_mobile::TestConvOp<int8_t, int32_t, 3, 5, 1>(
      in_channels, in_height, in_width, out_channels, groups);

  // kernel = 3, pad = 0, stride = 2
H
hjchen2 已提交
288
  std::cerr << "int8, kernel=3, pad=0, stride=2" << std::endl;
289 290 291
  paddle_mobile::TestConvOp<int8_t, int32_t, 3, 0, 2>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 1, stride = 2
H
hjchen2 已提交
292
  std::cerr << "int8, kernel=3, pad=1, stride=2" << std::endl;
293 294 295
  paddle_mobile::TestConvOp<int8_t, int32_t, 3, 1, 2>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 2, stride = 2
H
hjchen2 已提交
296
  std::cerr << "int8, kernel=3, pad=2, stride=2" << std::endl;
297 298 299
  paddle_mobile::TestConvOp<int8_t, int32_t, 3, 2, 2>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 3, pad = 5, stride = 2
H
hjchen2 已提交
300
  std::cerr << "int8, kernel=3, pad=5, stride=2" << std::endl;
301 302
  paddle_mobile::TestConvOp<int8_t, int32_t, 3, 5, 2>(
      in_channels, in_height, in_width, out_channels, groups);
H
hjchen2 已提交
303
#endif  // __aarch64__
304

305
  // kernel = 5, pad = 0, stride = 1
H
hjchen2 已提交
306
  std::cerr << "float, kernel=5, pad=0, stride=1" << std::endl;
307 308 309
  paddle_mobile::TestConvOp<float, float, 5, 0, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 5, pad = 1, stride = 1
H
hjchen2 已提交
310
  std::cerr << "float, kernel=5, pad=1, stride=1" << std::endl;
311 312 313
  paddle_mobile::TestConvOp<float, float, 5, 1, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 5, pad = 2, stride = 1
H
hjchen2 已提交
314
  std::cerr << "float, kernel=5, pad=2, stride=1" << std::endl;
315 316 317
  paddle_mobile::TestConvOp<float, float, 5, 2, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 5, pad = 5, stride = 1
H
hjchen2 已提交
318
  std::cerr << "float, kernel=5, pad=5, stride=1" << std::endl;
319 320 321
  paddle_mobile::TestConvOp<float, float, 5, 5, 1>(
      in_channels, in_height, in_width, out_channels, groups);

H
hjchen2 已提交
322
#ifndef __aarch64__
323
  // kernel = 5, pad = 0, stride = 1
H
hjchen2 已提交
324
  std::cerr << "int8, kernel=5, pad=0, stride=1" << std::endl;
325 326 327
  paddle_mobile::TestConvOp<int8_t, int32_t, 5, 0, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 5, pad = 1, stride = 1
H
hjchen2 已提交
328
  std::cerr << "int8, kernel=5, pad=1, stride=1" << std::endl;
329 330 331
  paddle_mobile::TestConvOp<int8_t, int32_t, 5, 1, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 5, pad = 2, stride = 1
H
hjchen2 已提交
332
  std::cerr << "int8, kernel=5, pad=2, stride=1" << std::endl;
333 334 335
  paddle_mobile::TestConvOp<int8_t, int32_t, 5, 2, 1>(
      in_channels, in_height, in_width, out_channels, groups);
  // kernel = 5, pad = 5, stride = 1
H
hjchen2 已提交
336
  std::cerr << "int8, kernel=5, pad=5, stride=1" << std::endl;
337 338
  paddle_mobile::TestConvOp<int8_t, int32_t, 5, 5, 1>(
      in_channels, in_height, in_width, out_channels, groups);
H
hjchen2 已提交
339
#endif  // __aarch64__
340 341

  return 0;
342
}
343 344

int main() {
H
hjchen2 已提交
345
  TestAll(16, 10, 10, 16, 16);
346 347 348 349 350 351 352 353 354 355 356 357 358
  TestAll(1, 5, 5, 1, 1);
  TestAll(1, 5, 5, 10, 1);
  TestAll(10, 5, 5, 10, 10);

  TestAll(5, 33, 33, 5, 1);
  TestAll(5, 33, 33, 13, 1);
  TestAll(13, 33, 33, 13, 13);

  TestAll(5, 33, 13, 5, 1);
  TestAll(5, 33, 13, 13, 1);
  TestAll(13, 33, 13, 13, 13);
  return 0;
}