backward.cc 15.2 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 21
#include <memory>

F
fengjiayi 已提交
22
#include "paddle/framework/block_desc.h"
23
#include "paddle/framework/op_registry.h"
Y
Yan Chunwei 已提交
24
#include "paddle/operators/net_op.h"
Y
Yan Chunwei 已提交
25
#include "paddle/operators/recurrent_op.h"
Y
Yu Yang 已提交
26 27 28 29

namespace paddle {
namespace framework {

Y
Yu Yang 已提交
30
static inline std::unique_ptr<OperatorBase> CreateGradOp(
31 32
    const OperatorBase& op,
    const std::unordered_set<std::string>& no_grad_set) {
Y
Yu Yang 已提交
33 34 35 36 37 38
  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());
39
  auto grad_descs = info.GradOpMaker()(op_desc, no_grad_set);
Y
Yu Yang 已提交
40 41
  std::vector<std::unique_ptr<OperatorBase>> grad_ops;
  grad_ops.reserve(grad_descs.size());
Y
Yu Yang 已提交
42 43 44
  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
45
                   return OpRegistry::CreateOp(*grad_desc);
Y
Yu Yang 已提交
46
                 });
Y
Yu Yang 已提交
47
  PADDLE_ENFORCE(!grad_ops.empty());
Y
Yu Yang 已提交
48 49 50 51 52 53 54
  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
55
    net_op->CompleteAddOp();
Y
Yu Yang 已提交
56 57 58 59
    return std::unique_ptr<OperatorBase>(net_op);
  }
}

Y
Yu Yang 已提交
60
template <typename Map, typename T>
Q
qiaolongfei 已提交
61
static void ForEachVarName(const Map& names, T callback) {
Y
Yu Yang 已提交
62
  for (auto& name : names) {
Y
Yu Yang 已提交
63
    for (auto& n : name.second) {
64
      if (callback(n)) return;
Y
Yu Yang 已提交
65 66
    }
  }
Y
Yu Yang 已提交
67 68
}

Y
Yan Chunwei 已提交
69
// return whether all the names + suffixes in the set
Y
Yu Yang 已提交
70
static bool AllInSet(
Y
Yu Yang 已提交
71
    const std::map<std::string, std::vector<std::string>>& names,
Y
Yu Yang 已提交
72
    const std::string& suffix, const std::unordered_set<std::string>& set) {
73 74 75 76
  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
77
  });
78
  return all_in_set;
Y
Yu Yang 已提交
79 80
}

Y
Yu Yang 已提交
81 82
static std::unique_ptr<OperatorBase> NOP() {
  auto net_op = new operators::NetOp();
Q
qiaolongfei 已提交
83
  net_op->SetType("@NOP@");
Y
Yu Yang 已提交
84
  net_op->CompleteAddOp();
Y
Yu Yang 已提交
85
  return std::unique_ptr<OperatorBase>(net_op);
Y
Yu Yang 已提交
86 87
}

Y
Yan Chunwei 已提交
88
//  Get backward operator from a forward operator, a recursive implementation.
Y
Yu Yang 已提交
89 90 91
//
//  no_grad_names the gradient variable names without gradient calculating.
//
92 93 94
//  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
95
//
Y
Yan Chunwei 已提交
96 97
//  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
98 99
//
//  See Backward.h for details
Y
Yu Yang 已提交
100
static std::unique_ptr<OperatorBase> BackwardRecursive(
Y
Yu Yang 已提交
101 102
    const OperatorBase& forwardOp,
    std::unordered_set<std::string>& no_grad_names, size_t& uniq_id) {
Y
Yu Yang 已提交
103 104
  //  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
105
  //  too much time for calculation, but it is useful for simplifying logic.
106
  if (AllInSet(forwardOp.Inputs() /*names*/, kGradVarSuffix /*suffix*/,
Y
Yan Chunwei 已提交
107
               no_grad_names /*set*/)) {
Y
Yu Yang 已提交
108
    return NOP();
Y
Yu Yang 已提交
109 110
  }

111 112
  //  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
113
  //  `no_grad_names` set. Return an NOP.
Q
qiaolongfei 已提交
114
  if (AllInSet(forwardOp.Outputs() /*names*/, kGradVarSuffix /*suffix*/,
Y
Yan Chunwei 已提交
115
               no_grad_names /*set*/)) {
Q
qiaolongfei 已提交
116
    ForEachVarName(forwardOp.Inputs(),
Y
Yu Yang 已提交
117 118 119 120
                   [&no_grad_names](const std::string& name) -> bool {
                     no_grad_names.insert(GradVarName(name));
                     return false;
                   });
Y
Yu Yang 已提交
121
    return NOP();
Y
Yu Yang 已提交
122 123
  }

Y
Yu Yang 已提交
124
  // Returned gradient network
Y
Yu Yang 已提交
125
  auto net = std::unique_ptr<operators::NetOp>(new operators::NetOp());
Y
Yu Yang 已提交
126 127

  if (forwardOp.IsNetOp()) {
Y
Yu Yang 已提交
128
    // Because forwardOp is a net op, it can static_cast.
Y
Yan Chunwei 已提交
129
    auto& forwardNet = static_cast<const operators::NetOp&>(forwardOp);
Y
Yu Yang 已提交
130

131
    // Map from output gradient variable name to operator's indices in
Y
Yan Chunwei 已提交
132
    // backward net's ops_. That operator generates that variable.
Y
Yu Yang 已提交
133 134 135
    std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;

    size_t local_op_id = 0;
Y
Yan Chunwei 已提交
136
    // reversely travel forwardNet and collect all duplicate outputs.
Y
Yu Yang 已提交
137
    for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
Y
Yu Yang 已提交
138
         ++it, ++local_op_id) {
Y
Yu Yang 已提交
139
      auto& fwd = *it;
Y
Yu Yang 已提交
140
      auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id);
Q
qiaolongfei 已提交
141
      ForEachVarName(bwd->Outputs(),
Y
Yu Yang 已提交
142 143 144 145
                     [&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
146
      net->AppendOp(std::move(bwd));
D
dongzhihong 已提交
147
    }
Y
Yu Yang 已提交
148
    // Get unique ID for this method.
D
dongzhihong 已提交
149
    auto uid = uniq_id++;
D
dongzhihong 已提交
150
    // TODO(dzh): more comment
Y
Yan Chunwei 已提交
151 152 153 154 155
    // 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
156
    using Pos = std::pair<size_t, std::unique_ptr<OperatorBase>>;
Y
Yu Yang 已提交
157
    std::list<Pos> insert_position;
D
dongzhihong 已提交
158
    for (auto& dup_output_op : dup_output_ops) {
D
dongzhihong 已提交
159
      const std::string& name = dup_output_op.first;
Q
qijun 已提交
160 161 162
      // duplicate @Empty@ don't need to be added
      if (name == kEmptyVarName) continue;

D
dongzhihong 已提交
163
      auto& dup_op = dup_output_op.second;
Y
Yan Chunwei 已提交
164
      // no duplicate output
D
dongzhihong 已提交
165 166
      if (dup_op.size() == 1) continue;

Y
Yan Chunwei 已提交
167 168
      // process the duplicate outputs
      std::vector<std::string> dup_outputs;
D
dongzhihong 已提交
169
      for (size_t i = 0; i < dup_op.size(); ++i) {
Y
Yan Chunwei 已提交
170
        // rename each duplicate output to an alias
D
dongzhihong 已提交
171
        auto op_offset = dup_op[i];
D
dongzhihong 已提交
172 173 174
        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
175
      }
176 177 178 179 180
      // 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
181
    }
Y
Yu Yang 已提交
182

183
    // make sure the inserted `sum` ops follow the BFS order.
Y
Yu Yang 已提交
184
    insert_position.sort(
D
dongzhihong 已提交
185
        [](const Pos& l, const Pos& r) { return l.first > r.first; });
Y
Yu Yang 已提交
186 187

    for (auto& pos : insert_position) {
Y
Yu Yang 已提交
188
      net->InsertOp(pos.first + 1, std::move(pos.second));
D
dongzhihong 已提交
189
    }
Y
Yu Yang 已提交
190
  } else {
191 192
    std::unique_ptr<OperatorBase> grad_op(
        CreateGradOp(forwardOp, no_grad_names));
Y
Yu Yang 已提交
193

Y
Yu Yang 已提交
194 195
    ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
                                          const std::string& grad_input) {
196
      if (no_grad_names.count(grad_input)) {
Y
Yu Yang 已提交
197
        // +1 for \0
198
        std::string prefix = grad_input.substr(
Y
Yu Yang 已提交
199
            0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
Q
qiaolongfei 已提交
200
        grad_op->Rename(grad_input, prefix + kZeroVarSuffix);
Y
Yu Yang 已提交
201 202 203

        // If part of input gradient of that operator is not calculated, fill
        // zero variables to that input gradient.
D
dangqingqing 已提交
204 205
        net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", {{"X", {prefix}}},
                                           {{"Y", {grad_input}}}, {}));
206
      }
Y
Yu Yang 已提交
207 208 209
      return false;
    });

Q
qiaolongfei 已提交
210 211
    ForEachVarName(grad_op->Outputs(),
                   [&no_grad_names, &grad_op](const std::string& grad_output) {
Y
Yu Yang 已提交
212
                     if (no_grad_names.count(grad_output)) {
Q
qiaolongfei 已提交
213
                       grad_op->Rename(grad_output, kEmptyVarName);
Y
Yu Yang 已提交
214 215 216
                     }
                     return false;
                   });
Y
Yu Yang 已提交
217

Y
Yan Chunwei 已提交
218
    // process recurrent gradient op as a special operator.
219
    if (forwardOp.Type() == "recurrent") {
F
Fix bug  
fengjiayi 已提交
220 221
      // NOTE clean up cycle call somewhere (RNN's stepnet constains itself),
      // or
Y
Yan Chunwei 已提交
222 223 224 225 226 227 228 229 230
      // this will result in infinite loop.
      const auto& rnnop =
          *static_cast<const operators::RecurrentOp*>(&forwardOp);
      auto rnn_grad_op =
          static_cast<operators::RecurrentGradientOp*>(grad_op.get());
      const auto& stepnet_op =
          *static_cast<const OperatorBase*>(&rnnop.stepnet());
      // create stepnet's gradient op
      rnn_grad_op->set_stepnet(
Y
Yu Yang 已提交
231
          BackwardRecursive(stepnet_op, no_grad_names, uniq_id));
Y
Yan Chunwei 已提交
232 233
    }

Y
Yu Yang 已提交
234 235 236
    if (net->ops_.empty()) {  // Current no aux op is added to network
      return grad_op;
    }
Y
Yu Yang 已提交
237
    net->AppendOp(std::move(grad_op));
Y
Yu Yang 已提交
238
  }
Q
qiaolongfei 已提交
239
  net->SetType("@GENERATED_BACKWARD@");
Y
Yu Yang 已提交
240
  net->CompleteAddOp();
Y
Yu Yang 已提交
241 242 243
  return std::unique_ptr<OperatorBase>(
      static_cast<OperatorBase*>(net.release()));
}
Y
Yu Yang 已提交
244

Y
Yu Yang 已提交
245
// See header for comments
Y
Yu Yang 已提交
246
std::unique_ptr<OperatorBase> Backward(
Y
Yu Yang 已提交
247
    const OperatorBase& forwardOp,
Y
Yu Yang 已提交
248 249
    const std::unordered_set<std::string>& no_grad_vars) {
  std::unordered_set<std::string> no_grad_names;
Q
qijun 已提交
250
  no_grad_names.reserve(no_grad_vars.size() + 1);
Y
Yu Yang 已提交
251

252
  no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
253

Y
Yu Yang 已提交
254
  for (auto& name : no_grad_vars) {
255
    no_grad_names.insert(name + kGradVarSuffix);
Y
Yu Yang 已提交
256
  }
Y
Yu Yang 已提交
257
  size_t uid = 0;
Y
Yu Yang 已提交
258
  return BackwardRecursive(forwardOp, no_grad_names, uid);
Y
Yu Yang 已提交
259
}
Y
Yi Wang 已提交
260

F
fengjiayi 已提交
261 262 263 264 265 266 267 268 269 270 271 272
// ====================================  //

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;
    }
  }
  return true;
}

F
fengjiayi 已提交
273
std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
F
Update  
fengjiayi 已提交
274
    const std::unique_ptr<OpDescBind>& op_desc,
F
fengjiayi 已提交
275
    std::unordered_set<std::string>& no_grad_vars) {
F
Update  
fengjiayi 已提交
276
  std::vector<std::unique_ptr<OpDescBind>> grad_op_descs;
277
  // All input gradients of forwarding operator do not need to calculate.
F
fengjiayi 已提交
278
  const std::vector<std::string>& inputs = op_desc->InputArgumentNames();
279
  if (AllGradInSet(inputs, no_grad_vars)) {
F
fengjiayi 已提交
280 281 282
    return grad_op_descs;  // empty vector
  }
  // All output gradients of forwarding operator do not need to calculate.
F
fengjiayi 已提交
283 284
  const std::vector<std::string>& outputs = op_desc->OutputArgumentNames();
  if (AllGradInSet(outputs, no_grad_vars)) {
285
    for (const std::string& name : inputs) {
F
fengjiayi 已提交
286 287 288 289 290
      no_grad_vars.insert(GradVarName(name));
    }
    return grad_op_descs;  // empty vector
  }

291 292 293
  grad_op_descs = OpInfoMap::Instance()
                      .Get(op_desc->Type())
                      .GradOpMaker()(*op_desc, no_grad_vars);
F
fengjiayi 已提交
294

F
Update  
fengjiayi 已提交
295 296 297
  std::list<std::unique_ptr<OpDescBind>> pending_fill_zeros_ops;
  for (auto& desc : grad_op_descs) {
    for (const std::string& in_name : desc->InputArgumentNames()) {
F
fengjiayi 已提交
298 299 300 301
      if (no_grad_vars.count(in_name)) {
        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
302
        desc->Rename(in_name, new_name);
F
fengjiayi 已提交
303 304 305
        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
306 307 308
      }
    }
  }
F
fengjiayi 已提交
309

F
fengjiayi 已提交
310
  for (auto& p : pending_fill_zeros_ops) {
F
fengjiayi 已提交
311
    grad_op_descs.insert(grad_op_descs.begin(), std::move(p));
F
fengjiayi 已提交
312
  }
F
fengjiayi 已提交
313 314 315
  return grad_op_descs;
}

F
fengjiayi 已提交
316 317 318 319 320
std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
    ProgramDescBind& program_desc, int block_idx,
    std::unordered_set<std::string>& no_grad_vars) {
  BlockDescBind* cur_block = program_desc.Block(block_idx);
  std::deque<std::unique_ptr<OpDescBind>>& op_descs = cur_block->ops_;
F
Update  
fengjiayi 已提交
321 322
  std::unordered_map<std::string, std::vector<size_t>> dup_out_ops;
  size_t grad_desc_idx = 0;
F
Update  
fengjiayi 已提交
323
  std::vector<std::unique_ptr<OpDescBind>> backward_descs;
F
fengjiayi 已提交
324
  for (auto it = op_descs.rbegin(); it != op_descs.rend(); ++it) {
F
Update  
fengjiayi 已提交
325
    std::vector<std::unique_ptr<OpDescBind>> op_grads =
F
fengjiayi 已提交
326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341
        MakeOpGrad(*it, no_grad_vars);

    if ((*it)->Type() == "recurrent") {
      PADDLE_ENFORCE_EQ(
          op_grads.size(), size_t(1),
          "rnn_op's gradient process should contain only one op.");
      int step_block_idx = (*it)->GetBlockAttr("stop_block");
      auto backward_block_op_descs =
          MakeBlockBackward(program_desc, step_block_idx, no_grad_vars);
      BlockDescBind* backward_block = program_desc.AppendBlock(*cur_block);
      for (auto& ptr : backward_block_op_descs) {
        backward_block->ops_.push_back(std::move(ptr));
      }
      op_grads[0]->SetBlockAttr("step_block", *backward_block);
    }

F
Update  
fengjiayi 已提交
342
    for (const auto& desc : op_grads) {
F
fengjiayi 已提交
343
      for (const std::string& out_name : desc->OutputArgumentNames()) {
F
Update  
fengjiayi 已提交
344 345 346 347
        dup_out_ops[out_name].emplace_back(grad_desc_idx);
      }
      ++grad_desc_idx;
    }
F
fengjiayi 已提交
348 349 350
    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
351 352
  }
  // Check whether some variables are written more than once
F
Update  
fengjiayi 已提交
353
  std::list<std::pair<size_t, std::unique_ptr<OpDescBind>>> pending_sum_ops;
F
Update  
fengjiayi 已提交
354 355 356 357 358 359 360
  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;
      for (size_t i = 0; i < dup_op.size(); ++i) {
        std::string new_name = out_name + "@RENAME@" + std::to_string(i);
F
Update  
fengjiayi 已提交
361
        backward_descs[dup_op[i]]->Rename(out_name, new_name);
F
Update  
fengjiayi 已提交
362 363
        sum_op_inputs.emplace_back(new_name);
      }
F
fengjiayi 已提交
364 365 366
      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
367 368 369
    }
  }
  pending_sum_ops.sort(
F
Update  
fengjiayi 已提交
370 371 372 373
      [](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
374
  for (auto& p : pending_sum_ops) {
F
Update  
fengjiayi 已提交
375 376
    backward_descs.insert(backward_descs.begin() + p.first + 1,
                          std::move(p.second));
F
Update  
fengjiayi 已提交
377
  }
F
fengjiayi 已提交
378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394
  return backward_descs;
}

void AppendBackward(ProgramDescBind& program_desc,
                    const std::unordered_set<std::string>& no_grad_vars) {
  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));
  }
  const int root_block_idx = 0;
  auto backward_op_descs =
      MakeBlockBackward(program_desc, root_block_idx, no_grad_var_names);
  auto& forw_op_descs = program_desc.Block(root_block_idx)->ops_;
  for (auto& ptr : backward_op_descs) {
    forw_op_descs.push_back(std::move(ptr));
F
fengjiayi 已提交
395
  }
F
Update  
fengjiayi 已提交
396 397
}

Y
Yu Yang 已提交
398 399
}  // namespace framework
}  // namespace paddle