backward.cc 21.5 KB
Newer Older
Y
Yu Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

15
#include "paddle/framework/backward.h"
Y
Yu Yang 已提交
16
#include "paddle/operators/net_op.h"
D
dongzhihong 已提交
17

F
fengjiayi 已提交
18
#include <deque>
D
dongzhihong 已提交
19
#include <list>
Y
Yu Yang 已提交
20
#include <memory>
Y
Yu Yang 已提交
21
#include <unordered_set>
Y
Yu Yang 已提交
22

F
fengjiayi 已提交
23
#include "paddle/framework/block_desc.h"
24
#include "paddle/framework/op_registry.h"
25
#include "paddle/operators/dynamic_recurrent_op.h"
Y
Yan Chunwei 已提交
26
#include "paddle/operators/net_op.h"
Y
Yu Yang 已提交
27 28 29 30

namespace paddle {
namespace framework {

Y
Yu Yang 已提交
31
static inline std::unique_ptr<OperatorBase> CreateGradOp(
32 33
    const OperatorBase& op, const std::unordered_set<std::string>& no_grad_set,
    std::unordered_map<std::string, std::string>* grad_to_var) {
Y
Yu Yang 已提交
34 35 36 37 38 39
  OpDescBind op_desc;
  op_desc.SetInputMap(op.Inputs());
  op_desc.SetOutputMap(op.Outputs());
  op_desc.SetType(op.Type());
  op_desc.SetAttrMap(op.Attrs());
  auto& info = OpInfoMap::Instance().Get(op.Type());
Y
Yu Yang 已提交
40
  auto grad_descs = info.GradOpMaker()(op_desc, no_grad_set, grad_to_var, {});
Y
Yu Yang 已提交
41 42
  std::vector<std::unique_ptr<OperatorBase>> grad_ops;
  grad_ops.reserve(grad_descs.size());
Y
Yu Yang 已提交
43 44 45
  std::transform(grad_descs.begin(), grad_descs.end(),
                 std::back_inserter(grad_ops),
                 [](const std::unique_ptr<OpDescBind>& grad_desc) {
Y
Yu Yang 已提交
46
                   return OpRegistry::CreateOp(*grad_desc);
Y
Yu Yang 已提交
47
                 });
Y
Yu Yang 已提交
48
  PADDLE_ENFORCE(!grad_ops.empty());
Y
Yu Yang 已提交
49 50 51 52 53 54 55
  if (grad_ops.size() == 1) {
    return std::move(grad_ops[0]);
  } else {
    auto net_op = new operators::NetOp();
    for (auto& grad_op : grad_ops) {
      net_op->AppendOp(std::move(grad_op));
    }
Y
Yu Yang 已提交
56
    net_op->CompleteAddOp();
Y
Yu Yang 已提交
57 58 59 60
    return std::unique_ptr<OperatorBase>(net_op);
  }
}

Y
Yu Yang 已提交
61
template <typename Map, typename T>
Q
qiaolongfei 已提交
62
static void ForEachVarName(const Map& names, T callback) {
Y
Yu Yang 已提交
63
  for (auto& name : names) {
Y
Yu Yang 已提交
64
    for (auto& n : name.second) {
65
      if (callback(n)) return;
Y
Yu Yang 已提交
66 67
    }
  }
Y
Yu Yang 已提交
68 69
}

Y
Yan Chunwei 已提交
70
// return whether all the names + suffixes in the set
Y
Yu Yang 已提交
71
static bool AllInSet(
Y
Yu Yang 已提交
72
    const std::map<std::string, std::vector<std::string>>& names,
Y
Yu Yang 已提交
73
    const std::string& suffix, const std::unordered_set<std::string>& set) {
74 75 76 77
  bool all_in_set = true;
  ForEachVarName(names, [&all_in_set, &set, &suffix](const std::string& n) {
    all_in_set = set.find(n + suffix) != set.end();
    return !all_in_set;
Y
Yu Yang 已提交
78
  });
79
  return all_in_set;
Y
Yu Yang 已提交
80 81
}

Y
Yu Yang 已提交
82 83
static std::unique_ptr<OperatorBase> NOP() {
  auto net_op = new operators::NetOp();
Q
qiaolongfei 已提交
84
  net_op->SetType("@NOP@");
Y
Yu Yang 已提交
85
  net_op->CompleteAddOp();
Y
Yu Yang 已提交
86
  return std::unique_ptr<OperatorBase>(net_op);
Y
Yu Yang 已提交
87 88
}

Y
Yan Chunwei 已提交
89
//  Get backward operator from a forward operator, a recursive implementation.
Y
Yu Yang 已提交
90 91 92
//
//  no_grad_names the gradient variable names without gradient calculating.
//
93 94 95
//  uniq_id is a unique index used inside recursively calling
//  BackwardRecursive. use `uid = uniq_id++;` to get the unique index, and
//  pass `uniq_id` through recursive calling.
Y
Yu Yang 已提交
96
//
Y
Yan Chunwei 已提交
97 98
//  returns The backward operator. In a simple situation, it may be a simple
//  operator, in a complex situation, it maybe a NetOp.
Y
Yu Yang 已提交
99 100
//
//  See Backward.h for details
Y
Yu Yang 已提交
101
static std::unique_ptr<OperatorBase> BackwardRecursive(
Y
Yu Yang 已提交
102
    const OperatorBase& forwardOp,
103 104 105
    std::unordered_set<std::string>& no_grad_names,
    std::unordered_map<std::string, std::string>* grad_to_var,
    size_t& uniq_id) {
Y
Yu Yang 已提交
106 107
  //  If all input gradients of forwarding operator do not need to calculate,
  //  just return an NOP. Not return null ptr because NOP does not take
Q
typo  
qiaolongfei 已提交
108
  //  too much time for calculation, but it is useful for simplifying logic.
109
  if (AllInSet(forwardOp.Inputs() /*names*/, kGradVarSuffix /*suffix*/,
Y
Yan Chunwei 已提交
110
               no_grad_names /*set*/)) {
Y
Yu Yang 已提交
111
    return NOP();
Y
Yu Yang 已提交
112 113
  }

114 115
  //  All output gradients of forwarding operator do not need to calculate.
  //  Then all input gradients cannot be computed at all, and we put them into
Y
Yu Yang 已提交
116
  //  `no_grad_names` set. Return an NOP.
Q
qiaolongfei 已提交
117
  if (AllInSet(forwardOp.Outputs() /*names*/, kGradVarSuffix /*suffix*/,
Y
Yan Chunwei 已提交
118
               no_grad_names /*set*/)) {
Q
qiaolongfei 已提交
119
    ForEachVarName(forwardOp.Inputs(),
Y
Yu Yang 已提交
120 121 122 123
                   [&no_grad_names](const std::string& name) -> bool {
                     no_grad_names.insert(GradVarName(name));
                     return false;
                   });
Y
Yu Yang 已提交
124
    return NOP();
Y
Yu Yang 已提交
125 126
  }

Y
Yu Yang 已提交
127
  // Returned gradient network
Y
Yu Yang 已提交
128
  auto net = std::unique_ptr<operators::NetOp>(new operators::NetOp());
Y
Yu Yang 已提交
129 130

  if (forwardOp.IsNetOp()) {
Y
Yu Yang 已提交
131
    // Because forwardOp is a net op, it can static_cast.
Y
Yan Chunwei 已提交
132
    auto& forwardNet = static_cast<const operators::NetOp&>(forwardOp);
Y
Yu Yang 已提交
133

134
    // Map from output gradient variable name to operator's indices in
Y
Yan Chunwei 已提交
135
    // backward net's ops_. That operator generates that variable.
Y
Yu Yang 已提交
136 137 138
    std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;

    size_t local_op_id = 0;
Y
Yan Chunwei 已提交
139
    // reversely travel forwardNet and collect all duplicate outputs.
Y
Yu Yang 已提交
140
    for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
Y
Yu Yang 已提交
141
         ++it, ++local_op_id) {
Y
Yu Yang 已提交
142
      auto& fwd = *it;
143
      auto bwd = BackwardRecursive(*fwd, no_grad_names, grad_to_var, uniq_id);
Q
qiaolongfei 已提交
144
      ForEachVarName(bwd->Outputs(),
Y
Yu Yang 已提交
145 146 147 148
                     [&dup_output_ops, local_op_id](const std::string& out) {
                       dup_output_ops[out].emplace_back(local_op_id);
                       return false;
                     });
Y
Yu Yang 已提交
149
      net->AppendOp(std::move(bwd));
D
dongzhihong 已提交
150
    }
Y
Yu Yang 已提交
151
    // Get unique ID for this method.
D
dongzhihong 已提交
152
    auto uid = uniq_id++;
D
dongzhihong 已提交
153
    // TODO(dzh): more comment
Y
Yan Chunwei 已提交
154 155 156 157 158
    // multiple operators which have the same output (y for example) may
    // overwrite the same y variable when backward, special operations are token
    // to handle this case. For each duplicate output, rename it to an alias
    // (original name with a offset), append an `add` op for its operator,
    // and finally sum all the alias variable to the final output variable y.
Y
Yu Yang 已提交
159
    using Pos = std::pair<size_t, std::unique_ptr<OperatorBase>>;
Y
Yu Yang 已提交
160
    std::list<Pos> insert_position;
D
dongzhihong 已提交
161
    for (auto& dup_output_op : dup_output_ops) {
D
dongzhihong 已提交
162
      const std::string& name = dup_output_op.first;
Q
qijun 已提交
163 164 165
      // duplicate @Empty@ don't need to be added
      if (name == kEmptyVarName) continue;

D
dongzhihong 已提交
166
      auto& dup_op = dup_output_op.second;
Y
Yan Chunwei 已提交
167
      // no duplicate output
D
dongzhihong 已提交
168 169
      if (dup_op.size() == 1) continue;

Y
Yan Chunwei 已提交
170 171
      // process the duplicate outputs
      std::vector<std::string> dup_outputs;
D
dongzhihong 已提交
172
      for (size_t i = 0; i < dup_op.size(); ++i) {
Y
Yan Chunwei 已提交
173
        // rename each duplicate output to an alias
D
dongzhihong 已提交
174
        auto op_offset = dup_op[i];
D
dongzhihong 已提交
175 176 177
        dup_outputs.push_back(name + "@RENAME@" + std::to_string(uid) + "@" +
                              std::to_string(i));
        net->ops_[op_offset]->Rename(name, dup_outputs.back());
D
dongzhihong 已提交
178
      }
179 180 181 182 183
      // collect all the offset for each alias,
      // insert a sum operator to add all aliases to output
      insert_position.push_back(
          {dup_op.back(), OpRegistry::CreateOp("sum", {{"X", dup_outputs}},
                                               {{"Out", {name}}}, {})});
D
dongzhihong 已提交
184
    }
Y
Yu Yang 已提交
185

186
    // make sure the inserted `sum` ops follow the BFS order.
Y
Yu Yang 已提交
187
    insert_position.sort(
D
dongzhihong 已提交
188
        [](const Pos& l, const Pos& r) { return l.first > r.first; });
Y
Yu Yang 已提交
189 190

    for (auto& pos : insert_position) {
Y
Yu Yang 已提交
191
      net->InsertOp(pos.first + 1, std::move(pos.second));
D
dongzhihong 已提交
192
    }
Y
Yu Yang 已提交
193
  } else {
194
    std::unique_ptr<OperatorBase> grad_op(
195
        CreateGradOp(forwardOp, no_grad_names, grad_to_var));
Y
Yu Yang 已提交
196

Y
Yu Yang 已提交
197 198
    ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
                                          const std::string& grad_input) {
199
      if (no_grad_names.count(grad_input)) {
Y
Yu Yang 已提交
200
        // +1 for \0
201
        std::string prefix = grad_input.substr(
Y
Yu Yang 已提交
202
            0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
Q
qiaolongfei 已提交
203
        grad_op->Rename(grad_input, prefix + kZeroVarSuffix);
Y
Yu Yang 已提交
204 205 206

        // If part of input gradient of that operator is not calculated, fill
        // zero variables to that input gradient.
D
dangqingqing 已提交
207 208
        net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", {{"X", {prefix}}},
                                           {{"Y", {grad_input}}}, {}));
209
      }
Y
Yu Yang 已提交
210 211 212
      return false;
    });

Q
qiaolongfei 已提交
213 214
    ForEachVarName(grad_op->Outputs(),
                   [&no_grad_names, &grad_op](const std::string& grad_output) {
Y
Yu Yang 已提交
215
                     if (no_grad_names.count(grad_output)) {
Q
qiaolongfei 已提交
216
                       grad_op->Rename(grad_output, kEmptyVarName);
Y
Yu Yang 已提交
217 218 219
                     }
                     return false;
                   });
Y
Yu Yang 已提交
220

Y
Yan Chunwei 已提交
221
    // process recurrent gradient op as a special operator.
Y
Yu Yang 已提交
222
    if (forwardOp.Type() == "dynamic_recurrent") {
223 224 225 226 227 228 229 230 231 232 233
      // NOTE clean up cycle call somewhere (RNN's stepnet constains itself),
      // or this will result in infinite loop.
      const auto& rnnop =
          *static_cast<const operators::DynamicRecurrentOp*>(&forwardOp);
      auto rnn_grad_op =
          static_cast<operators::DynamicRecurrentGradientOp*>(grad_op.get());
      const auto& stepnet_op =
          *static_cast<const OperatorBase*>(&rnnop.rnn.GetStepUnit());
      // create stepnet's gradient op
      rnn_grad_op->rnn.SetStepUnit(
          BackwardRecursive(stepnet_op, no_grad_names, grad_to_var, uniq_id));
Y
Yan Chunwei 已提交
234 235
    }

Y
Yu Yang 已提交
236 237 238
    if (net->ops_.empty()) {  // Current no aux op is added to network
      return grad_op;
    }
Y
Yu Yang 已提交
239
    net->AppendOp(std::move(grad_op));
Y
Yu Yang 已提交
240
  }
Q
qiaolongfei 已提交
241
  net->SetType("@GENERATED_BACKWARD@");
Y
Yu Yang 已提交
242
  net->CompleteAddOp();
Y
Yu Yang 已提交
243 244 245
  return std::unique_ptr<OperatorBase>(
      static_cast<OperatorBase*>(net.release()));
}
Y
Yu Yang 已提交
246

Y
Yu Yang 已提交
247
// See header for comments
Y
Yu Yang 已提交
248
std::unique_ptr<OperatorBase> Backward(
Y
Yu Yang 已提交
249
    const OperatorBase& forwardOp,
Y
Yu Yang 已提交
250 251
    const std::unordered_set<std::string>& no_grad_vars) {
  std::unordered_set<std::string> no_grad_names;
Q
qijun 已提交
252
  no_grad_names.reserve(no_grad_vars.size() + 1);
Y
Yu Yang 已提交
253

254
  no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
255

Y
Yu Yang 已提交
256
  for (auto& name : no_grad_vars) {
257
    no_grad_names.insert(name + kGradVarSuffix);
Y
Yu Yang 已提交
258
  }
Y
Yu Yang 已提交
259
  size_t uid = 0;
260 261
  std::unordered_map<std::string, std::string> grad_to_var;
  return BackwardRecursive(forwardOp, no_grad_names, &grad_to_var, uid);
Y
Yu Yang 已提交
262
}
Y
Yi Wang 已提交
263

F
fengjiayi 已提交
264 265 266 267 268 269 270 271 272
// ====================================  //

static bool AllGradInSet(const std::vector<std::string>& names,
                         const std::unordered_set<std::string>& set) {
  for (const std::string& name : names) {
    if (!set.count(GradVarName(name))) {
      return false;
    }
  }
Y
Yang Yang(Tony) 已提交
273 274 275 276 277 278 279 280 281 282 283 284 285
  if (VLOG_IS_ON(10)) {
    std::ostringstream sout;
    sout << "All input {";
    for (auto& name : names) {
      sout << name << ",";
    }
    sout << "} is in {";
    for (auto& name : set) {
      sout << name << ",";
    }
    sout << "}";
    VLOG(10) << sout.str();
  }
F
fengjiayi 已提交
286 287 288
  return true;
}

Y
Yu Yang 已提交
289 290 291 292 293 294 295 296 297
static std::string FwdName(const std::string& grad_name) {
  auto pos = grad_name.find("@GRAD");
  if (pos == std::string::npos) {
    return "";
  } else {
    return grad_name.substr(0, pos);
  }
}

Y
Yu Yang 已提交
298
static void CreateGradVarInBlock(
299 300 301 302
    size_t grad_op_start_index,
    const std::unordered_map<std::string, std::string>& param_name_map,
    BlockDescBind* block_desc,
    std::unordered_map<std::string, GradVarInfo>* grad_var_record) {
303 304 305
  auto ops = block_desc->AllOps();
  for (size_t op_index = grad_op_start_index; op_index < ops.size();
       ++op_index) {
Y
Yu Yang 已提交
306
    std::unordered_set<std::string> new_vars;
Y
Yu Yang 已提交
307 308 309 310 311
    ForEachVarName(ops[op_index]->Outputs(),
                   [&](const std::string& grad_var_name) {
                     if (block_desc->HasVar(grad_var_name)) {
                       return false;
                     }
Q
Qiao Longfei 已提交
312
                     auto var = block_desc->Var(grad_var_name);
Y
Yu Yang 已提交
313
                     new_vars.insert(var->Name());
Y
Yu Yang 已提交
314 315 316 317 318 319 320 321 322 323 324
                     auto it = param_name_map.find(grad_var_name);
                     if (it == param_name_map.end()) {
                       return false;
                     }
                     auto param_var_name = it->second;
                     auto& grad_record = (*grad_var_record)[param_var_name];
                     grad_record.name_ = grad_var_name;
                     grad_record.block_idx_ = block_desc->ID();
                     grad_record.op_idx_ = static_cast<int>(op_index);
                     return false; /* not break */
                   });
Y
Yang Yang(Tony) 已提交
325 326 327 328 329 330 331 332 333 334 335 336
    ops[op_index]->InferVarType(block_desc);
    for (auto& arg : ops[op_index]->OutputArgumentNames()) {
      if (new_vars.find(arg) == new_vars.end()) {
        continue;
      }
      auto pname = FwdName(arg);
      auto* param = block_desc->FindVarRecursive(pname);
      auto* grad = block_desc->FindVar(arg);
      if (param == nullptr) {
        grad->SetDataType(DataType::FP32);
      } else {
        grad->SetDataType(param->GetDataType());
Y
Yu Yang 已提交
337
      }
Q
Qiao Longfei 已提交
338
    }
Y
Yang Yang(Tony) 已提交
339
    ops[op_index]->InferShape(*block_desc);
340 341 342
  }
}

F
fengjiayi 已提交
343
std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
344
    const OpDescBind* op_desc, std::unordered_set<std::string>* no_grad_vars,
Y
Yu Yang 已提交
345 346 347
    std::unordered_map<std::string, std::string>* grad_to_var,
    const std::vector<BlockDescBind*>& grad_block =
        std::vector<BlockDescBind*>()) {
F
Update  
fengjiayi 已提交
348
  std::vector<std::unique_ptr<OpDescBind>> grad_op_descs;
349
  // All input gradients of forwarding operator do not need to calculate.
F
fengjiayi 已提交
350
  const std::vector<std::string>& inputs = op_desc->InputArgumentNames();
351
  if (AllGradInSet(inputs, *no_grad_vars)) {
F
fengjiayi 已提交
352 353 354
    return grad_op_descs;  // empty vector
  }
  // All output gradients of forwarding operator do not need to calculate.
F
fengjiayi 已提交
355
  const std::vector<std::string>& outputs = op_desc->OutputArgumentNames();
356
  if (AllGradInSet(outputs, *no_grad_vars)) {
357
    for (const std::string& name : inputs) {
358
      no_grad_vars->insert(GradVarName(name));
F
fengjiayi 已提交
359 360 361 362
    }
    return grad_op_descs;  // empty vector
  }

Y
Yu Yang 已提交
363 364 365 366
  grad_op_descs =
      OpInfoMap::Instance()
          .Get(op_desc->Type())
          .GradOpMaker()(*op_desc, *no_grad_vars, grad_to_var, grad_block);
F
fengjiayi 已提交
367

F
Update  
fengjiayi 已提交
368 369 370
  std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
  for (auto& desc : grad_op_descs) {
    for (const std::string& in_name : desc->InputArgumentNames()) {
371
      if (no_grad_vars->count(in_name)) {
F
fengjiayi 已提交
372 373 374
        std::string prefix = in_name.substr(
            0, in_name.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
        std::string new_name = prefix + kZeroVarSuffix;
F
Update  
fengjiayi 已提交
375
        desc->Rename(in_name, new_name);
F
fengjiayi 已提交
376 377 378
        std::unique_ptr<OpDescBind> fill_zeros_op(new OpDescBind(
            "fill_zeros_like", {{"X", {prefix}}}, {{"Y", {new_name}}}, {}));
        pending_fill_zeros_ops.push_back(std::move(fill_zeros_op));
F
fengjiayi 已提交
379 380 381
      }
    }
  }
F
fengjiayi 已提交
382

F
fengjiayi 已提交
383
  for (auto& p : pending_fill_zeros_ops) {
F
fengjiayi 已提交
384
    grad_op_descs.insert(grad_op_descs.begin(), std::move(p));
F
fengjiayi 已提交
385
  }
F
fengjiayi 已提交
386 387 388
  return grad_op_descs;
}

Y
Yu Yang 已提交
389 390 391 392 393 394
static BlockDescBind* CreateStepBlock(
    ProgramDescBind& program_desc,
    std::unordered_set<std::string>* no_grad_vars,
    std::unordered_map<std::string, std::string>* grad_to_var,
    int step_block_idx);

F
fengjiayi 已提交
395 396
std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
    ProgramDescBind& program_desc, int block_idx,
397 398
    std::unordered_set<std::string>* no_grad_vars,
    std::unordered_map<std::string, std::string>* grad_to_var) {
Y
Yang Yang(Tony) 已提交
399
  VLOG(5) << "MakeBlockBackward";
400
  BlockDescBind* cur_block = program_desc.MutableBlock(block_idx);
401
  std::vector<OpDescBind*> op_descs = cur_block->AllOps();
F
Update  
fengjiayi 已提交
402 403
  std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
  size_t grad_desc_idx = 0;
F
Update  
fengjiayi 已提交
404
  std::vector<std::unique_ptr<OpDescBind>> backward_descs;
405

F
fengjiayi 已提交
406
  for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) {
Y
Yang Yang(Tony) 已提交
407
    VLOG(5) << "Making backward " << (*it)->Type() << " op";
Y
Yu Yang 已提交
408
    std::vector<std::unique_ptr<OpDescBind>> op_grads;
F
fengjiayi 已提交
409

Y
Yang Yang(Tony) 已提交
410
    if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") {
411
      int step_block_idx = (*it)->GetBlockAttr("step_block");
Y
Yu Yang 已提交
412 413 414 415
      BlockDescBind* backward_block = CreateStepBlock(
          program_desc, no_grad_vars, grad_to_var, step_block_idx);
      op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
    } else if ((*it)->Type() == "conditional_block") {
Y
Yu Yang 已提交
416
      BlockDescBind* backward_block =
Y
Yu Yang 已提交
417 418
          CreateStepBlock(program_desc, no_grad_vars, grad_to_var,
                          (*it)->GetBlockAttr("block"));
Y
Yu Yang 已提交
419 420 421
      op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
    } else {
      op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var);
F
fengjiayi 已提交
422 423
    }

Y
Yang Yang(Tony) 已提交
424 425 426 427 428 429 430 431 432
    if (VLOG_IS_ON(10)) {
      std::ostringstream sout;
      sout << "Made ";
      for (auto& op_grad : op_grads) {
        sout << op_grad->Type() << " ";
      }
      VLOG(10) << sout.str();
    }

F
Update  
fengjiayi 已提交
433
    for (const auto& desc : op_grads) {
F
fengjiayi 已提交
434
      for (const std::string& out_name : desc->OutputArgumentNames()) {
435 436 437 438 439
        if (out_name.find("@GRAD") == std::string::npos) {
          // Not all outputs of a backward operator is a gradient. Only gradient
          // need to be sum. Skip variables are not gradient.
          continue;
        }
F
Update  
fengjiayi 已提交
440 441 442 443
        dup_out_ops[out_name].emplace_back(grad_desc_idx);
      }
      ++grad_desc_idx;
    }
F
fengjiayi 已提交
444 445 446
    std::transform(
        op_grads.begin(), op_grads.end(), std::back_inserter(backward_descs),
        [](std::unique_ptr<OpDescBind>& ptr) { return std::move(ptr); });
F
Update  
fengjiayi 已提交
447
  }
Y
Yang Yang(Tony) 已提交
448 449

  VLOG(5) << "Appending Sums";
F
Update  
fengjiayi 已提交
450
  // Check whether some variables are written more than once
F
Update  
fengjiayi 已提交
451
  std::list<std::pair<size_t, std::unique_ptr<OpDescBind>>> pending_sum_ops;
F
Update  
fengjiayi 已提交
452 453 454 455 456
  for (const auto& dup : dup_out_ops) {
    const std::string& out_name = dup.first;
    const std::vector<size_t> dup_op = dup.second;
    if (out_name != kEmptyVarName && dup_op.size() > 1) {
      std::vector<std::string> sum_op_inputs;
Y
Yang Yang(Tony) 已提交
457
      std::string next_g_name = out_name;
F
Update  
fengjiayi 已提交
458
      for (size_t i = 0; i < dup_op.size(); ++i) {
Y
Yang Yang(Tony) 已提交
459 460
        VLOG(10) << backward_descs[dup_op[i]]->Type() << " has " << out_name
                 << " duplicated";
F
Update  
fengjiayi 已提交
461
        std::string new_name = out_name + "@RENAME@" + std::to_string(i);
Y
Yang Yang(Tony) 已提交
462 463
        backward_descs[dup_op[i]]->RenameOutput(out_name, new_name);
        backward_descs[dup_op[i]]->RenameInput(out_name, next_g_name);
F
Update  
fengjiayi 已提交
464
        sum_op_inputs.emplace_back(new_name);
Y
Yang Yang(Tony) 已提交
465
        next_g_name = sum_op_inputs.back();
F
Update  
fengjiayi 已提交
466
      }
F
fengjiayi 已提交
467 468 469
      std::unique_ptr<OpDescBind> sum_op(new OpDescBind(
          "sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {}));
      pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)});
F
Update  
fengjiayi 已提交
470 471
    }
  }
Y
Yang Yang(Tony) 已提交
472

F
Update  
fengjiayi 已提交
473
  pending_sum_ops.sort(
F
Update  
fengjiayi 已提交
474 475 476 477
      [](const std::pair<size_t, std::unique_ptr<OpDescBind>>& a,
         const std::pair<size_t, std::unique_ptr<OpDescBind>>& b) {
        return a.first > b.first;
      });
F
Update  
fengjiayi 已提交
478
  for (auto& p : pending_sum_ops) {
F
Update  
fengjiayi 已提交
479 480
    backward_descs.insert(backward_descs.begin() + p.first + 1,
                          std::move(p.second));
F
Update  
fengjiayi 已提交
481
  }
482

Y
Yang Yang(Tony) 已提交
483 484
  VLOG(5) << "MakeBlockBackward Finished";

F
fengjiayi 已提交
485 486 487
  return backward_descs;
}

Y
Yu Yang 已提交
488 489 490 491 492 493 494 495 496 497 498 499 500 501 502
static BlockDescBind* CreateStepBlock(
    ProgramDescBind& program_desc,
    std::unordered_set<std::string>* no_grad_vars,
    std::unordered_map<std::string, std::string>* grad_to_var,
    int step_block_idx) {
  auto backward_block_op_descs = MakeBlockBackward(program_desc, step_block_idx,
                                                   no_grad_vars, grad_to_var);
  BlockDescBind* backward_block =
      program_desc.AppendBlock(*program_desc.MutableBlock(step_block_idx));
  for (auto& ptr : backward_block_op_descs) {
    backward_block->AppendAllocatedOp(move(ptr));
  }
  return backward_block;
}

Q
qiaolongfei 已提交
503 504 505
ParamGradInfoMap AppendBackward(
    ProgramDescBind& program_desc, const VarDescBind& target,
    const std::unordered_set<std::string>& no_grad_vars) {
F
fengjiayi 已提交
506 507 508 509 510 511
  std::unordered_set<std::string> no_grad_var_names;
  no_grad_var_names.reserve(no_grad_vars.size() + 1);
  no_grad_var_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
  for (auto& name : no_grad_vars) {
    no_grad_var_names.insert(GradVarName(name));
  }
512

F
fengjiayi 已提交
513
  const int root_block_idx = 0;
514
  auto root_block = program_desc.MutableBlock(root_block_idx);
515 516

  std::string fill_one_op_out = GradVarName(target.Name());
517 518
  bool is_scalar = target.Shape() == std::vector<int64_t>{1};
  PADDLE_ENFORCE(is_scalar, "target should be scalar");
Y
Yu Yang 已提交
519 520
  VLOG(3) << "backward from loss=" << target.Name()
          << " data_type=" << target.GetDataType();
521 522
  std::unique_ptr<OpDescBind> fill_one_op(
      new OpDescBind("fill_constant", {}, {{"Out", {fill_one_op_out}}},
523
                     {{"shape", std::vector<int>{1}},
524
                      {"value", static_cast<float>(1.0)},
Y
Yu Yang 已提交
525
                      {"data_type", target.GetDataType()}}));
Q
QI JUN 已提交
526 527 528
  // infer var type of fill_one_op
  fill_one_op->InferVarType(root_block);

529 530
  root_block->AppendAllocatedOp(std::move(fill_one_op));
  size_t forward_op_num = root_block->OpSize();
531
  size_t forward_block_num = program_desc.Size();
Y
Yu Yang 已提交
532 533

  // Insert backward operators
534 535 536
  std::unordered_map<std::string, std::string> grad_to_var;
  auto backward_op_descs = MakeBlockBackward(program_desc, root_block_idx,
                                             &no_grad_var_names, &grad_to_var);
Y
Yu Yang 已提交
537

F
fengjiayi 已提交
538
  for (auto& ptr : backward_op_descs) {
539
    root_block->AppendAllocatedOp(std::move(ptr));
540
  }
Q
Qiao Longfei 已提交
541 542 543 544 545 546
  // Create Variable

  // Create target gradient variable
  std::unordered_map<std::string, GradVarInfo> retv;

  auto var = root_block->Var(fill_one_op_out);
Y
Yu Yang 已提交
547
  var->SetDataType(target.GetDataType());
Q
Qiao Longfei 已提交
548 549 550 551 552
  var->SetShape(target.Shape());
  auto& target_grad = retv[target.Name()];
  target_grad.name_ = fill_one_op_out;
  target_grad.block_idx_ = root_block_idx;
  target_grad.op_idx_ = static_cast<int>(forward_op_num);
553 554

  // create grad_var for all blocks in this program
555
  CreateGradVarInBlock(forward_op_num, grad_to_var, root_block, &retv);
556 557
  for (size_t block_index = forward_block_num;
       block_index < program_desc.Size(); ++block_index) {
558
    CreateGradVarInBlock(0, grad_to_var, program_desc.MutableBlock(block_index),
559
                         &retv);
F
fengjiayi 已提交
560
  }
Y
Yu Yang 已提交
561
  return retv;
F
Update  
fengjiayi 已提交
562 563
}

Y
Yu Yang 已提交
564 565
}  // namespace framework
}  // namespace paddle