test_hooks.cc 9.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
// Copyright (c) 2020 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 <memory>
#include <set>
#include <string>
#include <vector>

#include "glog/logging.h"
#include "gtest/gtest.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/imperative/basic_engine.h"
#include "paddle/fluid/imperative/hooks.h"
#include "paddle/fluid/imperative/tracer.h"
#include "paddle/fluid/memory/memcpy.h"
27 28 29 30
#include "paddle/phi/core/kernel_registry.h"

PD_DECLARE_KERNEL(add, CPU, ALL_LAYOUT);
PD_DECLARE_KERNEL(add_grad, CPU, ALL_LAYOUT);
31 32
PD_DECLARE_KERNEL(matmul_with_flatten, CPU, ALL_LAYOUT);
PD_DECLARE_KERNEL(matmul_with_flatten_grad, CPU, ALL_LAYOUT);
33 34 35 36 37 38 39 40 41 42 43 44 45

namespace platform = paddle::platform;
namespace framework = paddle::framework;
namespace memory = paddle::memory;

DECLARE_bool(sort_sum_gradient);

namespace paddle {
namespace imperative {

using vb_vector = std::vector<std::shared_ptr<imperative::VarBase>>;
using var_pair = std::pair<std::string, vb_vector>;

46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69
std::shared_ptr<imperative::VariableWrapper> DoubleHook(
    const std::shared_ptr<imperative::VariableWrapper>& var) {
  // 1. create out var
  auto out_var = std::make_shared<imperative::VariableWrapper>(var->Name());
  out_var->SetType(var->Type());
  out_var->SetDataType(var->DataType());
  out_var->SetForwardDataType(var->ForwardDataType());
  out_var->InnerSetOverridedStopGradient(var->InnerOverridedStopGradient());

  // 2. get input and output var's tensor
  auto* out_tensor = out_var->MutableVar()->GetMutable<framework::LoDTensor>();
  auto& tensor = var->Var().Get<framework::LoDTensor>();
  out_tensor->Resize(tensor.dims());

  // 3. double calc
  auto* data = tensor.data<float>();
  auto* out_data = out_tensor->mutable_data<float>(platform::CPUPlace());
  for (int64_t i = 0; i < out_tensor->numel(); ++i) {
    out_data[i] = data[i] * 2.0;
  }

  return out_var;
}

70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86
TEST(TestHooks, TestGradVarLeafBackwardHook) {
  // 1. prepare
  Tracer tracer;
  std::shared_ptr<VarBase> x(new VarBase(true, "x"));
  std::shared_ptr<VarBase> y(new VarBase(true, "y"));
  std::shared_ptr<VarBase> out(new VarBase(true, "out"));
  x->SetOverridedStopGradient(false);
  y->SetOverridedStopGradient(false);

  platform::CPUPlace place;
  std::vector<float> src_data(10, 2.0);
  std::vector<int64_t> x_dims = {2, 5};
  std::vector<int64_t> y_dims = {5, 2};

  auto* x_tensor = x->MutableVar()->GetMutable<framework::LoDTensor>();
  auto* y_tensor = y->MutableVar()->GetMutable<framework::LoDTensor>();

87
  x_tensor->Resize(phi::make_ddim(x_dims));
88 89 90 91
  auto* mutable_x = x_tensor->mutable_data<float>(place);
  memory::Copy(place, mutable_x, place, src_data.data(),
               sizeof(float) * src_data.size());

92
  y_tensor->Resize(phi::make_ddim(y_dims));
93 94 95 96 97 98 99 100 101 102 103 104 105
  auto* mutable_y = y_tensor->mutable_data<float>(place);
  memory::Copy(place, mutable_y, place, src_data.data(),
               sizeof(float) * src_data.size());

  var_pair x_pair = var_pair("X", vb_vector(1, x));
  var_pair y_pair = var_pair("Y", vb_vector(1, y));
  var_pair out_pair = var_pair("Out", vb_vector(1, out));

  NameVarBaseMap ins = {x_pair, y_pair};
  NameVarBaseMap outs = {out_pair};
  framework::AttributeMap mul_attr_map;
  mul_attr_map["use_mkldnn"] = false;

106 107 108 109 110 111 112 113
  // add VariableWrapper hook
  x->GradVarBase()->AddVariableWrapperHook(
      std::make_shared<imperative::CppVariableWrapperHook>(DoubleHook));

  // add Void hook
  int64_t hook_value = 0;
  x->GradVarBase()->AddVoidHook(
      std::make_shared<std::function<void()>>([&]() { hook_value = 10; }));
114 115

  // 2. forward
J
Jiabin Yang 已提交
116
  tracer.TraceOp<VarBase>("mul", ins, outs, mul_attr_map, place, true);
117 118 119 120 121 122

  ASSERT_EQ(x->GradVarBase()->GradOpNum(), 0UL);
  ASSERT_EQ(y->GradVarBase()->GradOpNum(), 0UL);
  ASSERT_EQ(out->GradVarBase()->GradOpNum(), 1UL);

  // 3. backward
123 124
  std::vector<std::shared_ptr<imperative::VarBase>> tensors{out};
  std::vector<std::shared_ptr<imperative::VarBase>> grad_tensors{nullptr};
125
  BasicEngine engine;
126
  engine.Init(tensors, grad_tensors);
127 128
  engine.Execute();

129
  // verify VariableWrapper hook result
130 131 132 133 134 135
  framework::LoDTensor x_grad;
  framework::TensorCopySync(x->GradVar().Get<framework::LoDTensor>(), place,
                            &x_grad);
  for (int i = 0; i < x_grad.numel(); ++i) {
    ASSERT_EQ(x_grad.data<float>()[i], 8.0);
  }
136 137
  // verify Void hook result
  ASSERT_EQ(hook_value, 10);
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170

  framework::LoDTensor y_grad;
  framework::TensorCopySync(y->GradVar().Get<framework::LoDTensor>(), place,
                            &y_grad);

  for (int i = 0; i < y_grad.numel(); ++i) {
    ASSERT_EQ(y_grad.data<float>()[i], 4.0);
  }
}

void GradVarLeafBackwardHookWithGradAccmulatedTest() {
  // 1. prepare
  Tracer tracer;
  std::shared_ptr<VarBase> x(new VarBase(true, "x"));
  std::shared_ptr<VarBase> y(new VarBase(true, "y"));
  std::shared_ptr<VarBase> z(new VarBase(true, "z"));
  std::shared_ptr<VarBase> out_xy(new VarBase(true, "out_xy"));
  std::shared_ptr<VarBase> out_xz(new VarBase(true, "out_xz"));
  std::shared_ptr<VarBase> out(new VarBase(true, "out"));
  x->SetOverridedStopGradient(false);
  y->SetOverridedStopGradient(false);
  z->SetOverridedStopGradient(false);

  platform::CPUPlace place;
  std::vector<float> src_data(10, 2.0);
  std::vector<int64_t> x_dims = {2, 5};
  std::vector<int64_t> y_dims = {5, 2};
  std::vector<int64_t> z_dims = {5, 2};

  auto* x_tensor = x->MutableVar()->GetMutable<framework::LoDTensor>();
  auto* y_tensor = y->MutableVar()->GetMutable<framework::LoDTensor>();
  auto* z_tensor = z->MutableVar()->GetMutable<framework::LoDTensor>();

171
  x_tensor->Resize(phi::make_ddim(x_dims));
172 173 174 175
  auto* mutable_x = x_tensor->mutable_data<float>(place);
  memory::Copy(place, mutable_x, place, src_data.data(),
               sizeof(float) * src_data.size());

176
  y_tensor->Resize(phi::make_ddim(y_dims));
177 178 179 180
  auto* mutable_y = y_tensor->mutable_data<float>(place);
  memory::Copy(place, mutable_y, place, src_data.data(),
               sizeof(float) * src_data.size());

181
  z_tensor->Resize(phi::make_ddim(z_dims));
182 183 184 185
  auto* mutable_z = z_tensor->mutable_data<float>(place);
  memory::Copy(place, mutable_z, place, src_data.data(),
               sizeof(float) * src_data.size());

186 187 188 189 190 191 192 193
  // add VariableWrapper hook
  x->GradVarBase()->AddVariableWrapperHook(
      std::make_shared<imperative::CppVariableWrapperHook>(DoubleHook));

  // add Void hook
  int64_t hook_value = 0;
  x->GradVarBase()->AddVoidHook(
      std::make_shared<std::function<void()>>([&]() { hook_value = 100; }));
194 195 196 197 198 199 200 201 202

  // 2. forward
  var_pair x_pair = var_pair("X", vb_vector(1, x));
  var_pair y_pair = var_pair("Y", vb_vector(1, y));
  var_pair out_xy_pair = var_pair("Out", vb_vector(1, out_xy));
  NameVarBaseMap ins = {x_pair, y_pair};
  NameVarBaseMap outs = {out_xy_pair};
  framework::AttributeMap mul_attr_map;
  mul_attr_map["use_mkldnn"] = false;
J
Jiabin Yang 已提交
203
  tracer.TraceOp<VarBase>("mul", ins, outs, mul_attr_map, place, true);
204 205 206 207 208

  var_pair z_pair = var_pair("Y", vb_vector(1, z));
  var_pair out_xz_pair = var_pair("Out", vb_vector(1, out_xz));
  ins = {x_pair, z_pair};
  outs = {out_xz_pair};
J
Jiabin Yang 已提交
209
  tracer.TraceOp<VarBase>("mul", ins, outs, mul_attr_map, place, true);
210 211 212 213 214 215 216

  var_pair xy_pair = var_pair("X", vb_vector(1, out_xy));
  var_pair xz_pair = var_pair("Y", vb_vector(1, out_xz));
  var_pair out_pair = var_pair("Out", vb_vector(1, out));
  ins = {xy_pair, xz_pair};
  outs = {out_pair};
  framework::AttributeMap add_attr_map;
J
Jiabin Yang 已提交
217 218
  tracer.TraceOp<VarBase>("elementwise_add", ins, outs, add_attr_map, place,
                          true);
219 220 221 222 223 224 225

  ASSERT_EQ(x->GradVarBase()->GradOpNum(), 0UL);
  ASSERT_EQ(y->GradVarBase()->GradOpNum(), 0UL);
  ASSERT_EQ(z->GradVarBase()->GradOpNum(), 0UL);
  ASSERT_EQ(out->GradVarBase()->GradOpNum(), 1UL);

  // 3. backward
226 227
  std::vector<std::shared_ptr<imperative::VarBase>> tensors{out};
  std::vector<std::shared_ptr<imperative::VarBase>> grad_tensors{nullptr};
228
  BasicEngine engine;
229
  engine.Init(tensors, grad_tensors);
230 231
  engine.Execute();

232
  // verify VariableWrapper hook result
233 234 235 236 237 238
  framework::LoDTensor x_grad;
  framework::TensorCopySync(x->GradVar().Get<framework::LoDTensor>(), place,
                            &x_grad);
  for (int i = 0; i < x_grad.numel(); ++i) {
    ASSERT_EQ(x_grad.data<float>()[i], 16.0);
  }
239 240
  // verify Void hook result
  ASSERT_EQ(hook_value, 100);
241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271

  framework::LoDTensor y_grad;
  framework::TensorCopySync(y->GradVar().Get<framework::LoDTensor>(), place,
                            &y_grad);

  for (int i = 0; i < y_grad.numel(); ++i) {
    ASSERT_EQ(y_grad.data<float>()[i], 4.0);
  }

  framework::LoDTensor z_grad;
  framework::TensorCopySync(z->GradVar().Get<framework::LoDTensor>(), place,
                            &z_grad);

  for (int i = 0; i < z_grad.numel(); ++i) {
    ASSERT_EQ(z_grad.data<float>()[i], 4.0);
  }
}

TEST(TestHooks, TestGradVarLeafBackwardHookWithGradAccmulated) {
  GradVarLeafBackwardHookWithGradAccmulatedTest();
}

TEST(TestHooks, TestGradVarLeafBackwardHookWithSortedGradAccmulated) {
  FLAGS_sort_sum_gradient = true;
  GradVarLeafBackwardHookWithGradAccmulatedTest();
  FLAGS_sort_sum_gradient = false;
}

}  // namespace imperative
}  // namespace paddle

272 273
USE_OP_ITSELF(mul);
USE_OP_ITSELF(mul_grad);
274 275
USE_OP_ITSELF(elementwise_add);
USE_OP_ITSELF(elementwise_add_grad);