layer.h 7.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
// 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.

#pragma once

X
Xin Pan 已提交
17 18 19 20 21 22 23
// clang-format off
#include "paddle/fluid/framework/python_headers.h"
// clang-format on

#include <map>     // NOLINT
#include <string>  // NOLINT
#include <vector>  // NOLINT
M
minqiyang 已提交
24
#include <memory>  // NOLINT
M
minqiyang 已提交
25

26 27 28 29
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/platform/enforce.h"
M
minqiyang 已提交
30
#include "paddle/fluid/platform/device_context.h"
31

M
minqiyang 已提交
32 33
#include "paddle/fluid/imperative/type_defs.h"

34 35 36
namespace paddle {
namespace imperative {

M
minqiyang 已提交
37 38
class VarBase;

X
Xin Pan 已提交
39 40
namespace py = ::pybind11;

X
Xin Pan 已提交
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
class PreparedOp {
 public:
  PreparedOp(const framework::OperatorBase& op,
             const framework::RuntimeContext& ctx,
             framework::OperatorWithKernel::OpKernelFunc func,
             platform::DeviceContext* dev_ctx)
      : op(op), ctx(ctx), func(func), dev_ctx(dev_ctx) {}

  static PreparedOp Prepare(const framework::RuntimeContext& ctx,
                            const framework::OperatorWithKernel& op,
                            const platform::Place& place) {
    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
    auto* dev_ctx = pool.Get(place);

    // check if op[type] has kernel registered.
    auto& all_op_kernels = op.AllOpKernels();
    auto kernels_iter = all_op_kernels.find(op.Type());
    if (kernels_iter == all_op_kernels.end()) {
      PADDLE_THROW(
          "There are no kernels which are registered in the %s operator.",
          op.Type());
    }

    framework::OperatorWithKernel::OpKernelMap& kernels = kernels_iter->second;

    auto expected_kernel_key = op.GetExpectedKernelType(
X
Xin Pan 已提交
67
        framework::ExecutionContext(op, framework::Scope(), *dev_ctx, ctx));
X
Xin Pan 已提交
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
    VLOG(3) << "expected_kernel_key:" << expected_kernel_key;

    auto kernel_iter = kernels.find(expected_kernel_key);
#ifdef PADDLE_WITH_MKLDNN
    // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
    if (kernel_iter == kernels.end() &&
        expected_kernel_key.library_type_ == framework::LibraryType::kMKLDNN) {
      VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
      expected_kernel_key.library_type_ = framework::LibraryType::kPlain;
      expected_kernel_key.data_layout_ = framework::DataLayout::kAnyLayout;
      kernel_iter = kernels.find(expected_kernel_key);
    }
#endif
    if (kernel_iter == kernels.end()) {
      PADDLE_THROW("op %s does not have kernel for %s", op.Type(),
                   KernelTypeToString(expected_kernel_key));
    }
    return PreparedOp(op, ctx, kernel_iter->second, dev_ctx);
  }

M
minqiyang 已提交
88 89
  inline platform::DeviceContext* GetDeviceContext() const { return dev_ctx; }

X
Xin Pan 已提交
90 91 92 93 94
  const framework::OperatorBase& op;
  const framework::RuntimeContext& ctx;
  framework::OperatorWithKernel::OpKernelFunc func;
  platform::DeviceContext* dev_ctx;
};
X
polish  
Xin Pan 已提交
95

96 97
class OpBase;

M
minqiyang 已提交
98 99 100 101 102
/* The wrapper for Variable which holds a Variable and a VarBase of its
 * gradient. This object should be managed totally by Python intepreter.
 *
 * Nearly all interface should be implemented in C++.
 */
103 104
class VarBase {
 public:
105
  VarBase() : VarBase(new framework::Variable(), new VarBase(true)) {}
X
polish  
Xin Pan 已提交
106 107

  // Owns `var` and `grad`
108
  VarBase(framework::Variable* var, VarBase* grad)
X
Xin Pan 已提交
109
      : var_desc_(nullptr),
X
polish  
Xin Pan 已提交
110 111
        var_(var),
        grads_(grad),
X
Xin Pan 已提交
112 113 114
        stop_gradient_(false),
        pre_op_(nullptr),
        pre_op_out_idx_(-1) {}
M
minqiyang 已提交
115 116

  explicit VarBase(bool stop_gradient)
X
Xin Pan 已提交
117
      : var_desc_(nullptr),
X
Xin Pan 已提交
118
        var_(new framework::Variable()),
M
minqiyang 已提交
119
        grads_(stop_gradient ? nullptr : new VarBase(true)),
X
Xin Pan 已提交
120 121 122
        stop_gradient_(stop_gradient),
        pre_op_(nullptr),
        pre_op_out_idx_(-1) {}
123

M
minqiyang 已提交
124 125 126 127 128 129 130 131 132
  virtual ~VarBase() {
    if (var_) {
      delete var_;
    }

    if (grads_) {
      delete grads_;
    }
  }
133

X
Xin Pan 已提交
134 135 136 137 138 139
  OpBase* PreOp() const { return pre_op_; }
  int PreOpOutIdx() const { return pre_op_out_idx_; }

  void SetStopGradient(bool stop_gradient) { stop_gradient_ = stop_gradient; }
  bool IsStopGradient() const { return stop_gradient_; }

X
Xin Pan 已提交
140
  void RunBackward();
141

X
Xin Pan 已提交
142 143 144 145 146 147 148 149 150
  void TrackPreOp(OpBase* pre_op, const std::string& pre_op_out_name,
                  int pre_op_out_idx, bool stop_gradient) {
    pre_op_ = pre_op;
    pre_op_out_name_ = pre_op_out_name;
    pre_op_out_idx_ = pre_op_out_idx;
    stop_gradient_ = stop_gradient;
  }

  void ClearGradient() {
X
Xin Pan 已提交
151 152 153 154
    delete grads_;
    grads_ = new VarBase(true);
  }

M
minqiyang 已提交
155
  framework::LoDTensor& GradValue();
156

M
minqiyang 已提交
157 158
  std::unique_ptr<VarBase> NewVarBase(const platform::Place& dst_place,
                                      const bool blocking) const;
M
minqiyang 已提交
159

M
minqiyang 已提交
160 161 162 163 164 165 166
  inline std::string GradName() const {
    PADDLE_ENFORCE(
        var_desc_,
        "Couldn't get gradient variable's name, please call backward() first");
    return string::Sprintf("%s@IGrad", var_desc_->Name());
  }

167
  framework::VarDesc* var_desc_;
M
minqiyang 已提交
168

M
minqiyang 已提交
169 170
  framework::Variable* var_;
  VarBase* grads_;
171

X
Xin Pan 已提交
172
 private:
173
  bool stop_gradient_;
X
Xin Pan 已提交
174 175 176
  OpBase* pre_op_;
  std::string pre_op_out_name_;
  int pre_op_out_idx_;
177 178
};

M
minqiyang 已提交
179 180 181
/* The wrapper for OpDesc which holds a OpDesc and a OpDesc of its
 * gradient. This object should be managed totally by Python intepreter.
 */
182 183
class OpBase {
 public:
X
Xin Pan 已提交
184 185 186
  OpBase()
      : op_desc_(nullptr),
        forward_id_(-1),
M
minqiyang 已提交
187
        backward_id_(-1),
P
Paddle CI 已提交
188
        place_(platform::CPUPlace()) {}
189 190

  virtual ~OpBase() {
X
Xin Pan 已提交
191 192 193
    for (framework::OpDesc* desc : grad_op_descs_) {
      delete desc;
    }
194 195
  }

X
Xin Pan 已提交
196
  std::map<std::string, std::vector<VarBase*>> ApplyGrad();
197

X
polish  
Xin Pan 已提交
198 199
  // One of `op_desc_` or `forward_id_` is set, not both.
  // For pure python PyLayer, use `forward_id_`, otherwise, use op_desc_.
200
  framework::OpDesc* op_desc_;
X
Xin Pan 已提交
201
  int forward_id_;
X
polish  
Xin Pan 已提交
202

X
Xin Pan 已提交
203
  // When has backward, one of `grad_op_descs_` or `backward_id_` is set,
X
polish  
Xin Pan 已提交
204
  // not both.
X
polish  
Xin Pan 已提交
205
  // Note: each fwd op corresponds to a vector of bwd ops.
X
Xin Pan 已提交
206
  std::vector<framework::OpDesc*> grad_op_descs_;
X
Xin Pan 已提交
207
  int backward_id_;
X
Xin Pan 已提交
208

P
Paddle CI 已提交
209
  platform::Place place_;
M
minqiyang 已提交
210

M
minqiyang 已提交
211 212 213
  VarBasePtrMap input_vars_;
  VarBasePtrMap output_vars_;
  OpBasePtrMap pre_ops_;
X
Xin Pan 已提交
214
  std::map<std::string, std::vector<int>> pre_ops_out_idx_;
215

X
polish  
Xin Pan 已提交
216
  // Inputs to a vector of bwd ops.
X
Xin Pan 已提交
217
  std::vector<framework::VariableValueMap> grad_input_vars_;
X
polish  
Xin Pan 已提交
218
  // Outputs to a vector of bwd ops.
X
Xin Pan 已提交
219
  std::vector<framework::VariableValueMap> grad_output_vars_;
X
polish  
Xin Pan 已提交
220

221 222 223 224 225 226 227 228 229 230 231
  framework::BlockDesc* block_;
};

class Layer {
 public:
  virtual ~Layer() {}

  virtual std::vector<VarBase> Forward(const std::vector<VarBase>& inputs) {
    std::vector<VarBase> vars;
    return vars;
  }
X
Xin Pan 已提交
232
};
233

X
Xin Pan 已提交
234 235 236 237
class PyLayer {
 public:
  virtual ~PyLayer() {}

X
polish  
Xin Pan 已提交
238 239
  static const char* kFwdInp;
  static const char* kFwdOut;
X
Xin Pan 已提交
240

X
Xin Pan 已提交
241
  static void RegisterFunc(int func_id, const py::object& py_func);
X
Xin Pan 已提交
242

X
polish  
Xin Pan 已提交
243 244
  static int NumFuncs();

X
Xin Pan 已提交
245
  static std::vector<VarBase*> Apply(int func_id,
X
Xin Pan 已提交
246 247
                                     const std::vector<VarBase*>& inputs);

X
polish  
Xin Pan 已提交
248 249
  static std::vector<framework::Variable*> ApplyGrad(
      int func_id, const std::vector<framework::Variable*>& inputs);
250

X
polish  
Xin Pan 已提交
251 252 253
 private:
  static std::vector<framework::Variable*> CallPythonFunc(
      const py::object& callable, const std::vector<framework::Variable*>& ins);
254 255 256 257
};

}  // namespace imperative
}  // namespace paddle