backward.py 62.4 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# 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
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# 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 16
from __future__ import print_function

17
from paddle.fluid import framework as framework
F
update  
fengjiayi 已提交
18
from . import core
F
update  
fengjiayi 已提交
19
import collections
20
import copy
21
import six
22
import logging
M
minqiyang 已提交
23
from .. import compat as cpt
24
from . import unique_name
25
from . import log_helper
L
liym27 已提交
26
import paddle.fluid
M
mapingshuo 已提交
27 28 29 30 31
__all__ = [
    'append_backward',
    'gradients',
]

32 33 34
_logger = log_helper.get_logger(
    __name__, logging.INFO, fmt='%(asctime)s-%(levelname)s: %(message)s')

M
mapingshuo 已提交
35 36 37 38 39 40 41 42 43 44 45

class ProgramStats(object):
    def __init__(self, block, ops):
        self.block = block
        self.ops = ops
        self.op_deps = {}  # op-> in_ops, out_ops
        self.var_op_deps = {}  # var as input op, var as output op

    def get_input_nodes(self):
        input_names = []
        for name in self.var_op_deps:
46
            if len(self.var_op_deps[name]["var_as_output_ops"]) == 0 and \
M
mapingshuo 已提交
47 48 49 50 51 52 53 54 55 56 57 58
               len(self.var_op_deps[name]["var_as_input_ops"]) > 0:
                if self.block.var(name).persistable:
                    continue
                input_names.append(name)
        for op in self.ops:
            if op.desc.type() == "read":
                input_names.extend(op.desc.output_arg_names())
        return input_names

    def get_reserved_vars(self):
        var_name = []
        for op in self.ops:
M
mapingshuo 已提交
59
            if op.desc.type() == "seed":
M
mapingshuo 已提交
60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
                var_name.extend(op.desc.output_arg_names())
        return var_name

    def get_out_of_subgraph_vars(self, begin_op_idx, end_op_idx):
        var_name = []
        for i in range(begin_op_idx, end_op_idx, 1):
            for name in self.ops[i].desc.output_arg_names():
                if name in self.var_op_deps:
                    for idx in self.var_op_deps[name]["var_as_input_ops"]:
                        if idx >= end_op_idx:
                            var_name.append(name)
        return var_name

    def is_subgraph(self, var_group1, var_group2):
        # should traverse from var_group1 to var_group2
        # max op idx in var_group2
        # min op idx in var_group1
        min_op_idx = len(self.ops)
        max_op_idx = -1
        for name in var_group1:
            if name not in self.var_op_deps:
                return False, min_op_idx, max_op_idx
        for name in var_group2:
            if name not in self.var_op_deps:
                return False, min_op_idx, max_op_idx
        for name in var_group1:
            op_idx = self.var_op_deps[name]["var_as_input_ops"]
            for idx in op_idx:
                min_op_idx = min(min_op_idx, idx)
        for name in var_group2:
            op_idx = self.var_op_deps[name]["var_as_output_ops"]
            for idx in op_idx:
                max_op_idx = max(max_op_idx, idx)
        if min_op_idx >= max_op_idx:
            return False, min_op_idx, max_op_idx
        return True, min_op_idx, max_op_idx

    def build_stats(self):
        for i, op in enumerate(self.ops):
            self.op_deps[i] = {"in_ops": [], "out_ops": []}
            for j, name in enumerate(op.desc.input_arg_names()):
                if name in self.var_op_deps:
                    self.op_deps[i]["in_ops"].extend(self.var_op_deps[name][
                        "var_as_output_ops"])
            for j, name in enumerate(op.desc.input_arg_names()):
                if name in self.var_op_deps:
                    self.var_op_deps[name]["var_as_input_ops"].extend([i])
                else:
                    self.var_op_deps[name] = {}
                    self.var_op_deps[name]["var_as_input_ops"] = [i]
                    self.var_op_deps[name]["var_as_output_ops"] = []

            for j, name in enumerate(op.desc.output_arg_names()):
                if name in self.var_op_deps:
                    self.var_op_deps[name]["var_as_output_ops"].extend([i])
                else:
                    self.var_op_deps[name] = {}
                    self.var_op_deps[name]["var_as_input_ops"] = []
                    self.var_op_deps[name]["var_as_output_ops"] = [i]

            for op_idx in self.op_deps[i]["in_ops"]:
                self.op_deps[op_idx]["out_ops"].extend([i])

123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138
    def sort_checkpoints(self, checkpoints_name):
        sorted_checkpoints = []
        for name in checkpoints_name:
            if name not in self.var_op_deps:
                _logger.debug(
                    "Recompute Optimizer: deleted %s from checkpoints, because it is not used in paddle program."
                    % name)
            elif self.var_op_deps[name]["var_as_output_ops"] == []:
                # input nodes
                sorted_checkpoints.append((name, -1))
            else:
                sorted_checkpoints.append(
                    (name, max(self.var_op_deps[name]["var_as_output_ops"])))
        sorted_checkpoints = sorted(sorted_checkpoints, key=lambda x: x[1])
        return [x[0] for x in sorted_checkpoints]

M
mapingshuo 已提交
139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174
    def modify_forward_desc_for_recompute(self):
        op_types = [op.desc.type() for op in self.ops]
        if "dropout" not in op_types:
            return

        op_idx = 0
        while (op_idx < len(self.ops)):
            op = self.ops[op_idx]
            if op.desc.type() != "dropout":
                op_idx += 1
                continue
            # add a seed op so that the two dropout op can generate same output
            op_unique_name = unique_name.generate("seed")
            var_unique_name = unique_name.generate_with_ignorable_key(".".join(
                [op_unique_name, 'tmp']))
            added_var = self.block.create_var(
                name=var_unique_name,
                dtype='int32',
                type=core.VarDesc.VarType.LOD_TENSOR,
                persistable=False,
                stop_gradient=False)
            seed = 0 if op.attr("fix_seed") is False else int(op.attr("seed"))
            added_op = self.block._insert_op(
                index=op.idx,
                type='seed',
                inputs={},
                outputs={'Out': [added_var]},
                attrs={'seed': seed})
            self.ops.insert(op_idx, added_op)
            # modify dropout op desc so that it accept a seed var as input
            op.desc.set_input("Seed", [var_unique_name])
            op.desc.remove_attr("fix_seed")
            op.desc.remove_attr("seed")
            self.block._sync_with_cpp()
            op_idx += 2

M
mapingshuo 已提交
175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236

def _pretty_op_desc_(op_desc, prefix):
    out_s = "%s\tname:[%s]\n%s    \tinputs:[%s]\n%s    \toutputs:[%s]" % \
            (prefix + "_op", str(op_desc.type()), prefix + "_input", " ".join(op_desc.input_arg_names()),
             prefix + "_output", " ".join(op_desc.output_arg_names()))
    return out_s


def _add_needed_descs_to_block(descs, block, main_block, in_memory_vars):
    if len(descs) == 0:
        return []
    result_descs = []
    op_role_attr_name = \
            core.op_proto_and_checker_maker.kOpRoleAttrName()
    backward = core.op_proto_and_checker_maker.OpRole.Backward
    for desc in descs:
        if isinstance(desc, framework.Operator):
            desc = desc.desc
        if isinstance(desc, tuple):
            desc = desc[0]
        is_needed = False
        for name in desc.output_arg_names():
            if main_block.has_var(name) and main_block.var(name).persistable:
                continue
            if name not in in_memory_vars:
                is_needed = True
        if is_needed:
            new_op_desc = block.desc.append_op()
            new_op_desc.copy_from(desc)
            new_op_desc._set_attr(op_role_attr_name, backward)
            result_descs.append(new_op_desc)
    return result_descs


def _add_descs_to_block(descs, block):
    if len(descs) == 0:
        return []
    result_descs = []
    op_role_attr_name = \
        core.op_proto_and_checker_maker.kOpRoleAttrName()
    backward = core.op_proto_and_checker_maker.OpRole.Backward
    for desc in descs:
        if isinstance(desc, framework.Operator):
            desc = desc.desc
        if isinstance(desc, tuple):
            desc = desc[0]
        new_op_desc = block.desc.append_op()
        new_op_desc.copy_from(desc)
        new_op_desc._set_attr(op_role_attr_name, backward)
        result_descs.append(new_op_desc)
    return result_descs


def _find_loss_op_(loss):
    for op in reversed(loss.block.ops):
        assert isinstance(op, framework.Operator)
        if len(op.output_arg_names) == 1 and op.output_arg_names[
                0] == loss.name:
            loss.op = op
            break
    if loss.op is None:
        raise ValueError("loss.op is None. Should not happend")
237 238


239 240
def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None):
    """
241
    Traverse all ops in op_descs[begin_idx : end_idx],
242 243
    if any op has inputs/outputs named "old_name", rename it as 'new_name'
    """
F
update  
fengjiayi 已提交
244 245 246
    if begin_idx is None:
        begin_idx = 0
    if end_idx is None:
247
        end_idx = len(op_descs)
F
update  
fengjiayi 已提交
248
    for i in range(begin_idx, end_idx):
249
        op_desc = op_descs[i]
F
fengjiayi 已提交
250 251
        if isinstance(op_desc, tuple):
            op_desc = op_desc[0]
W
Wu Yi 已提交
252 253
        op_desc._rename_input(old_name, new_name)
        op_desc._rename_output(old_name, new_name)
F
update  
fengjiayi 已提交
254 255


F
fengjiayi 已提交
256
def _create_op_desc_(op_type, inputs, outputs, attrs):
257 258 259
    """
    Create a C++ OpDesc object with specified inputs, outputs and attributes.
    """
F
fengjiayi 已提交
260 261
    op_desc = core.OpDesc()
    op_desc.set_type(op_type)
M
minqiyang 已提交
262
    for para, args in six.iteritems(inputs):
263 264 265 266 267
        op_desc.set_input(
            para,
            list(
                map(lambda arg: arg.decode() if isinstance(arg, six.binary_type) else arg,
                    args)))
M
minqiyang 已提交
268
    for para, args in six.iteritems(outputs):
269 270 271 272 273
        op_desc.set_output(
            para,
            list(
                map(lambda arg: arg.decode() if isinstance(arg, six.binary_type) else arg,
                    args)))
Y
yuyang18 已提交
274 275 276 277 278 279

    op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName()

    if op_role_attr_name not in attrs:
        attrs[
            op_role_attr_name] = core.op_proto_and_checker_maker.OpRole.Backward
M
minqiyang 已提交
280
    for name, val in six.iteritems(attrs):
F
fengjiayi 已提交
281 282 283
        if isinstance(val, framework.Block):
            op_desc.set_block_attr(name, val.desc)
        else:
W
Wu Yi 已提交
284
            op_desc._set_attr(name, val)
F
fengjiayi 已提交
285 286 287
    return op_desc


M
mapingshuo 已提交
288 289 290 291 292 293 294 295 296 297 298 299 300 301
def _create_loss_op_desc_(loss):
    op_desc = _create_op_desc_(
        "fill_constant", {}, {"Out": [_append_grad_suffix_(loss.name)]}, {
            "shape": [1],
            "value": 1.0,
            "dtype": loss.dtype,
            "force_cpu": False,
            core.op_proto_and_checker_maker.kOpRoleAttrName():
            int(core.op_proto_and_checker_maker.OpRole.Backward) |
            int(core.op_proto_and_checker_maker.OpRole.Loss),
        })
    return op_desc


302
def _infer_var_data_type_shape_(grad_var_name, block):
303
    """
304
    Infer the data type and shape of given grad variable
305
    """
M
minqiyang 已提交
306 307 308 309
    grad_var = block.desc.find_var(cpt.to_bytes(grad_var_name))
    fwd_name = _strip_grad_suffix_(grad_var_name)
    if block.desc.has_var_recursive(cpt.to_bytes(fwd_name)):
        fwd_var = block.desc.find_var_recursive(cpt.to_bytes(fwd_name))
F
fengjiayi 已提交
310
        grad_var.set_dtype(fwd_var.dtype())
311
        grad_var.set_shape(fwd_var.shape())
F
fengjiayi 已提交
312
    else:
313
        grad_var.set_dtype(core.VarDesc.VarType.FP32)
F
fengjiayi 已提交
314 315


F
fengjiayi 已提交
316
def _all_in_set_(cands, s):
317 318 319
    """
    Test if all elements of 'cands' are in set 's'
    """
F
fengjiayi 已提交
320 321
    if len(cands) == 0:
        return False
F
fengjiayi 已提交
322 323 324 325 326 327
    for c in cands:
        if not c in s:
            return False
    return True


328 329 330 331 332 333
def _some_in_set_(cands, s):
    """
    Test if some elements of 'cands' are in set 's'
    """
    if len(cands) == 0:
        return False
M
minqiyang 已提交
334 335
    literal_set = cpt.to_text(s)
    literal_cands = cpt.to_text(cands)
M
minqiyang 已提交
336 337
    for c in literal_cands:
        if c in literal_set:
338 339 340 341
            return True
    return False


F
fengjiayi 已提交
342
def _strip_grad_suffix_(name):
343
    """
M
mapingshuo 已提交
344
    Strip the grad suffix from the given variable name
345 346 347
    e.g. x@GRAD ==> x
         y@GRAD@RENAME@1 ==> y
    """
M
minqiyang 已提交
348
    name = cpt.to_text(name)
M
minqiyang 已提交
349
    pos = name.find(core.grad_var_suffix())
F
fengjiayi 已提交
350
    return name[:pos] if pos != -1 else name
F
fengjiayi 已提交
351 352 353


def _append_grad_suffix_(name):
354 355 356 357
    """
    Append grad suffix to the given variable name
    e.g. x ==> x@GRAD
    """
M
minqiyang 已提交
358
    return cpt.to_text(name) + core.grad_var_suffix()
F
fengjiayi 已提交
359 360


361
def _addup_repetitive_outputs_(op_descs, block_idx):
362 363
    """
    In backward part, an variable may be the output of more than one ops.
F
fengjiayi 已提交
364 365
    And one op may yield its multiple outputs to the same variable.
    In these cases, the variable should be the accumulation of all the outputs.
366 367
    `sum_op`s are added to implement the accumulate.
    """
F
update  
fengjiayi 已提交
368 369
    pending_sum_ops = []
    var_rename_count = collections.defaultdict(int)
F
fengjiayi 已提交
370
    renamed_vars = collections.defaultdict(list)
371
    renamed_var_start_idx = collections.defaultdict(list)
F
fengjiayi 已提交
372
    for idx, op_desc in enumerate(op_descs):
F
update  
fengjiayi 已提交
373
        for var_name in op_desc.input_arg_names():
M
mapingshuo 已提交
374 375
            if "@GRAD" not in var_name:
                continue
F
fengjiayi 已提交
376
            if len(renamed_vars[var_name]) > 1:
377 378 379
                pending_sum_ops.append((_create_op_desc_(
                    "sum", {"X": renamed_vars[var_name]}, {"Out": [var_name]},
                    {"use_mkldnn": False}), idx))
F
fengjiayi 已提交
380
                renamed_vars[var_name] = [var_name]
F
update  
fengjiayi 已提交
381
        for param_idx, param_name in enumerate(op_desc.output_names()):
F
fengjiayi 已提交
382 383
            arg_names = op_desc.output(param_name)
            for arg_idx, var_name in enumerate(arg_names):
M
mapingshuo 已提交
384 385 386 387
                if "@GRAD" not in var_name:
                    continue
                #if "@RENAME@" in var_name:
                #    continue
F
fengjiayi 已提交
388 389 390 391 392 393 394
                if var_name == core.empty_var_name(
                ) or var_name in op_desc.input_arg_names():
                    # empty variable or inplace op
                    continue
                if len(renamed_vars[var_name]) == 0:
                    # it's the first time we get the variable
                    renamed_vars[var_name] = [var_name]
395
                    renamed_var_start_idx[var_name] = idx
F
fengjiayi 已提交
396 397
                else:
                    if len(renamed_vars[var_name]) == 1:
398
                        new_name = var_name + "@RENAME@block" + str(block_idx) + "@" + \
F
fengjiayi 已提交
399 400 401 402
                            str(var_rename_count[var_name])
                        var_rename_count[var_name] += 1
                        # rename original var_name
                        renamed_vars[var_name][0] = new_name
403 404 405 406 407 408
                        # before change: _rename_arg_(op_descs, var_name,
                        #                             new_name, 0, idx)
                        # rename arg from idx of the first appearance
                        # in backward, not always from 0
                        _rename_arg_(op_descs, var_name, new_name,
                                     renamed_var_start_idx[var_name], idx)
F
fengjiayi 已提交
409 410
                        _rename_arg_(pending_sum_ops, var_name, new_name)

F
update  
fengjiayi 已提交
411 412 413 414 415 416 417 418 419 420 421 422 423
                        for p in op_desc.output_names()[:param_idx]:
                            p_arg_names = op_desc.output(p)
                            if var_name in p_arg_names:
                                op_desc.set_output(p, [
                                    new_name if x == var_name else x
                                    for x in p_arg_names
                                ])

                        arg_names = [
                            new_name if x == var_name else x
                            for x in arg_names[:arg_idx]
                        ] + arg_names[arg_idx:]

424
                    new_name = var_name + "@RENAME@block" + str(block_idx) + "@" + \
F
fengjiayi 已提交
425
                        str(var_rename_count[var_name])
F
fengjiayi 已提交
426
                    var_rename_count[var_name] += 1
F
fengjiayi 已提交
427 428 429
                    arg_names[arg_idx] = new_name
                    op_desc.set_output(param_name, arg_names)
                    renamed_vars[var_name].append(new_name)
F
update  
fengjiayi 已提交
430

M
minqiyang 已提交
431
    for var_name, inputs in six.iteritems(renamed_vars):
F
update  
fengjiayi 已提交
432
        if len(inputs) > 1:
433 434 435
            pending_sum_ops.append(
                (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]},
                                  {"use_mkldnn": False}), len(op_descs)))
F
fengjiayi 已提交
436
    # sum_op descs are sorted according to their insert position
F
update  
fengjiayi 已提交
437
    for p in reversed(pending_sum_ops):
F
fengjiayi 已提交
438 439 440 441 442 443
        op_descs.insert(p[1], p[0])

    return op_descs


def _remove_no_grad_branch_(op_descs, no_grad_set):
444 445 446 447
    """
    Remove unnecessary grad ops
    A grad op can be removed in two cases:
        1. all outputs of the grad op are in 'no_grad_set'
F
fengjiayi 已提交
448
        2. all grad inputs of the grad op are in 'no_grad_set'
449
    """
F
fengjiayi 已提交
450 451

    def _op_can_be_removed_(op_desc, no_grad_set):
F
fengjiayi 已提交
452 453
        out_arg_names = op_desc.output_arg_names()
        if len(out_arg_names) == 0 or _all_in_set_(out_arg_names, no_grad_set):
F
fengjiayi 已提交
454
            return True
455 456 457 458
        if _all_in_set_([
                name for name in op_desc.input_arg_names()
                if name.find(core.grad_var_suffix()) != -1
        ], no_grad_set):
F
fengjiayi 已提交
459
            no_grad_set.update(out_arg_names)
F
fengjiayi 已提交
460 461 462
            return True
        return False

F
fengjiayi 已提交
463
    # Remove ops whose outputs are all in no_grad_dict
464 465 466 467
    op_descs = [
        op_desc for op_desc in op_descs
        if not _op_can_be_removed_(op_desc, no_grad_set)
    ]
F
fengjiayi 已提交
468 469
    # Insert fill_zeros_like_op
    to_insert = []
F
fengjiayi 已提交
470
    for idx, op_desc in enumerate(op_descs):
F
fengjiayi 已提交
471
        for arg in op_desc.input_arg_names():
M
mapingshuo 已提交
472
            # arg is a gradient var name and arg should not have gradient
F
fengjiayi 已提交
473
            if core.grad_var_suffix() in arg and arg in no_grad_set:
474
                x_in = _strip_grad_suffix_(arg)
M
mapingshuo 已提交
475 476
                # the reason should be: arg can be input of another grad op
                # and the op is a not-to-remove op
477 478
                to_insert.append((_create_op_desc_(
                    "fill_zeros_like", {"X": [x_in]}, {"Out": [arg]}, {}), idx))
F
fengjiayi 已提交
479

480
    list([op_descs.insert(p[1], p[0]) for p in reversed(to_insert)])
F
fengjiayi 已提交
481 482 483 484

    return op_descs


C
chengduo 已提交
485 486 487 488 489 490 491 492 493 494 495 496 497 498 499
def _find_not_need_ops(grad_op_descs, forward_ops, input_grad_names_set):
    """
    Pruning Program with Structural Analysis Method of Computational Graph.
    The nodes of the computational graph composed of backward OPS should be
    interconnected. If there are unconnected sub-graphs in the computational graph,
    these sub-graphs should be cut off.

    Args:
        grad_op_descs(list[core.OpDesc]): The candidate backward OpDescs.
        forward_ops(list[Operator]): The forward ops.
        input_grad_names_set(set): this set is used to store the gradients' name
            which is generated by backward ops, and input_grad_names_set can help
            to prune the unnecessary backward ops.

    Return:
500
        (set[core.OpDesc]): A set of OpDescs which should be pruned.
C
chengduo 已提交
501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599
    """

    class Var(object):
        def __init__(self, var_name):
            self.var_name = var_name
            self.gen_op = None
            self.pendding_ops = []

        def set_gen_op(self, gen_op):
            assert isinstance(gen_op, Op)
            assert self.gen_op is None
            self.gen_op = gen_op

        def add_pending_op(self, op):
            assert isinstance(op, Op)
            self.pendding_ops.append(op)

    class Op(object):
        def __init__(self, op_desc):
            self.op_desc = op_desc
            self.inputs = []
            self.outputs = []

        def insert_input(self, var):
            assert isinstance(var, Var)
            self.inputs.append(var)

        def insert_output(self, var):
            assert isinstance(var, Var)
            self.outputs.append(var)

    var_versions = dict()

    def _create_node(name):
        if name not in var_versions.keys():
            var_versions[name] = [Var(name)]
        else:
            var_versions[name].append(Var(name))
        return var_versions[name][-1]

    def _create_or_get_last_version_node(name):
        if name not in var_versions.keys():
            var_versions[name] = [Var(name)]
        return var_versions[name][-1]

    def _create_op_node(op_desc):
        op_node = Op(op_desc)
        for input in op_desc.input_arg_names():
            var = _create_or_get_last_version_node(name=input)
            var.add_pending_op(op_node)
            op_node.insert_input(var)
        for output in op_desc.output_arg_names():
            var = _create_node(name=output)
            var.set_gen_op(op_node)
            op_node.insert_output(var)
        return op_node

    # Record the forward vars
    forward_vars_set = set() if input_grad_names_set is None else set(
        input_grad_names_set)
    for op in forward_ops:
        forward_vars_set.update(op.desc.input_arg_names())
        forward_vars_set.update(op.desc.output_arg_names())

    # Record the vars which are created during backward and is not generated by op.
    backward_vars_set = set()
    # special_op_nodes is the candidate sub-graph head node.
    special_op_nodes = set()
    for op_desc in grad_op_descs:
        input_set = set(op_desc.input_arg_names())
        # The new_vars are created during backward and is not generated by op.
        new_vars = input_set - forward_vars_set - backward_vars_set
        backward_vars_set.update(op_desc.output_arg_names())

        op_node = _create_op_node(op_desc)
        if len(new_vars) == len(input_set):
            special_op_nodes.add(op_node)

    not_need_op_descs = []
    # Start traversing all candidate sub-graph headers to check whether
    # they are connected to backward computational graphs, and if they are
    # not, list them in not_need_op_descs
    for special_op_node in special_op_nodes:
        op_list = [special_op_node]
        ready_vars = set(special_op_node.inputs)
        remove_ops = True
        candidate_ops = [special_op_node]
        while len(candidate_ops) > 0:
            op_node = candidate_ops.pop(0)
            if _all_in_set_(op_node.inputs, ready_vars):
                for out_var in op_node.outputs:
                    candidate_ops.extend(out_var.pendding_ops)
                    op_list.extend(out_var.pendding_ops)
                ready_vars.update(op_node.outputs)
            else:
                remove_ops = False
                break
        if remove_ops:
            not_need_op_descs.extend([node.op_desc for node in op_list])
600 601 602 603 604 605 606
    not_need_op_descs_set = set(not_need_op_descs)
    grad_op_descs_set = set(grad_op_descs)
    # If a backward computational graph is simply one sub-graph header, the
    # not_need_op_descs will be whole graph, this IF clause avoids it. 
    if grad_op_descs_set == not_need_op_descs_set:
        return set()
    return not_need_op_descs_set
C
chengduo 已提交
607 608


609
from .proto import framework_pb2
Y
Yang Yang 已提交
610 611 612 613


def serialize_op_decs(op_desc):
    protostr = op_desc.serialize_to_string()
M
minqiyang 已提交
614
    proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang 已提交
615 616 617
    return proto.__str__()


M
mapingshuo 已提交
618 619 620 621 622 623 624 625 626 627 628 629 630 631 632
def _append_backward_ops_with_checkpoints_(
        block, ops, target_block, no_grad_dict, grad_to_var, checkpoints):
    """
    Create grad ops with forward ops, and insert them into given block

    Args:
        block(Block): the block where forward ops are
        ops(Op): the forward operators whose forward recomputation backward ops need to be added
        target_block(Block): the block which is going to hold new generated grad ops
        no_grad_dict(dict):
            key(int) block index
            val(str): corresponding forward variable name
        checkpoints: variables that a user defined as checkpoint for forward recomputation

    Algorithms:
M
mapingshuo 已提交
633
        0) deal with forward recomputing program descs
M
mapingshuo 已提交
634 635 636 637 638
        1) find ops between checkpoints, i.e. recompute_segments
        2) go through all forward ops and induct all variables that will be hold in memory
            a. variables that are used across segments will be held in memory
            b. output of dropout op will be held in memory
            c. input variables will be held in memory
M
mapingshuo 已提交
639 640 641
        3) go through each recompute_segments, add backward ops with forward recomputation
            a. add ops in current recompute_segment as forward recomputation ops
            b. rename all non-checkpoint variables in recomputation ops
M
mapingshuo 已提交
642 643
            c. add backward ops of current recomputation ops
            d. add sum op for repetitive_outputs
M
mapingshuo 已提交
644 645
        4) remove no grad branch as it is in _remove_no_grad_branch_
        5) Note1: all appended ops' OpRole are Backward
M
mapingshuo 已提交
646 647
        6) Note2: all variables with new name should be returned so that _append_backward_vars_ can be called
        7) Note3: current forward recomputation backpropagation does not handle programs with subblock
M
mapingshuo 已提交
648
    """
M
mapingshuo 已提交
649 650

    checkpoints_name = [x.name for x in checkpoints]
651
    checkpoints_name = list(set(checkpoints_name))
M
mapingshuo 已提交
652 653
    local_block = block.program._create_block()
    buffer_block = block.program._create_block()
M
mapingshuo 已提交
654
    # 0) deal with forward recomputing program descs  
M
mapingshuo 已提交
655
    program_stat = ProgramStats(block, ops)
M
mapingshuo 已提交
656
    program_stat.modify_forward_desc_for_recompute()
M
mapingshuo 已提交
657
    program_stat.build_stats()
M
mapingshuo 已提交
658 659

    # 1) find ops between checkpoints, i.e. recompute_segments
660
    checkpoints_name = program_stat.sort_checkpoints(checkpoints_name)
M
mapingshuo 已提交
661 662
    segments = []

663
    if len(checkpoints_name) == 1:
M
mapingshuo 已提交
664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690
        # only one checkpoint
        max_op_idx = -1
        var_group = [checkpoints_name[0]]
        for name in var_group:
            if name not in program_stat.var_op_deps:
                break
            op_idx = program_stat.var_op_deps[name]["var_as_output_ops"]
            for idx in op_idx:
                max_op_idx = max(max_op_idx, idx)
        if max_op_idx > 0:
            segments.append([0, max_op_idx + 1])
    else:
        start_idx = 0
        while True:
            if start_idx >= len(checkpoints_name) - 1:
                break
            flag, min_idx, max_idx = program_stat.is_subgraph(
                [checkpoints_name[start_idx]],
                [checkpoints_name[start_idx + 1]])
            if flag:
                segments.append([min_idx, max_idx + 1])
            start_idx += 1

    if segments != [] and segments[0][0] != 0:
        recompute_segments = [[0, segments[0][0]]] + segments
    else:
        recompute_segments = segments
M
mapingshuo 已提交
691 692

    # 2) go through all forward ops and induct all variables that will be hold in memory
M
mapingshuo 已提交
693
    vars_should_be_hold = []
694
    # a. variables that are used across segments will be held in memory
M
mapingshuo 已提交
695 696 697
    for segment in recompute_segments:
        vars_should_be_hold.extend(
            program_stat.get_out_of_subgraph_vars(segment[0], segment[1]))
M
mapingshuo 已提交
698
    # b. output of dropout op will be held in memory
M
mapingshuo 已提交
699
    vars_should_be_hold.extend(program_stat.get_reserved_vars())
M
mapingshuo 已提交
700
    # c. input variables are checkpoints
M
mapingshuo 已提交
701 702 703
    vars_should_be_hold.extend(program_stat.get_input_nodes())
    vars_should_be_hold = list(set(vars_should_be_hold))

M
mapingshuo 已提交
704
    # 3) go through each recompute_segments, add backward ops with forward recomputation
M
mapingshuo 已提交
705 706 707 708 709 710 711
    grad_op_descs = []
    var_name_dict = {}

    vars_in_memory = vars_should_be_hold + checkpoints_name

    max_calculated_op_position = len(ops)
    if recompute_segments == []:
712
        # if there is no recompute segment, add backward ops like
M
mapingshuo 已提交
713
        # _append_backward_ops_ function
M
mapingshuo 已提交
714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758
        gap_ops = ops[0:max_calculated_op_position]
        for op in reversed(gap_ops):
            if op.has_attr("sub_block"):
                raise Exception("Recompute don't support ops with sub_block"
                                "invoke op: %s" %
                                _pretty_op_desc_(op.desc, "with_sub_block"))
            grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
                op.desc, cpt.to_text(no_grad_dict[block.idx]), [])
            added_descs = _add_descs_to_block(grad_op_desc, local_block)
            grad_op_descs.extend(added_descs)
            grad_to_var.update(op_grad_to_var)

    for i, segment in enumerate(recompute_segments[::-1]):
        # add grad op for ops not in any segments
        gap_ops = ops[segment[1]:max_calculated_op_position]
        max_calculated_op_position = segment[0]
        for op in reversed(gap_ops):
            if op.has_attr("sub_block"):
                raise Exception("Recompute don't support ops with sub_block"
                                "invoke op: %s" %
                                _pretty_op_desc_(op.desc, "with_sub_block"))
            grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
                op.desc, cpt.to_text(no_grad_dict[block.idx]), [])
            added_descs = _add_descs_to_block(grad_op_desc, local_block)
            grad_op_descs.extend(added_descs)
            grad_to_var.update(op_grad_to_var)

        ff_ops = ops[segment[0]:segment[1]]
        var_suffix = ".subprog_%d" % i

        for op in ff_ops:
            if op.has_attr("sub_block"):
                raise Exception("Recompute don't support ops with sub_block"
                                "invoke op: %s" %
                                _pretty_op_desc_(op.desc, "with_sub_block"))
            input_and_output_names = []
            input_and_output_names.extend(op.desc.input_arg_names())
            input_and_output_names.extend(op.desc.output_arg_names())
            for name in input_and_output_names:
                if block.var(name).persistable or name in checkpoints_name:
                    continue
                if name in vars_should_be_hold:
                    continue
                if name not in var_name_dict:
                    var_name_dict[name] = name + var_suffix
M
mapingshuo 已提交
759
        # 3.a. add ops in current recompute_segment as forward recomputation ops
M
mapingshuo 已提交
760 761 762 763
        buffer_descs = _add_needed_descs_to_block(ff_ops, buffer_block, block,
                                                  vars_in_memory)
        added_descs = _add_descs_to_block(ff_ops, local_block)

M
mapingshuo 已提交
764
        # 3.b. rename all non-checkpoint variables in recomputation ops
M
mapingshuo 已提交
765 766 767 768 769 770
        for key in var_name_dict:
            _rename_arg_(buffer_descs, key, var_name_dict[key])

        # added_descs should be in grad_op_descs because it is backward op desc
        grad_op_descs.extend(buffer_descs)

M
mapingshuo 已提交
771
        # 3.c. add backward ops of current recomputation ops
M
mapingshuo 已提交
772 773 774 775 776 777 778 779
        for op_desc in reversed(added_descs):
            grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
                op_desc, cpt.to_text(no_grad_dict[block.idx]), [])
            for key in var_name_dict:
                _rename_arg_(grad_op_desc, key, var_name_dict[key])
            grad_op_descs.extend(grad_op_desc)
            grad_to_var.update(op_grad_to_var)

M
mapingshuo 已提交
780
    # 3.d. add sum op for repetitive_outputs
781
    grad_op_descs = _addup_repetitive_outputs_(grad_op_descs, block.idx)
M
mapingshuo 已提交
782
    # 4) remove no grad branch as it is in _remove_no_grad_branch_
M
mapingshuo 已提交
783 784 785 786 787 788
    grad_op_descs = _remove_no_grad_branch_(grad_op_descs,
                                            no_grad_dict[block.idx])
    added_descs = _add_descs_to_block(grad_op_descs, target_block)
    return program_stat, checkpoints_name, vars_should_be_hold, recompute_segments


789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804
def _get_sub_block_path(sub_block, sub_block_op_desc, no_grad_set):
    """
    Get output vars in subblock which will be assigned to parent block.
    It is used to find the grad path in subblock
    """
    assert sub_block_op_desc.has_attr(
        "sub_block") and sub_block.idx == sub_block_op_desc._block_attr_id(
            "sub_block")
    # TODO(huihuangzheng): add support for recurrent op and while op
    if sub_block_op_desc.type == "conditional_block":
        sub_outputs = []
        sub_assign_to_out_ops = []
        for var in sub_block_op_desc.output_arg_names:
            for op_desc in sub_block.ops:
                if op_desc.type == "assign" and var in op_desc.output_arg_names:
                    sub_assign_to_out_ops.append(op_desc)
805 806 807 808
                    for name in op_desc.input_arg_names:
                        if sub_block.has_var(name):
                            sub_outputs.append(sub_block.var(name))

809 810 811 812 813 814 815 816 817 818
        sub_block_op_path = _find_op_path_(sub_block, sub_outputs, [],
                                           no_grad_set)
        # TODO better way than finding in list
        for op_desc in sub_assign_to_out_ops:
            if op_desc not in sub_block_op_path:
                sub_block_op_path.append(op_desc)
        return sub_block_op_path
    return sub_block.ops


819 820
def _append_backward_ops_(block,
                          ops,
F
fengjiayi 已提交
821 822 823
                          target_block,
                          no_grad_dict,
                          grad_to_var,
824 825
                          callbacks=None,
                          input_grad_names_set=None):
826 827 828 829 830
    """
    Create all grad ops, and insert them into given block

    Args:
        block(Block): the block where forward ops are
831
        ops(Op): the forward operators whose backward ops need to be added
832
        target_block(Block): the block which is going to hold new generated grad ops
833
        no_grad_dict(dict):
834 835 836 837 838
            key(int)  block index
            val(set) a set of varibale names. These varibales have no gradient
        grad_to_var(dict)(output argument):
            key(str): grad variable name
            val(str): corresponding forward variable name
C
chengduo 已提交
839 840 841 842
        callbacks(callable object): a callable object used to decorate new generated grad ops
        input_grad_names_set(set): this set is used to store the gradients' name which is
            generated by backward ops, and input_grad_names_set can help to prune the unnecessary
            backward ops.
843
    """
Y
Yang Yang 已提交
844
    if callbacks is not None:
Y
Yang Yang 已提交
845 846 847 848
        assert (isinstance(callbacks, list))
        for cb in callbacks:
            if not hasattr(cb, '__call__'):
                raise ValueError("'callback' must be a callable object.")
F
fengjiayi 已提交
849

F
fengjiayi 已提交
850
    # grad_op_descs holds created grad_op, and will be appended to target_block
F
fengjiayi 已提交
851 852
    grad_op_descs = []
    program = block.program
853 854

    # add grad_op_desc by reversed ops
855
    for op in reversed(ops):
F
fengjiayi 已提交
856 857 858
        grad_sub_block_list = []
        # If the op has its own sub-block, deal with the sub-block first
        if op.has_attr("sub_block"):
W
Wu Yi 已提交
859
            sub_block = program.block(op._block_attr_id("sub_block"))
W
Wu Yi 已提交
860
            grad_sub_block = program._create_block()
W
Wu Yi 已提交
861
            grad_sub_block._set_forward_block_idx(sub_block.idx)
862 863 864
            # see follwing comments for why set None here.
            pre_input_grad_names_set = copy.copy(input_grad_names_set)
            input_grad_names_set = None
865 866 867
            sub_block_path = _get_sub_block_path(sub_block, op,
                                                 no_grad_dict[sub_block.idx])
            _append_backward_ops_(sub_block, sub_block_path, grad_sub_block,
868 869 870
                                  no_grad_dict, grad_to_var, callbacks,
                                  input_grad_names_set)
            input_grad_names_set = pre_input_grad_names_set
Y
Yu Yang 已提交
871

W
Wu Yi 已提交
872
            program._rollback()
F
fengjiayi 已提交
873 874
            grad_sub_block_list.append(grad_sub_block.desc)

F
fengjiayi 已提交
875
        # Getting op's corresponding grad_op
F
fengjiayi 已提交
876
        grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
M
minqiyang 已提交
877
            op.desc, cpt.to_text(no_grad_dict[block.idx]), grad_sub_block_list)
Y
Yang Yu 已提交
878

879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905
        # If input_grad_names_set is not None, extend grad_op_descs only when
        # any input grad in outputs of previous grad ops.
        # But this strategy is not suited for while op for some control flow,
        # for example, for while op, the grads maybe generated in next loop.
        if input_grad_names_set is not None:
            is_append_grad = False
            for op_desc in grad_op_desc:
                input_grad_names = [
                    name for name in op_desc.input_arg_names()
                    if name.find(core.grad_var_suffix()) != -1
                ]
                # some code of gradient ops, like increment, are not very
                # standard, there is no @GRAD in these ops' inputs.
                if len(input_grad_names) == 0:
                    is_append_grad = True
                    break

                if _some_in_set_(input_grad_names, input_grad_names_set):
                    grad_op_descs.append(op_desc)
                    is_append_grad = True
                    for name in op_desc.output_arg_names():
                        input_grad_names_set.add(name)
            if is_append_grad:
                grad_to_var.update(op_grad_to_var)
        else:
            grad_op_descs.extend(grad_op_desc)
            grad_to_var.update(op_grad_to_var)
F
fengjiayi 已提交
906

M
mapingshuo 已提交
907
    # sum parameter's gradients' var given multiple var gradient
908
    grad_op_descs = _addup_repetitive_outputs_(grad_op_descs, block.idx)
F
fengjiayi 已提交
909

M
mapingshuo 已提交
910 911
    # if all outputs of the grad op are in no_grad_set, then just remove and fill zero
    # if all inputs of the grad op are in no_grad_set, just remove this op
F
fengjiayi 已提交
912 913
    grad_op_descs = _remove_no_grad_branch_(grad_op_descs,
                                            no_grad_dict[block.idx])
F
fengjiayi 已提交
914

M
mapingshuo 已提交
915
    # remove some backward ops
C
chengduo 已提交
916
    not_need_ops = _find_not_need_ops(grad_op_descs, ops, input_grad_names_set)
M
mapingshuo 已提交
917

C
chengduo 已提交
918 919 920
    grad_op_descs = [
        op_desc for op_desc in grad_op_descs if op_desc not in not_need_ops
    ]
921

F
fengjiayi 已提交
922
    # append op_desc in grad_op_descs to target_block
Y
yuyang18 已提交
923 924
    op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
    backward = core.op_proto_and_checker_maker.OpRole.Backward
F
update  
fengjiayi 已提交
925
    for op_desc in grad_op_descs:
F
fengjiayi 已提交
926 927
        new_op_desc = target_block.desc.append_op()
        new_op_desc.copy_from(op_desc)
W
Wu Yi 已提交
928
        new_op_desc._set_attr(op_role_attr_name, backward)
Y
Yang Yang 已提交
929
        grad_to_var["__current_op_desc__"] = new_op_desc
Y
Yang Yang 已提交
930 931 932 933
        if callbacks is not None:
            assert (isinstance(callbacks, list))
            for cb in callbacks:
                cb(block=target_block, context=grad_to_var)
F
update  
fengjiayi 已提交
934

F
fengjiayi 已提交
935 936

def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
937 938 939 940 941 942 943 944 945 946 947 948
    """
    Create new variables required by backward pass.

    Args:
        block(Block): the block where new variables will be created
        start_op_idx(int): Only variables required by ops in block.ops[start_op_idx : ] will be created
        grad_to_var(dict):
            key(str): grad variable name
            val(str): corresponding forward variable name
            In most cases, this dict is generated by _append_backward_ops_()
        grad_info_map(dict)(output argument):
            key(str): forward variable name
949
            val(tuple): a tuple of (str, Block), str is the corresponding grad name, Block is the block containing grad variable
950
    """
F
fengjiayi 已提交
951 952 953
    for op_idx in range(start_op_idx, block.desc.op_size()):
        op_desc = block.desc.op(op_idx)
        if op_desc.has_attr("sub_block"):
W
Wu Yi 已提交
954
            sub_block = block.program.block(op_desc._block_attr_id("sub_block"))
F
fengjiayi 已提交
955 956 957 958
            _append_backward_vars_(sub_block, 0, grad_to_var, grad_info_map)
        new_vars = set()
        # create new gradient variables
        for grad_var_name in op_desc.output_arg_names():
M
minqiyang 已提交
959 960
            if block.desc.has_var_recursive(cpt.to_bytes(
                    grad_var_name)) or grad_var_name == core.empty_var_name():
F
fengjiayi 已提交
961
                continue
M
minqiyang 已提交
962
            block.desc.var(cpt.to_bytes(grad_var_name))
F
fengjiayi 已提交
963
            new_vars.add(grad_var_name)
964
            if grad_var_name not in grad_to_var:
F
fengjiayi 已提交
965 966 967 968 969
                continue
            grad_info_map[grad_to_var[grad_var_name]] = (grad_var_name, block)
        # infer_shape and infer_type
        op_desc.infer_var_type(block.desc)
        op_desc.infer_shape(block.desc)
970

F
fengjiayi 已提交
971 972
        for arg in op_desc.output_arg_names():
            if arg in new_vars:
973
                _infer_var_data_type_shape_(arg, block)
F
update  
fengjiayi 已提交
974 975


976 977 978 979 980 981
def _rename_grad_(block, start_op_idx, grad_to_var, target_grad_map):
    var_map = copy.copy(target_grad_map)
    for op_idx in range(start_op_idx, block.desc.op_size()):
        op_desc = block.desc.op(op_idx)
        for name in op_desc.input_arg_names():
            if name in var_map:
W
Wu Yi 已提交
982
                op_desc._rename_input(name, var_map[name])
983 984

        for name in op_desc.output_arg_names():
M
mapingshuo 已提交
985 986
            if "@GRAD" not in name:
                continue
987
            if block.desc.find_var(name.encode("ascii")):
Y
Yu Yang 已提交
988
                new_name = unique_name.generate(name)
W
Wu Yi 已提交
989
                op_desc._rename_output(name, new_name)
990 991
                var_map[name] = new_name

M
minqiyang 已提交
992
    for g, ng in six.iteritems(var_map):
993 994 995 996 997 998 999 1000 1001 1002 1003
        if g in grad_to_var:
            grad_to_var[ng] = grad_to_var[g]
            grad_to_var.pop(g)


def _get_stop_gradients_(program):
    no_grad_dict = dict()
    assert isinstance(program, framework.Program)
    for block in program.blocks:
        assert isinstance(block, framework.Block)
        block_no_grad_set = set()
1004
        for var in list(block.vars.values()):
1005 1006 1007 1008 1009 1010 1011
            assert isinstance(var, framework.Variable)
            if var.stop_gradient:
                block_no_grad_set.add(_append_grad_suffix_(var.name))
        no_grad_dict[block.idx] = block_no_grad_set
    return no_grad_dict


1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022
def _get_son_parent_block_idx_dict(program, current_block_idx):

    son_parent_block_idx_dict = collections.OrderedDict()
    while current_block_idx >= 0:
        parent_block_idx = program.block(current_block_idx).parent_idx
        son_parent_block_idx_dict[current_block_idx] = parent_block_idx
        current_block_idx = parent_block_idx

    return son_parent_block_idx_dict


M
mapingshuo 已提交
1023 1024 1025 1026 1027
def append_backward(loss,
                    parameter_list=None,
                    no_grad_set=None,
                    callbacks=None,
                    checkpoints=None):
1028
    """
1029
    This function appends backward part to main_program.
F
fengjiayi 已提交
1030

1031 1032
    A complete neural network training is made up of forward and backward
    propagation. However, when we configure a network, we only need to
1033 1034
    specify its forward part. This function uses the chain rule to automatically
    generate the backward part according to the forward part.
F
fengjiayi 已提交
1035

1036 1037
    In most cases, users do not need to invoke this function manually.
    It will be automatically invoked by the optimizer's `minimize` function.
F
fengjiayi 已提交
1038

1039 1040
    Parameters:
        loss( :ref:`api_guide_Variable_en` ): The loss variable of the network.
1041 1042
        parameter_list(list[Variable|str], optional): List of Parameters or Parameter.names
                                           that need to be updated by optimizers.
1043
                                           If it is None, all parameters
F
fengjiayi 已提交
1044
                                           will be updated.
1045
                                           Default: None.
1046
        no_grad_set(set[str], optional): Variable names in the :ref:`api_guide_Block_en` 0 whose gradients
1047
                               should be ignored. All variables with
1048
                               `stop_gradient=True` from all blocks will
F
fengjiayi 已提交
1049
                               be automatically added into this set.
1050 1051
                               If this parameter is not None, the names in this set will be added to the default set.
                               Default: None.
1052
        callbacks(list[callable object], optional): List of callback functions.
1053
                                               The callbacks are used for
1054 1055 1056 1057 1058 1059
                                               doing some custom jobs during
                                               backward part building. All
                                               callable objects in it will
                                               be invoked once each time a
                                               new gradient operator is added
                                               into the program. The callable
Z
zhangchunle 已提交
1060
                                               object must have two input
1061
                                               parameters: 'block' and 'context'.
1062
                                               The 'block' is the :ref:`api_guide_Block_en` which
1063 1064 1065 1066
                                               the new gradient operator will
                                               be added to. The 'context' is a
                                               map, whose keys are gradient
                                               variable names and values are
1067
                                               corresponding original :ref:`api_guide_Variable_en` .
1068 1069 1070 1071 1072 1073
                                               In addition to this, the 'context'
                                               has another special key-value pair:
                                               the key is string '__current_op_desc__'
                                               and the value is the op_desc of the
                                               gradient operator who has just
                                               triggered the callable object.
1074
                                               Default: None.
F
fengjiayi 已提交
1075 1076

    Returns:
1077 1078
        list of tuple ( :ref:`api_guide_Variable_en` , :ref:`api_guide_Variable_en` ): Pairs of parameter and its corresponding gradients.
        The key is the parameter and the value is gradient variable.
F
fengjiayi 已提交
1079 1080 1081 1082 1083 1084 1085

    Raises:
        AssertionError: If `loss` is not an instance of Variable.

    Examples:
        .. code-block:: python

1086
            import paddle.fluid as fluid
1087 1088
            x = fluid.data(name='x', shape=[None, 13], dtype='float32')
            y = fluid.data(name='y', shape=[None, 1], dtype='float32')
L
lujun 已提交
1089 1090 1091 1092

            y_predict = fluid.layers.fc(input=x, size=1, act=None)
            loss = fluid.layers.square_error_cost(input=y_predict, label=y)

F
fengjiayi 已提交
1093 1094
            avg_loss = fluid.layers.mean(loss)
            param_grad_list = fluid.backward.append_backward(loss=avg_loss)
1095 1096 1097 1098 1099
            p_g_list1 = fluid.backward.append_backward(loss=avg_loss)  # len(p_g_list1) == 2
            p_g_list2 = fluid.backward.append_backward(loss=avg_loss, parameter_list=[p_g_list1[0][0].name])  # len(p_g_list1) == 1
            p_g_list3 = fluid.backward.append_backward(loss=avg_loss, no_grad_set=set([p_g_list1[0][0].name]))  # len(p_g_list1) == 1
            p_g_list4 = fluid.backward.append_backward(loss=avg_loss, parameter_list=[p_g_list1[0][0].name], no_grad_set=set([p_g_list1[0][0].name]))  # len(p_g_list1) == 0

1100 1101
    """
    assert isinstance(loss, framework.Variable)
Y
yuyang18 已提交
1102

Y
Fix bug  
yuyang18 已提交
1103 1104
    if loss.op is None:
        # the loss is from a cloned program. Find loss op manually.
M
mapingshuo 已提交
1105
        _find_loss_op_(loss)
Y
Fix bug  
yuyang18 已提交
1106

W
Wu Yi 已提交
1107 1108 1109
    loss.op._set_attr(core.op_proto_and_checker_maker.kOpRoleAttrName(),
                      int(core.op_proto_and_checker_maker.OpRole.Forward) |
                      int(core.op_proto_and_checker_maker.OpRole.Loss))
Y
yuyang18 已提交
1110

Y
Yang Yang 已提交
1111 1112
    if callbacks is not None:
        isinstance(callbacks, list)
Y
Yu Yang 已提交
1113

F
fengjiayi 已提交
1114
    program = loss.block.program
1115 1116 1117 1118 1119 1120 1121 1122 1123 1124
    root_block = program.block(0)
    current_block_idx = program.current_block_idx
    current_block = program.block(current_block_idx)

    is_in_control_flow = current_block_idx != 0

    # Double grad is not supported in sub-block (control flow)
    if not is_in_control_flow:
        # _appending_grad_times used for double grad
        program._appending_grad_times += 1
1125

F
fengjiayi 已提交
1126
    if no_grad_set is None:
1127 1128 1129
        no_grad_set = set()
    no_grad_set = copy.copy(no_grad_set)
    no_grad_dict = _get_stop_gradients_(program)
1130 1131
    # no_grad_set only contains vars in block 0
    # Todo(liym27): support vars in sub block
1132
    no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set)))
Y
Yu Yang 已提交
1133

1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152
    # Currently it is only to support the optimizer.minimize
    # in a switch branch, which can append_backward in a sub_block.
    # Note: while_loop is in control flow, but it makes no sense to call optimizer in while.
    # Todo: report error when it is in while_loop
    if is_in_control_flow:
        # create grad block if in switch control flow.
        target_grad_block = program._create_block(
            parent_idx=current_block.parent_idx)
        target_grad_block._set_forward_block_idx(current_block_idx)
        # after _create_block, program.current_block changes
    else:
        target_grad_block = root_block

    son_parent_block_idx_dict = _get_son_parent_block_idx_dict(
        program, current_block_idx)

    block_fwd_op_num_dict = {}  # block_id: fwd_op_num
    for idx in son_parent_block_idx_dict:
        block_fwd_op_num_dict[idx] = program.block(idx).desc.op_size()
F
fengjiayi 已提交
1153

F
fengjiayi 已提交
1154 1155
    grad_to_var = dict()

M
mapingshuo 已提交
1156
    op_desc = _create_loss_op_desc_(loss)
1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215
    target_grad_block.desc.append_op().copy_from(op_desc)

    for block_idx in son_parent_block_idx_dict:
        block = program.block(block_idx)

        block_no_grad_set = set(
            map(_strip_grad_suffix_, no_grad_dict[block_idx]))
        op_path = _find_op_path_(block, [loss], [], block_no_grad_set)

        no_grad_vars = _find_no_grad_vars(block, op_path, [loss],
                                          block_no_grad_set)

        block_no_grad_set.update(no_grad_vars)
        no_grad_dict[block_idx].update(
            list(map(_append_grad_suffix_, block_no_grad_set)))

        input_grad_names_set = None
        # For double backward, input_grad_names is used for filtering
        # some non-used gradients op(s).

        # Todo(liym27): need a better design.
        # not support double grad in control flow sub-block now.
        if not is_in_control_flow:
            if program._appending_grad_times > 1:
                input_grad_names_set = set([_append_grad_suffix_(loss.name)])

        # Todo: support _append_backward_ops_with_checkpoints_ in
        #  sub-block (control flow)
        if checkpoints != None and \
                isinstance(checkpoints, list) and \
                len(checkpoints) > 0:
            program_stat, checkpoint_names, \
            vars_should_be_hold, \
            recompute_segments = \
                _append_backward_ops_with_checkpoints_(
                    root_block,
                    op_path,
                    root_block,
                    no_grad_dict,
                    grad_to_var,
                    checkpoints)
        else:
            _append_backward_ops_(
                block,  # the block where forward ops are in
                op_path,
                target_grad_block,
                no_grad_dict,
                grad_to_var,
                callbacks,
                input_grad_names_set=input_grad_names_set)

    grad_info_map = dict()

    # if in control flow, target_grad_block is a created new block which only contains grad ops,
    # so fwd_op_num is set to 0.
    fwd_op_num = block_fwd_op_num_dict[
        current_block_idx] if not is_in_control_flow else 0

    # Because append_backward may be called multiple times,
1216 1217
    # we need rename the internal gradient variables so that they have
    # different names.
1218
    _rename_grad_(target_grad_block, fwd_op_num, grad_to_var, {})
1219

1220 1221
    _append_backward_vars_(target_grad_block, fwd_op_num, grad_to_var,
                           grad_info_map)
F
fengjiayi 已提交
1222

F
fengjiayi 已提交
1223
    program.current_block_idx = current_block_idx
W
Wu Yi 已提交
1224
    program._sync_with_cpp()
F
fengjiayi 已提交
1225

1226
    if parameter_list is not None:
1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240
        if not isinstance(parameter_list, (list, tuple, set)):
            raise TypeError(
                "The type of parameter_list argument must be list or tuple or set, but received %s."
                % (type(parameter_list)))
        parameters = []
        for i, param in enumerate(parameter_list):
            if isinstance(param, framework.Variable):
                parameters.append(param.name)
            elif isinstance(param, six.string_types):
                parameters.append(param)
            else:
                raise TypeError(
                    "The type of parameter_list's member must be paddle.fluid.Variable or str, but received %s."
                    % (type(param)))
1241
    else:
F
fengjiayi 已提交
1242
        params = program.global_block().all_parameters()
C
chengduo 已提交
1243
        parameters = [param.name for param in params if param.trainable]
1244

1245
    params_and_grads = []
1246
    op_role_var_attr_name = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
1247
    for param in parameters:
M
minqiyang 已提交
1248
        if cpt.to_text(param) not in grad_info_map:
F
fengjiayi 已提交
1249
            continue
F
update  
fengjiayi 已提交
1250
        grad_info = grad_info_map[param]
F
fengjiayi 已提交
1251
        grad_block = grad_info[1]
1252 1253 1254 1255
        if not grad_block.has_var(grad_info[0]):
            raise ValueError("grad block[{0}] did not have grad var {1}".format(
                grad_info[1], grad_info[0]))
        # Get the param var from the global block
F
fengjiayi 已提交
1256
        param_var = program.global_block().var(param)
1257
        grad_var = grad_block.var(grad_info[0])
1258 1259 1260 1261 1262
        if not is_in_control_flow:
            if loss.block.has_var(grad_info[0]):
                params_and_grads.append((param_var, grad_var))
            else:
                params_and_grads.append((param_var, None))
1263
        else:
1264
            params_and_grads.append((param_var, grad_var))
Y
yuyang18 已提交
1265 1266 1267 1268

    for p, g in params_and_grads:
        if g is None:
            continue
1269 1270 1271
        ops = grad_block.ops if is_in_control_flow else program.global_block(
        ).ops
        for op in reversed(ops):
Y
yuyang18 已提交
1272 1273 1274 1275 1276 1277 1278
            assert isinstance(op, framework.Operator)
            if g.name in op.output_arg_names:
                g.op = op
                break

        if g.op is None:
            raise ValueError("Unexpected branch")
Y
yuyang18 已提交
1279
        attr_val = [p.name, g.name]
Y
yuyang18 已提交
1280 1281
        if g.op.has_attr(op_role_var_attr_name):
            attr_val.extend(g.op.attr(op_role_var_attr_name))
W
Wu Yi 已提交
1282
        g.op._set_attr(op_role_var_attr_name, attr_val)
Y
yuyang18 已提交
1283

1284
    return params_and_grads
1285 1286 1287 1288 1289 1290 1291 1292


def _as_list(x):
    if x is None:
        return []
    return list(x) if isinstance(x, collections.Sequence) else [x]


1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342
def _is_ancestor_block(ancestor_block, block):
    prog = block.program
    ancestor_idx = ancestor_block.idx
    parent_idx = block.parent_idx

    while parent_idx != -1:
        if parent_idx == ancestor_idx:
            return True
        parent_idx = prog.block(parent_idx).parent_idx

    return False


def _get_output_names(cur_block, targets):
    """
    In `cur_block`, get output names those linked to targets.
    NOTE:
    1. `targets` can be in `cur_block`;
    Usually, `targets` is in `cur_block`. However, considering control flow,
    2. `targets` may be in sub-block but `cur_block` is an ancestor of `targets[0].block`;
    3. `targets` may be in the block which is ancestor of `cur_block`.
    """

    block = targets[0].block if targets else cur_block
    prog = cur_block.program
    if _is_ancestor_block(block, cur_block):
        return set()

    current_output_names = set([out.name for out in targets])

    # if `cur_block` is an ancestor of `targets[0].block`, run while loop
    while block.idx != cur_block.idx:
        assert block.parent_idx != -1
        parent_block = prog.block(block.parent_idx)

        parent_block_output_names = set()
        for op in reversed(block.ops):
            if _some_in_set_(op.desc.output_arg_names(), current_output_names):
                for name in op.desc.input_arg_names():
                    current_output_names.add(name)
                    if not block.desc.find_var(cpt.to_bytes(name)) \
                            and parent_block.desc.find_var(cpt.to_bytes(name)):
                        parent_block_output_names.add(name)

        block = parent_block
        current_output_names = parent_block_output_names

    return current_output_names


1343 1344 1345
def _find_no_grad_vars(block, op_path, targets, no_grad_set):
    """
    Find the vars which is not used in the program, and
1346
    those vars belong to no_grad_var.
1347
    """
1348
    output_names = _get_output_names(block, targets)
1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362
    no_grad_var = []
    for i, op in reversed(list(enumerate(op_path))):
        # If the op has sub_block, it is too complicated to find the correct no_grad_var.
        if not op.has_attr("sub_block"):
            for out_var in op.desc.output_arg_names():
                if out_var not in output_names and out_var not in op.desc.input_arg_names(
                ) and not block.vars[out_var].stop_gradient:
                    no_grad_var.append(out_var)
        for name in op.desc.input_arg_names():
            if name not in no_grad_set:
                output_names.add(name)
    return set(no_grad_var)


1363 1364 1365 1366 1367
def _find_op_path_(block, outputs, inputs, no_grad_set):
    """
    no_grad_set will also be changed
    """
    input_names = set([inp.name for inp in inputs])
1368
    output_names = _get_output_names(block, outputs)
1369 1370 1371 1372 1373 1374

    relevant_op_flags = [True] * len(block.ops)

    # All the inputs of the block are used if inputs is empty,
    if inputs:
        for i, op in enumerate(block.ops):
1375 1376 1377
            if _some_in_set_(
                    op.desc.input_arg_names(),
                    input_names) and core.has_non_empty_grad_op_maker(op.type):
1378 1379 1380 1381 1382 1383 1384
                for name in op.desc.output_arg_names():
                    if name not in no_grad_set:
                        input_names.add(name)
            else:
                relevant_op_flags[i] = False

    for i, op in reversed(list(enumerate(block.ops))):
1385 1386 1387
        if _some_in_set_(
                op.desc.output_arg_names(),
                output_names) and core.has_non_empty_grad_op_maker(op.type):
1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400
            for name in op.desc.input_arg_names():
                if name not in no_grad_set:
                    output_names.add(name)
        else:
            relevant_op_flags[i] = False

    op_path = [
        block.ops[i] for i in range(len(block.ops)) if relevant_op_flags[i]
    ]

    if inputs:
        for op in op_path:
            for name in op.desc.input_arg_names():
1401
                if name not in input_names and block.vars[name].stop_gradient:
1402 1403 1404 1405 1406 1407 1408
                    no_grad_set.add(name)

    return op_path


def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
    """
1409
    Backpropagate the gradients of targets to inputs.
1410 1411 1412 1413

    Args:
        targets(Variable|list[Variable]): The target variables
        inputs(Variable|list[Variable]): The input variables
1414 1415 1416
        target_gradients (Variable|list[Variable]|None): The gradient variables
            of targets which has the same shape with targets, If None, ones will
            be created for them.
1417 1418 1419 1420 1421
        no_grad_set(set[string]): The names of variables that have no gradients
            in Block 0. All variables with `stop_gradient=True` from all blocks
            will be automatically added.

    Return:
1422
        (list[Variable]): A list of gradients for inputs
1423 1424 1425 1426 1427 1428 1429 1430 1431
        If an input does not affect targets, the corresponding gradient variable
        will be None
    """
    targets = _as_list(targets)
    inputs = _as_list(inputs)
    target_gradients = _as_list(target_gradients)

    block = targets[0].block
    prog = block.program
1432 1433
    # increase appending gradients times
    prog._appending_grad_times += 1
1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446
    block_idx = block.idx

    if not target_gradients:
        target_gradients = [None] * len(targets)

    if len(targets) != len(target_gradients):
        raise ValueError(
            "Should have the same number of target_gradients as targets")

    if no_grad_set is None:
        no_grad_set = set()
    no_grad_set = copy.copy(no_grad_set)
    no_grad_dict = _get_stop_gradients_(prog)
1447
    no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set)))
1448 1449 1450

    fwd_op_num = block.desc.op_size()

1451 1452
    input_grad_names_set = set()

1453 1454 1455 1456 1457
    target_grad_map = {}
    for i, grad in enumerate(target_gradients):
        target = targets[i]
        if grad is None:
            grad_name = _append_grad_suffix_(target.name)
L
lvmengsi 已提交
1458 1459 1460 1461 1462
            target_shape = target.name + '_shape'
            block.desc.append_op().copy_from(
                _create_op_desc_("shape", {'Input': [target.name]},
                                 {"Out": [target_shape]}, {}))
            input_grad_names_set.add(target_shape)
L
liym27 已提交
1463
            op_desc = _create_op_desc_("fill_constant",
L
lvmengsi 已提交
1464
                                       {"ShapeTensor": [target_shape]},
1465
                                       {"Out": [grad_name]}, {
1466
                                           "shape": target.shape,
1467 1468 1469
                                           "value": 1.0,
                                           "dtype": target.dtype,
                                       })
L
liym27 已提交
1470

1471
            block.desc.append_op().copy_from(op_desc)
1472
            input_grad_names_set.add(grad_name)
1473 1474 1475 1476 1477 1478 1479 1480
        else:
            if target.block.idx != block_idx or target.block.program != prog:
                raise ValueError("all targets must be in the same block")
            if target.shape != grad.shape:
                raise ValueError(
                    "The shapes of target and grad are different: %s %s" % (
                        target.name, grad.name))
            target_grad_map[_append_grad_suffix_(target.name)] = grad.name
1481 1482 1483 1484 1485 1486
            input_grad_names_set.add(grad.name)

    # For double backward, input_grad_names is used for filter
    # some non-used gradients op.
    if prog._appending_grad_times == 1:
        input_grad_names_set = None
1487 1488 1489 1490 1491 1492 1493

    for input in inputs:
        if input.block.program != prog:
            raise "input must be in the same program as targets"

    block_no_grad_set = set(map(_strip_grad_suffix_, no_grad_dict[0]))
    op_path = _find_op_path_(block, targets, inputs, block_no_grad_set)
1494
    no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set)))
1495 1496
    grad_to_var = dict()
    grad_info_map = dict()
1497 1498 1499 1500 1501 1502 1503
    _append_backward_ops_(
        block,
        op_path,
        block,
        no_grad_dict,
        grad_to_var,
        input_grad_names_set=input_grad_names_set)
1504 1505 1506 1507 1508 1509 1510

    # Because calc_gradient may be called multiple times,
    # we need rename the internal gradient variables so that they have
    # different names.
    _rename_grad_(block, fwd_op_num, grad_to_var, target_grad_map)

    _append_backward_vars_(block, fwd_op_num, grad_to_var, grad_info_map)
W
Wu Yi 已提交
1511
    prog._sync_with_cpp()
1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526

    grad_vars = []
    for input_var in inputs:
        if input_var.name not in grad_info_map:
            grad_vars.append(None)
        else:
            grad_info = grad_info_map[input_var.name]
            grad_block = grad_info[1]
            grad_var = grad_block.var(grad_info[0])
            grad_vars.append(grad_var)

    if len(grad_vars) == 1:
        return grad_vars[0]
    else:
        return grad_vars
1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563


def gradients(targets, inputs, target_gradients=None, no_grad_set=None):
    """
    Backpropagate the gradients of targets to inputs.

    Args:
        targets (Variable|list[Variable]): The target variables.
        inputs (Variable|list[Variable]): The input variables.
        target_gradients (Variable|list[Variable]|None): The gradient variables
            of targets which has the same shape with targets, If None, ones will
            be created for them.
        no_grad_set (set[string]): The names of variables that have no gradients
            in Block 0. All variables with `stop_gradient=True` from all blocks
            will be automatically added.

    Return:
        (list[Variable]): A list of gradients for inputs
        If an input does not affect targets, the corresponding gradient variable
        will be None.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            x = fluid.layers.data(name='x', shape=[2,8,8], dtype='float32')
            x.stop_gradient=False
            y = fluid.layers.conv2d(x, 4, 1, bias_attr=False)
            y = fluid.layers.relu(y)
            y = fluid.layers.conv2d(y, 4, 1, bias_attr=False)
            y = fluid.layers.relu(y)
            z = fluid.gradients([y], x)
            print(z)
    """
    outs = calc_gradient(targets, inputs, target_gradients, no_grad_set)
    return _as_list(outs)