backward.py 53.9 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 59 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
               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:
            if op.desc.type() == "dropout":
                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
139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200

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")
201 202


203 204
def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None):
    """
205
    Traverse all ops in op_descs[begin_idx : end_idx],
206 207
    if any op has inputs/outputs named "old_name", rename it as 'new_name'
    """
F
update  
fengjiayi 已提交
208 209 210
    if begin_idx is None:
        begin_idx = 0
    if end_idx is None:
211
        end_idx = len(op_descs)
F
update  
fengjiayi 已提交
212
    for i in range(begin_idx, end_idx):
213
        op_desc = op_descs[i]
F
fengjiayi 已提交
214 215
        if isinstance(op_desc, tuple):
            op_desc = op_desc[0]
W
Wu Yi 已提交
216 217
        op_desc._rename_input(old_name, new_name)
        op_desc._rename_output(old_name, new_name)
F
update  
fengjiayi 已提交
218 219


F
fengjiayi 已提交
220
def _create_op_desc_(op_type, inputs, outputs, attrs):
221 222 223
    """
    Create a C++ OpDesc object with specified inputs, outputs and attributes.
    """
F
fengjiayi 已提交
224 225
    op_desc = core.OpDesc()
    op_desc.set_type(op_type)
M
minqiyang 已提交
226
    for para, args in six.iteritems(inputs):
227 228 229 230 231
        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
232
    for para, args in six.iteritems(outputs):
233 234 235 236 237
        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
238 239 240 241 242 243

    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
244
    for name, val in six.iteritems(attrs):
F
fengjiayi 已提交
245 246 247
        if isinstance(val, framework.Block):
            op_desc.set_block_attr(name, val.desc)
        else:
W
Wu Yi 已提交
248
            op_desc._set_attr(name, val)
F
fengjiayi 已提交
249 250 251
    return op_desc


M
mapingshuo 已提交
252 253 254 255 256 257 258 259 260 261 262 263 264 265
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


266 267 268 269
def _infer_var_data_type_(grad_var_name, block):
    """
    Infer the data type of given grad variable
    """
M
minqiyang 已提交
270 271 272 273
    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
274 275
        grad_var.set_dtype(fwd_var.dtype())
    else:
276
        grad_var.set_dtype(core.VarDesc.VarType.FP32)
F
fengjiayi 已提交
277 278


F
fengjiayi 已提交
279
def _all_in_set_(cands, s):
280 281 282
    """
    Test if all elements of 'cands' are in set 's'
    """
F
fengjiayi 已提交
283 284
    if len(cands) == 0:
        return False
F
fengjiayi 已提交
285 286 287 288 289 290
    for c in cands:
        if not c in s:
            return False
    return True


291 292 293 294 295 296
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
297 298
    literal_set = cpt.to_text(s)
    literal_cands = cpt.to_text(cands)
M
minqiyang 已提交
299 300
    for c in literal_cands:
        if c in literal_set:
301 302 303 304
            return True
    return False


F
fengjiayi 已提交
305
def _strip_grad_suffix_(name):
306
    """
M
mapingshuo 已提交
307
    Strip the grad suffix from the given variable name
308 309 310
    e.g. x@GRAD ==> x
         y@GRAD@RENAME@1 ==> y
    """
M
minqiyang 已提交
311
    name = cpt.to_text(name)
M
minqiyang 已提交
312
    pos = name.find(core.grad_var_suffix())
F
fengjiayi 已提交
313
    return name[:pos] if pos != -1 else name
F
fengjiayi 已提交
314 315 316


def _append_grad_suffix_(name):
317 318 319 320
    """
    Append grad suffix to the given variable name
    e.g. x ==> x@GRAD
    """
M
minqiyang 已提交
321
    return cpt.to_text(name) + core.grad_var_suffix()
F
fengjiayi 已提交
322 323


F
fengjiayi 已提交
324
def _addup_repetitive_outputs_(op_descs):
325 326
    """
    In backward part, an variable may be the output of more than one ops.
F
fengjiayi 已提交
327 328
    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.
329 330
    `sum_op`s are added to implement the accumulate.
    """
F
update  
fengjiayi 已提交
331 332
    pending_sum_ops = []
    var_rename_count = collections.defaultdict(int)
F
fengjiayi 已提交
333
    renamed_vars = collections.defaultdict(list)
334
    renamed_var_start_idx = collections.defaultdict(list)
F
fengjiayi 已提交
335
    for idx, op_desc in enumerate(op_descs):
F
update  
fengjiayi 已提交
336
        for var_name in op_desc.input_arg_names():
M
mapingshuo 已提交
337 338
            if "@GRAD" not in var_name:
                continue
F
fengjiayi 已提交
339
            if len(renamed_vars[var_name]) > 1:
340 341 342
                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
343
                renamed_vars[var_name] = [var_name]
F
update  
fengjiayi 已提交
344
        for param_idx, param_name in enumerate(op_desc.output_names()):
F
fengjiayi 已提交
345 346
            arg_names = op_desc.output(param_name)
            for arg_idx, var_name in enumerate(arg_names):
M
mapingshuo 已提交
347 348 349 350
                if "@GRAD" not in var_name:
                    continue
                #if "@RENAME@" in var_name:
                #    continue
F
fengjiayi 已提交
351 352 353 354 355 356 357
                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]
358
                    renamed_var_start_idx[var_name] = idx
F
fengjiayi 已提交
359 360 361 362 363 364 365
                else:
                    if len(renamed_vars[var_name]) == 1:
                        new_name = var_name + "@RENAME@" + \
                            str(var_rename_count[var_name])
                        var_rename_count[var_name] += 1
                        # rename original var_name
                        renamed_vars[var_name][0] = new_name
366 367 368 369 370 371
                        # 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
372 373
                        _rename_arg_(pending_sum_ops, var_name, new_name)

F
update  
fengjiayi 已提交
374 375 376 377 378 379 380 381 382 383 384 385 386
                        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:]

F
update  
fengjiayi 已提交
387
                    new_name = var_name + "@RENAME@" + \
F
fengjiayi 已提交
388
                        str(var_rename_count[var_name])
F
fengjiayi 已提交
389
                    var_rename_count[var_name] += 1
F
fengjiayi 已提交
390 391 392
                    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
393

M
minqiyang 已提交
394
    for var_name, inputs in six.iteritems(renamed_vars):
F
update  
fengjiayi 已提交
395
        if len(inputs) > 1:
396 397 398
            pending_sum_ops.append(
                (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]},
                                  {"use_mkldnn": False}), len(op_descs)))
F
fengjiayi 已提交
399
    # sum_op descs are sorted according to their insert position
F
update  
fengjiayi 已提交
400
    for p in reversed(pending_sum_ops):
F
fengjiayi 已提交
401 402 403 404 405 406
        op_descs.insert(p[1], p[0])

    return op_descs


def _remove_no_grad_branch_(op_descs, no_grad_set):
407 408 409 410
    """
    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
411
        2. all grad inputs of the grad op are in 'no_grad_set'
412
    """
F
fengjiayi 已提交
413 414

    def _op_can_be_removed_(op_desc, no_grad_set):
F
fengjiayi 已提交
415 416
        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
417
            return True
418 419 420 421
        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
422
            no_grad_set.update(out_arg_names)
F
fengjiayi 已提交
423 424 425
            return True
        return False

F
fengjiayi 已提交
426
    # Remove ops whose outputs are all in no_grad_dict
427 428 429 430
    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
431 432
    # Insert fill_zeros_like_op
    to_insert = []
F
fengjiayi 已提交
433
    for idx, op_desc in enumerate(op_descs):
F
fengjiayi 已提交
434
        for arg in op_desc.input_arg_names():
M
mapingshuo 已提交
435
            # arg is a gradient var name and arg should not have gradient
F
fengjiayi 已提交
436
            if core.grad_var_suffix() in arg and arg in no_grad_set:
437
                x_in = _strip_grad_suffix_(arg)
M
mapingshuo 已提交
438 439
                # the reason should be: arg can be input of another grad op
                # and the op is a not-to-remove op
440 441
                to_insert.append((_create_op_desc_(
                    "fill_zeros_like", {"X": [x_in]}, {"Out": [arg]}, {}), idx))
F
fengjiayi 已提交
442

443
    list([op_descs.insert(p[1], p[0]) for p in reversed(to_insert)])
F
fengjiayi 已提交
444 445 446 447

    return op_descs


C
chengduo 已提交
448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566
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:
        (list[core.OpDesc]): A list of OpDescs which should be pruned.
    """

    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])

    return set(not_need_op_descs)


567
from .proto import framework_pb2
Y
Yang Yang 已提交
568 569 570 571


def serialize_op_decs(op_desc):
    protostr = op_desc.serialize_to_string()
M
minqiyang 已提交
572
    proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang 已提交
573 574 575
    return proto.__str__()


M
mapingshuo 已提交
576 577 578 579 580 581 582 583 584 585 586 587 588 589 590
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
591 592 593 594 595
        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
596 597 598
        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
599 600
            c. add backward ops of current recomputation ops
            d. add sum op for repetitive_outputs
M
mapingshuo 已提交
601 602
        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
603 604
        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
605
    """
M
mapingshuo 已提交
606 607

    checkpoints_name = [x.name for x in checkpoints]
608
    checkpoints_name = list(set(checkpoints_name))
M
mapingshuo 已提交
609 610 611
    local_block = block.program._create_block()
    buffer_block = block.program._create_block()

M
mapingshuo 已提交
612
    # 1) find ops between checkpoints, i.e. recompute_segments
M
mapingshuo 已提交
613 614
    program_stat = ProgramStats(block, ops)
    program_stat.build_stats()
615
    checkpoints_name = program_stat.sort_checkpoints(checkpoints_name)
M
mapingshuo 已提交
616 617
    segments = []

618
    if len(checkpoints_name) == 1:
M
mapingshuo 已提交
619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645
        # 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
646 647

    # 2) go through all forward ops and induct all variables that will be hold in memory
M
mapingshuo 已提交
648
    vars_should_be_hold = []
649
    # a. variables that are used across segments will be held in memory
M
mapingshuo 已提交
650 651 652
    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
653
    # b. output of dropout op will be held in memory
M
mapingshuo 已提交
654
    vars_should_be_hold.extend(program_stat.get_reserved_vars())
M
mapingshuo 已提交
655
    # c. input variables are checkpoints
M
mapingshuo 已提交
656 657 658
    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
659
    # 3) go through each recompute_segments, add backward ops with forward recomputation
M
mapingshuo 已提交
660 661 662 663 664 665 666
    grad_op_descs = []
    var_name_dict = {}

    vars_in_memory = vars_should_be_hold + checkpoints_name

    max_calculated_op_position = len(ops)
    if recompute_segments == []:
667
        # if there is no recompute segment, add backward ops like
M
mapingshuo 已提交
668
        # _append_backward_ops_ function
M
mapingshuo 已提交
669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713
        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
714
        # 3.a. add ops in current recompute_segment as forward recomputation ops
M
mapingshuo 已提交
715 716 717 718
        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
719
        # 3.b. rename all non-checkpoint variables in recomputation ops
M
mapingshuo 已提交
720 721 722 723 724 725
        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
726
        # 3.c. add backward ops of current recomputation ops
M
mapingshuo 已提交
727 728 729 730 731 732 733 734
        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
735
    # 3.d. add sum op for repetitive_outputs
M
mapingshuo 已提交
736
    grad_op_descs = _addup_repetitive_outputs_(grad_op_descs)
M
mapingshuo 已提交
737
    # 4) remove no grad branch as it is in _remove_no_grad_branch_
M
mapingshuo 已提交
738 739 740 741 742 743
    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


744 745
def _append_backward_ops_(block,
                          ops,
F
fengjiayi 已提交
746 747 748
                          target_block,
                          no_grad_dict,
                          grad_to_var,
749 750
                          callbacks=None,
                          input_grad_names_set=None):
751 752 753 754 755
    """
    Create all grad ops, and insert them into given block

    Args:
        block(Block): the block where forward ops are
756
        ops(Op): the forward operators whose backward ops need to be added
757
        target_block(Block): the block which is going to hold new generated grad ops
758
        no_grad_dict(dict):
759 760 761 762 763
            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
764 765 766 767
        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.
768
    """
Y
Yang Yang 已提交
769
    if callbacks is not None:
Y
Yang Yang 已提交
770 771 772 773
        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
774

F
fengjiayi 已提交
775
    # grad_op_descs holds created grad_op, and will be appended to target_block
F
fengjiayi 已提交
776 777
    grad_op_descs = []
    program = block.program
778
    for op in reversed(ops):
F
fengjiayi 已提交
779 780 781
        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
782
            sub_block = program.block(op._block_attr_id("sub_block"))
W
Wu Yi 已提交
783
            grad_sub_block = program._create_block()
W
Wu Yi 已提交
784
            grad_sub_block._set_forward_block_idx(sub_block.idx)
785 786 787
            # see follwing comments for why set None here.
            pre_input_grad_names_set = copy.copy(input_grad_names_set)
            input_grad_names_set = None
X
Xin Pan 已提交
788
            _append_backward_ops_(sub_block, sub_block.ops, grad_sub_block,
789 790 791
                                  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
792

W
Wu Yi 已提交
793
            program._rollback()
F
fengjiayi 已提交
794 795
            grad_sub_block_list.append(grad_sub_block.desc)

F
fengjiayi 已提交
796
        # Getting op's corresponding grad_op
F
fengjiayi 已提交
797
        grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
M
minqiyang 已提交
798
            op.desc, cpt.to_text(no_grad_dict[block.idx]), grad_sub_block_list)
Y
Yang Yu 已提交
799

800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826
        # 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
827

M
mapingshuo 已提交
828 829 830
    # add grad_op_desc by reversed ops

    # sum parameter's gradients' var given multiple var gradient
F
fengjiayi 已提交
831 832
    grad_op_descs = _addup_repetitive_outputs_(grad_op_descs)

M
mapingshuo 已提交
833 834
    # 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
835 836
    grad_op_descs = _remove_no_grad_branch_(grad_op_descs,
                                            no_grad_dict[block.idx])
F
fengjiayi 已提交
837

M
mapingshuo 已提交
838
    # remove some backward ops
C
chengduo 已提交
839
    not_need_ops = _find_not_need_ops(grad_op_descs, ops, input_grad_names_set)
M
mapingshuo 已提交
840

C
chengduo 已提交
841 842 843
    grad_op_descs = [
        op_desc for op_desc in grad_op_descs if op_desc not in not_need_ops
    ]
F
fengjiayi 已提交
844
    # append op_desc in grad_op_descs to target_block
Y
yuyang18 已提交
845 846
    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
847
    for op_desc in grad_op_descs:
F
fengjiayi 已提交
848 849
        new_op_desc = target_block.desc.append_op()
        new_op_desc.copy_from(op_desc)
W
Wu Yi 已提交
850
        new_op_desc._set_attr(op_role_attr_name, backward)
Y
Yang Yang 已提交
851
        grad_to_var["__current_op_desc__"] = new_op_desc
Y
Yang Yang 已提交
852 853 854 855
        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
856

F
fengjiayi 已提交
857 858

def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
859 860 861 862 863 864 865 866 867 868 869 870
    """
    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
871
            val(tuple): a tuple of (str, Block), str is the corresponding grad name, Block is the block containing grad variable
872
    """
F
fengjiayi 已提交
873 874 875
    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
876
            sub_block = block.program.block(op_desc._block_attr_id("sub_block"))
F
fengjiayi 已提交
877 878 879 880
            _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
881 882
            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
883
                continue
M
minqiyang 已提交
884
            block.desc.var(cpt.to_bytes(grad_var_name))
F
fengjiayi 已提交
885
            new_vars.add(grad_var_name)
886
            if grad_var_name not in grad_to_var:
F
fengjiayi 已提交
887 888 889 890 891 892 893 894
                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)
        for arg in op_desc.output_arg_names():
            if arg in new_vars:
                _infer_var_data_type_(arg, block)
F
update  
fengjiayi 已提交
895 896


897 898 899 900 901 902
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
903
                op_desc._rename_input(name, var_map[name])
904 905

        for name in op_desc.output_arg_names():
M
mapingshuo 已提交
906 907
            if "@GRAD" not in name:
                continue
908
            if block.desc.find_var(name.encode("ascii")):
Y
Yu Yang 已提交
909
                new_name = unique_name.generate(name)
W
Wu Yi 已提交
910
                op_desc._rename_output(name, new_name)
911 912
                var_map[name] = new_name

M
minqiyang 已提交
913
    for g, ng in six.iteritems(var_map):
914 915 916 917 918 919 920 921 922 923 924
        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()
925
        for var in list(block.vars.values()):
926 927 928 929 930 931 932
            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


M
mapingshuo 已提交
933 934 935 936 937
def append_backward(loss,
                    parameter_list=None,
                    no_grad_set=None,
                    callbacks=None,
                    checkpoints=None):
938
    """
939
    This function appends backward part to main_program.
F
fengjiayi 已提交
940

941 942
    A complete neural network training is made up of forward and backward
    propagation. However, when we configure a network, we only need to
943 944
    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
945

946 947
    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
948

949 950 951
    Parameters:
        loss( :ref:`api_guide_Variable_en` ): The loss variable of the network.
        parameter_list(list of str, optional): Names of parameters that need
952 953
                                           to be updated by optimizers.
                                           If it is None, all parameters
F
fengjiayi 已提交
954
                                           will be updated.
955 956
                                           Default: None.
        no_grad_set(set of str, optional): Variable names in the :ref:`api_guide_Block_en` 0 whose gradients
957
                               should be ignored. All variables with
958
                               `stop_gradient=True` from all blocks will
F
fengjiayi 已提交
959
                               be automatically added into this set.
960 961 962 963
                               If this parameter is not None, the names in this set will be added to the default set.
                               Default: None.
        callbacks(list of callable object, optional): List of callback functions.
                                               The callbacks are used for
964 965 966 967 968 969 970 971
                                               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
                                               object must has two input
                                               parameters: 'block' and 'context'.
972
                                               The 'block' is the :ref:`api_guide_Block_en` which
973 974 975 976
                                               the new gradient operator will
                                               be added to. The 'context' is a
                                               map, whose keys are gradient
                                               variable names and values are
977
                                               corresponding original :ref:`api_guide_Variable_en` .
978 979 980 981 982 983
                                               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.
984
                                               Default: None.
F
fengjiayi 已提交
985 986

    Returns:
987 988
        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
989 990 991 992 993 994 995

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

    Examples:
        .. code-block:: python

996
            import paddle.fluid as fluid
997 998
            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
999 1000 1001 1002

            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
1003 1004
            avg_loss = fluid.layers.mean(loss)
            param_grad_list = fluid.backward.append_backward(loss=avg_loss)
1005 1006 1007 1008 1009
            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

1010 1011
    """
    assert isinstance(loss, framework.Variable)
Y
yuyang18 已提交
1012

Y
Fix bug  
yuyang18 已提交
1013 1014
    if loss.op is None:
        # the loss is from a cloned program. Find loss op manually.
M
mapingshuo 已提交
1015
        _find_loss_op_(loss)
Y
Fix bug  
yuyang18 已提交
1016

W
Wu Yi 已提交
1017 1018 1019
    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
1020

Y
Yang Yang 已提交
1021 1022
    if callbacks is not None:
        isinstance(callbacks, list)
Y
Yu Yang 已提交
1023

F
fengjiayi 已提交
1024
    program = loss.block.program
1025 1026
    program._appending_grad_times += 1

F
fengjiayi 已提交
1027
    if no_grad_set is None:
1028 1029 1030
        no_grad_set = set()
    no_grad_set = copy.copy(no_grad_set)
    no_grad_dict = _get_stop_gradients_(program)
1031
    no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set)))
Y
Yu Yang 已提交
1032

F
update  
fengjiayi 已提交
1033
    grad_info_map = dict()
F
fengjiayi 已提交
1034
    root_block = program.block(0)
F
fengjiayi 已提交
1035

F
fengjiayi 已提交
1036 1037
    fwd_op_num = root_block.desc.op_size()
    current_block_idx = program.current_block_idx
F
fengjiayi 已提交
1038 1039
    grad_to_var = dict()

M
mapingshuo 已提交
1040
    op_desc = _create_loss_op_desc_(loss)
1041 1042 1043 1044
    root_block.desc.append_op().copy_from(op_desc)

    block_no_grad_set = set(map(_strip_grad_suffix_, no_grad_dict[0]))
    op_path = _find_op_path_(root_block, [loss], [], block_no_grad_set)
1045 1046 1047
    no_grad_vars = _find_no_grad_vars(root_block, op_path, [loss],
                                      block_no_grad_set)
    block_no_grad_set.update(no_grad_vars)
1048
    no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set)))
1049

1050 1051 1052 1053 1054 1055
    input_grad_names_set = None
    # For double backward, input_grad_names is used for filter
    # some non-used gradients op.
    if program._appending_grad_times > 1:
        input_grad_names_set = set([_append_grad_suffix_(loss.name)])

M
mapingshuo 已提交
1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078

    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_(
            root_block,
            op_path,
            root_block,
            no_grad_dict,
            grad_to_var,
            callbacks,
            input_grad_names_set=input_grad_names_set)
1079 1080 1081 1082 1083 1084

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

F
fengjiayi 已提交
1085
    _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map)
F
fengjiayi 已提交
1086

F
fengjiayi 已提交
1087
    program.current_block_idx = current_block_idx
W
Wu Yi 已提交
1088
    program._sync_with_cpp()
F
fengjiayi 已提交
1089

1090 1091 1092
    if parameter_list is not None:
        parameters = parameter_list
    else:
F
fengjiayi 已提交
1093
        params = program.global_block().all_parameters()
C
chengduo 已提交
1094
        parameters = [param.name for param in params if param.trainable]
1095

1096 1097
    params_and_grads = []
    for param in parameters:
M
minqiyang 已提交
1098
        if cpt.to_text(param) not in grad_info_map:
F
fengjiayi 已提交
1099
            continue
F
update  
fengjiayi 已提交
1100
        grad_info = grad_info_map[param]
F
fengjiayi 已提交
1101
        grad_block = grad_info[1]
1102 1103 1104 1105
        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
1106
        param_var = program.global_block().var(param)
1107 1108 1109 1110 1111
        grad_var = grad_block.var(grad_info[0])
        if loss.block.has_var(grad_info[0]):
            params_and_grads.append((param_var, grad_var))
        else:
            params_and_grads.append((param_var, None))
Y
yuyang18 已提交
1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124

    op_role_var_attr_name = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
    for p, g in params_and_grads:
        if g is None:
            continue
        for op in reversed(program.global_block().ops):
            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
1125
        attr_val = [p.name, g.name]
Y
yuyang18 已提交
1126 1127
        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
1128
        g.op._set_attr(op_role_var_attr_name, attr_val)
Y
yuyang18 已提交
1129

1130
    return params_and_grads
1131 1132 1133 1134 1135 1136 1137 1138


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


1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158
def _find_no_grad_vars(block, op_path, targets, no_grad_set):
    """
    Find the vars which is not used in the program, and
    those var belong to no_grad_var.
    """
    output_names = set([out.name for out in targets])
    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)


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
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])
    output_names = set([out.name for out in outputs])

    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):
            if _some_in_set_(op.desc.input_arg_names(), input_names):
                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))):
        if _some_in_set_(op.desc.output_arg_names(), output_names):
            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():
1193
                if name not in input_names and block.vars[name].stop_gradient:
1194 1195 1196 1197 1198 1199 1200
                    no_grad_set.add(name)

    return op_path


def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
    """
1201
    Backpropagate the gradients of targets to inputs.
1202 1203 1204 1205

    Args:
        targets(Variable|list[Variable]): The target variables
        inputs(Variable|list[Variable]): The input variables
1206 1207 1208
        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.
1209 1210 1211 1212 1213
        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:
1214
        (list[Variable]): A list of gradients for inputs
1215 1216 1217 1218 1219 1220 1221 1222 1223
        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
1224 1225
    # increase appending gradients times
    prog._appending_grad_times += 1
1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238
    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)
1239
    no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set)))
1240 1241 1242

    fwd_op_num = block.desc.op_size()

1243 1244
    input_grad_names_set = set()

1245 1246 1247 1248 1249
    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
1250 1251 1252 1253 1254
            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
1255
            op_desc = _create_op_desc_("fill_constant",
L
lvmengsi 已提交
1256
                                       {"ShapeTensor": [target_shape]},
1257
                                       {"Out": [grad_name]}, {
1258
                                           "shape": target.shape,
1259 1260 1261
                                           "value": 1.0,
                                           "dtype": target.dtype,
                                       })
L
liym27 已提交
1262

1263
            block.desc.append_op().copy_from(op_desc)
1264
            input_grad_names_set.add(grad_name)
1265 1266 1267 1268 1269 1270 1271 1272
        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
1273 1274 1275 1276 1277 1278
            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
1279 1280 1281 1282 1283 1284 1285

    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)
1286
    no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set)))
1287 1288
    grad_to_var = dict()
    grad_info_map = dict()
1289 1290 1291 1292 1293 1294 1295
    _append_backward_ops_(
        block,
        op_path,
        block,
        no_grad_dict,
        grad_to_var,
        input_grad_names_set=input_grad_names_set)
1296 1297 1298 1299 1300 1301 1302

    # 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
1303
    prog._sync_with_cpp()
1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318

    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
1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355


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)