layer.h 8.0 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"
M
minqiyang 已提交
31
#include "paddle/fluid/operators/math/math_function.h"
32

M
minqiyang 已提交
33 34
#include "paddle/fluid/imperative/type_defs.h"

35 36 37
namespace paddle {
namespace imperative {

M
minqiyang 已提交
38 39
class VarBase;

X
Xin Pan 已提交
40 41
namespace py = ::pybind11;

X
Xin Pan 已提交
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
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
68
        framework::ExecutionContext(op, framework::Scope(), *dev_ctx, ctx));
X
Xin Pan 已提交
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
    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
89 90
  inline platform::DeviceContext* GetDeviceContext() const { return dev_ctx; }

X
Xin Pan 已提交
91 92 93 94 95
  const framework::OperatorBase& op;
  const framework::RuntimeContext& ctx;
  framework::OperatorWithKernel::OpKernelFunc func;
  platform::DeviceContext* dev_ctx;
};
X
polish  
Xin Pan 已提交
96

97 98
class OpBase;

M
minqiyang 已提交
99 100 101 102 103
/* 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++.
 */
104 105
class VarBase {
 public:
106
  VarBase() : VarBase(new framework::Variable(), new VarBase(true)) {}
X
polish  
Xin Pan 已提交
107 108

  // Owns `var` and `grad`
109
  VarBase(framework::Variable* var, VarBase* grad)
X
Xin Pan 已提交
110
      : var_desc_(nullptr),
X
polish  
Xin Pan 已提交
111 112
        var_(var),
        grads_(grad),
X
Xin Pan 已提交
113 114 115
        stop_gradient_(false),
        pre_op_(nullptr),
        pre_op_out_idx_(-1) {}
M
minqiyang 已提交
116 117

  explicit VarBase(bool stop_gradient)
X
Xin Pan 已提交
118
      : var_desc_(nullptr),
X
Xin Pan 已提交
119
        var_(new framework::Variable()),
M
minqiyang 已提交
120
        grads_(stop_gradient ? nullptr : new VarBase(true)),
X
Xin Pan 已提交
121 122 123
        stop_gradient_(stop_gradient),
        pre_op_(nullptr),
        pre_op_out_idx_(-1) {}
124

M
minqiyang 已提交
125 126 127 128 129 130 131 132 133
  virtual ~VarBase() {
    if (var_) {
      delete var_;
    }

    if (grads_) {
      delete grads_;
    }
  }
134

X
Xin Pan 已提交
135 136 137 138 139 140
  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
141
  void RunBackward();
142

X
Xin Pan 已提交
143
  void TrackPreOp(OpBase* pre_op, const std::string& pre_op_out_name,
M
minqiyang 已提交
144
                  int pre_op_out_idx, bool pre_op_stop_gradient) {
X
Xin Pan 已提交
145 146 147
    pre_op_ = pre_op;
    pre_op_out_name_ = pre_op_out_name;
    pre_op_out_idx_ = pre_op_out_idx;
M
minqiyang 已提交
148 149 150
    if (pre_op_stop_gradient) {
      stop_gradient_ = pre_op_stop_gradient;
    }
X
Xin Pan 已提交
151 152 153
  }

  void ClearGradient() {
M
minqiyang 已提交
154
    VLOG(1) << "clear gradient of " << var_desc_->Name();
M
minqiyang 已提交
155 156 157 158 159 160 161
    if (grads_ && grads_->var_ && grads_->var_->IsInitialized()) {
      auto grads_t = grads_->var_->GetMutable<framework::LoDTensor>();
      operators::math::set_constant(
          *(platform::DeviceContextPool::Instance().Get(
              grads_->var_->Get<framework::LoDTensor>().place())),
          grads_t, 0.0);
    }
X
Xin Pan 已提交
162 163
  }

M
minqiyang 已提交
164
  framework::LoDTensor& GradValue();
165

M
minqiyang 已提交
166 167
  std::unique_ptr<VarBase> NewVarBase(const platform::Place& dst_place,
                                      const bool blocking) const;
M
minqiyang 已提交
168

M
minqiyang 已提交
169 170 171 172 173 174 175
  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());
  }

176
  framework::VarDesc* var_desc_;
M
minqiyang 已提交
177

M
minqiyang 已提交
178 179
  framework::Variable* var_;
  VarBase* grads_;
180

X
Xin Pan 已提交
181
 private:
182
  bool stop_gradient_;
X
Xin Pan 已提交
183 184 185
  OpBase* pre_op_;
  std::string pre_op_out_name_;
  int pre_op_out_idx_;
186 187
};

M
minqiyang 已提交
188 189 190
/* The wrapper for OpDesc which holds a OpDesc and a OpDesc of its
 * gradient. This object should be managed totally by Python intepreter.
 */
191 192
class OpBase {
 public:
X
Xin Pan 已提交
193 194 195
  OpBase()
      : op_desc_(nullptr),
        forward_id_(-1),
M
minqiyang 已提交
196
        backward_id_(-1),
P
Paddle CI 已提交
197
        place_(platform::CPUPlace()) {}
198 199

  virtual ~OpBase() {
X
Xin Pan 已提交
200 201 202
    for (framework::OpDesc* desc : grad_op_descs_) {
      delete desc;
    }
203 204
  }

X
Xin Pan 已提交
205
  std::map<std::string, std::vector<VarBase*>> ApplyGrad();
206

X
polish  
Xin Pan 已提交
207 208
  // One of `op_desc_` or `forward_id_` is set, not both.
  // For pure python PyLayer, use `forward_id_`, otherwise, use op_desc_.
209
  framework::OpDesc* op_desc_;
X
Xin Pan 已提交
210
  int forward_id_;
X
polish  
Xin Pan 已提交
211

X
Xin Pan 已提交
212
  // When has backward, one of `grad_op_descs_` or `backward_id_` is set,
X
polish  
Xin Pan 已提交
213
  // not both.
X
polish  
Xin Pan 已提交
214
  // Note: each fwd op corresponds to a vector of bwd ops.
X
Xin Pan 已提交
215
  std::vector<framework::OpDesc*> grad_op_descs_;
X
Xin Pan 已提交
216
  int backward_id_;
X
Xin Pan 已提交
217

P
Paddle CI 已提交
218
  platform::Place place_;
M
minqiyang 已提交
219

M
minqiyang 已提交
220 221 222
  VarBasePtrMap input_vars_;
  VarBasePtrMap output_vars_;
  OpBasePtrMap pre_ops_;
X
Xin Pan 已提交
223
  std::map<std::string, std::vector<int>> pre_ops_out_idx_;
224

X
polish  
Xin Pan 已提交
225
  // Inputs to a vector of bwd ops.
X
Xin Pan 已提交
226
  std::vector<framework::VariableValueMap> grad_input_vars_;
X
polish  
Xin Pan 已提交
227
  // Outputs to a vector of bwd ops.
X
Xin Pan 已提交
228
  std::vector<framework::VariableValueMap> grad_output_vars_;
X
polish  
Xin Pan 已提交
229

230 231 232 233 234 235 236 237 238 239 240
  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
241
};
242

X
Xin Pan 已提交
243 244 245 246
class PyLayer {
 public:
  virtual ~PyLayer() {}

X
polish  
Xin Pan 已提交
247 248
  static const char* kFwdInp;
  static const char* kFwdOut;
X
Xin Pan 已提交
249

X
Xin Pan 已提交
250
  static void RegisterFunc(int func_id, const py::object& py_func);
X
Xin Pan 已提交
251

X
polish  
Xin Pan 已提交
252 253
  static int NumFuncs();

X
Xin Pan 已提交
254
  static std::vector<VarBase*> Apply(int func_id,
X
Xin Pan 已提交
255 256
                                     const std::vector<VarBase*>& inputs);

X
polish  
Xin Pan 已提交
257 258
  static std::vector<framework::Variable*> ApplyGrad(
      int func_id, const std::vector<framework::Variable*>& inputs);
259

X
polish  
Xin Pan 已提交
260 261 262
 private:
  static std::vector<framework::Variable*> CallPythonFunc(
      const py::object& callable, const std::vector<framework::Variable*>& ins);
263 264 265 266
};

}  // namespace imperative
}  // namespace paddle