backward.py 37.4 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
M
minqiyang 已提交
22
from .. import compat as cpt
23
from . import unique_name
24

25
__all__ = ['append_backward', 'gradients']
26 27


28 29
def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None):
    """
30
    Traverse all ops in op_descs[begin_idx : end_idx],
31 32
    if any op has inputs/outputs named "old_name", rename it as 'new_name'
    """
F
update  
fengjiayi 已提交
33 34 35
    if begin_idx is None:
        begin_idx = 0
    if end_idx is None:
36
        end_idx = len(op_descs)
F
update  
fengjiayi 已提交
37
    for i in range(begin_idx, end_idx):
38
        op_desc = op_descs[i]
F
fengjiayi 已提交
39 40
        if isinstance(op_desc, tuple):
            op_desc = op_desc[0]
W
Wu Yi 已提交
41 42
        op_desc._rename_input(old_name, new_name)
        op_desc._rename_output(old_name, new_name)
F
update  
fengjiayi 已提交
43 44


F
fengjiayi 已提交
45
def _create_op_desc_(op_type, inputs, outputs, attrs):
46 47 48
    """
    Create a C++ OpDesc object with specified inputs, outputs and attributes.
    """
F
fengjiayi 已提交
49 50
    op_desc = core.OpDesc()
    op_desc.set_type(op_type)
M
minqiyang 已提交
51
    for para, args in six.iteritems(inputs):
52 53 54 55 56
        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
57
    for para, args in six.iteritems(outputs):
58 59 60 61 62
        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
63 64 65 66 67 68

    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
69
    for name, val in six.iteritems(attrs):
F
fengjiayi 已提交
70 71 72
        if isinstance(val, framework.Block):
            op_desc.set_block_attr(name, val.desc)
        else:
W
Wu Yi 已提交
73
            op_desc._set_attr(name, val)
F
fengjiayi 已提交
74 75 76
    return op_desc


77 78 79 80
def _infer_var_data_type_(grad_var_name, block):
    """
    Infer the data type of given grad variable
    """
M
minqiyang 已提交
81 82 83 84
    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
85 86
        grad_var.set_dtype(fwd_var.dtype())
    else:
87
        grad_var.set_dtype(core.VarDesc.VarType.FP32)
F
fengjiayi 已提交
88 89


F
fengjiayi 已提交
90
def _all_in_set_(cands, s):
91 92 93
    """
    Test if all elements of 'cands' are in set 's'
    """
F
fengjiayi 已提交
94 95
    if len(cands) == 0:
        return False
F
fengjiayi 已提交
96 97 98 99 100 101
    for c in cands:
        if not c in s:
            return False
    return True


102 103 104 105 106 107
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
108 109
    literal_set = cpt.to_text(s)
    literal_cands = cpt.to_text(cands)
M
minqiyang 已提交
110 111
    for c in literal_cands:
        if c in literal_set:
112 113 114 115
            return True
    return False


F
fengjiayi 已提交
116
def _strip_grad_suffix_(name):
117 118 119 120 121
    """
    Strip the grad suffix from the given varibale name
    e.g. x@GRAD ==> x
         y@GRAD@RENAME@1 ==> y
    """
M
minqiyang 已提交
122
    name = cpt.to_text(name)
M
minqiyang 已提交
123
    pos = name.find(core.grad_var_suffix())
F
fengjiayi 已提交
124
    return name[:pos] if pos != -1 else name
F
fengjiayi 已提交
125 126 127


def _append_grad_suffix_(name):
128 129 130 131
    """
    Append grad suffix to the given variable name
    e.g. x ==> x@GRAD
    """
M
minqiyang 已提交
132
    return cpt.to_text(name) + core.grad_var_suffix()
F
fengjiayi 已提交
133 134


F
fengjiayi 已提交
135
def _addup_repetitive_outputs_(op_descs):
136 137
    """
    In backward part, an variable may be the output of more than one ops.
F
fengjiayi 已提交
138 139
    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.
140 141
    `sum_op`s are added to implement the accumulate.
    """
F
update  
fengjiayi 已提交
142 143
    pending_sum_ops = []
    var_rename_count = collections.defaultdict(int)
F
fengjiayi 已提交
144
    renamed_vars = collections.defaultdict(list)
145
    renamed_var_start_idx = collections.defaultdict(list)
F
fengjiayi 已提交
146
    for idx, op_desc in enumerate(op_descs):
F
update  
fengjiayi 已提交
147
        for var_name in op_desc.input_arg_names():
F
fengjiayi 已提交
148
            if len(renamed_vars[var_name]) > 1:
149 150 151
                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
152
                renamed_vars[var_name] = [var_name]
F
update  
fengjiayi 已提交
153
        for param_idx, param_name in enumerate(op_desc.output_names()):
F
fengjiayi 已提交
154 155 156 157 158 159 160 161 162
            arg_names = op_desc.output(param_name)
            for arg_idx, var_name in enumerate(arg_names):
                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]
163
                    renamed_var_start_idx[var_name] = idx
F
fengjiayi 已提交
164 165 166 167 168 169 170
                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
171 172 173 174 175 176
                        # 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
177 178
                        _rename_arg_(pending_sum_ops, var_name, new_name)

F
update  
fengjiayi 已提交
179 180 181 182 183 184 185 186 187 188 189 190 191
                        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
192
                    new_name = var_name + "@RENAME@" + \
F
fengjiayi 已提交
193
                        str(var_rename_count[var_name])
F
fengjiayi 已提交
194
                    var_rename_count[var_name] += 1
F
fengjiayi 已提交
195 196 197
                    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
198

M
minqiyang 已提交
199
    for var_name, inputs in six.iteritems(renamed_vars):
F
update  
fengjiayi 已提交
200
        if len(inputs) > 1:
201 202 203
            pending_sum_ops.append(
                (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]},
                                  {"use_mkldnn": False}), len(op_descs)))
F
fengjiayi 已提交
204
    # sum_op descs are sorted according to their insert position
F
update  
fengjiayi 已提交
205
    for p in reversed(pending_sum_ops):
F
fengjiayi 已提交
206 207 208 209 210 211
        op_descs.insert(p[1], p[0])

    return op_descs


def _remove_no_grad_branch_(op_descs, no_grad_set):
212 213 214 215
    """
    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
216
        2. all grad inputs of the grad op are in 'no_grad_set'
217
    """
F
fengjiayi 已提交
218 219

    def _op_can_be_removed_(op_desc, no_grad_set):
F
fengjiayi 已提交
220 221
        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
222
            return True
223 224 225 226
        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
227
            no_grad_set.update(out_arg_names)
F
fengjiayi 已提交
228 229 230
            return True
        return False

F
fengjiayi 已提交
231
    # Remove ops whose outputs are all in no_grad_dict
232 233 234 235
    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
236 237
    # Insert fill_zeros_like_op
    to_insert = []
F
fengjiayi 已提交
238
    for idx, op_desc in enumerate(op_descs):
F
fengjiayi 已提交
239
        for arg in op_desc.input_arg_names():
F
fengjiayi 已提交
240
            if core.grad_var_suffix() in arg and arg in no_grad_set:
241
                x_in = _strip_grad_suffix_(arg)
242 243
                to_insert.append((_create_op_desc_(
                    "fill_zeros_like", {"X": [x_in]}, {"Out": [arg]}, {}), idx))
F
fengjiayi 已提交
244

245
    list([op_descs.insert(p[1], p[0]) for p in reversed(to_insert)])
F
fengjiayi 已提交
246 247 248 249

    return op_descs


C
chengduo 已提交
250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368
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)


369
from .proto import framework_pb2
Y
Yang Yang 已提交
370 371 372 373


def serialize_op_decs(op_desc):
    protostr = op_desc.serialize_to_string()
M
minqiyang 已提交
374
    proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang 已提交
375 376 377
    return proto.__str__()


378 379
def _append_backward_ops_(block,
                          ops,
F
fengjiayi 已提交
380 381 382
                          target_block,
                          no_grad_dict,
                          grad_to_var,
383 384
                          callbacks=None,
                          input_grad_names_set=None):
385 386 387 388 389
    """
    Create all grad ops, and insert them into given block

    Args:
        block(Block): the block where forward ops are
390
        ops(Op): the forward operators whose backward ops need to be added
391
        target_block(Block): the block which is going to hold new generated grad ops
392
        no_grad_dict(dict):
393 394 395 396 397
            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
398 399 400 401
        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.
402
    """
Y
Yang Yang 已提交
403
    if callbacks is not None:
Y
Yang Yang 已提交
404 405 406 407
        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
408

F
fengjiayi 已提交
409
    # grad_op_descs holds created grad_op, and will be appended to target_block
F
fengjiayi 已提交
410 411
    grad_op_descs = []
    program = block.program
412
    for op in reversed(ops):
F
fengjiayi 已提交
413 414 415
        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
416
            sub_block = program.block(op._block_attr_id("sub_block"))
W
Wu Yi 已提交
417
            grad_sub_block = program._create_block()
W
Wu Yi 已提交
418
            grad_sub_block._set_forward_block_idx(sub_block.idx)
419 420 421
            # 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
422
            _append_backward_ops_(sub_block, sub_block.ops, grad_sub_block,
423 424 425
                                  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
426

W
Wu Yi 已提交
427
            program._rollback()
F
fengjiayi 已提交
428 429
            grad_sub_block_list.append(grad_sub_block.desc)

F
fengjiayi 已提交
430
        # Getting op's corresponding grad_op
F
fengjiayi 已提交
431
        grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
M
minqiyang 已提交
432
            op.desc, cpt.to_text(no_grad_dict[block.idx]), grad_sub_block_list)
Y
Yang Yu 已提交
433

434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460
        # 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
461 462 463 464 465

    grad_op_descs = _addup_repetitive_outputs_(grad_op_descs)

    grad_op_descs = _remove_no_grad_branch_(grad_op_descs,
                                            no_grad_dict[block.idx])
F
fengjiayi 已提交
466

C
chengduo 已提交
467 468 469 470
    not_need_ops = _find_not_need_ops(grad_op_descs, ops, input_grad_names_set)
    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
471
    # append op_desc in grad_op_descs to target_block
Y
yuyang18 已提交
472 473
    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
474
    for op_desc in grad_op_descs:
F
fengjiayi 已提交
475 476
        new_op_desc = target_block.desc.append_op()
        new_op_desc.copy_from(op_desc)
W
Wu Yi 已提交
477
        new_op_desc._set_attr(op_role_attr_name, backward)
Y
Yang Yang 已提交
478
        grad_to_var["__current_op_desc__"] = new_op_desc
Y
Yang Yang 已提交
479 480 481 482
        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
483

F
fengjiayi 已提交
484 485

def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
486 487 488 489 490 491 492 493 494 495 496 497
    """
    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
498
            val(tuple): a tuple of (str, Block), str is the corresponding grad name, Block is the block containing grad variable
499
    """
F
fengjiayi 已提交
500 501 502
    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
503
            sub_block = block.program.block(op_desc._block_attr_id("sub_block"))
F
fengjiayi 已提交
504 505 506 507
            _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
508 509
            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
510
                continue
M
minqiyang 已提交
511
            block.desc.var(cpt.to_bytes(grad_var_name))
F
fengjiayi 已提交
512
            new_vars.add(grad_var_name)
513
            if grad_var_name not in grad_to_var:
F
fengjiayi 已提交
514 515 516 517 518 519 520 521
                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
522 523


524 525 526 527 528 529
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
530
                op_desc._rename_input(name, var_map[name])
531 532 533

        for name in op_desc.output_arg_names():
            if block.desc.find_var(name.encode("ascii")):
Y
Yu Yang 已提交
534
                new_name = unique_name.generate(name)
W
Wu Yi 已提交
535
                op_desc._rename_output(name, new_name)
536 537
                var_map[name] = new_name

M
minqiyang 已提交
538
    for g, ng in six.iteritems(var_map):
539 540 541 542 543 544 545 546 547 548 549
        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()
550
        for var in list(block.vars.values()):
551 552 553 554 555 556 557
            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


Y
Yang Yang 已提交
558 559
def append_backward(loss, parameter_list=None, no_grad_set=None,
                    callbacks=None):
560
    """
F
fengjiayi 已提交
561 562
    Append backward part to main_program.

563 564 565
    A complete neural network training is made up of forward and backward
    propagation. However, when we configure a network, we only need to
    specify its forwrd part. The backward part is generated automatically
F
fengjiayi 已提交
566 567
    according to the forward part by this function.

568
    In most cases, users do not need to invoke this function manually. It
F
fengjiayi 已提交
569
    will be automatically invoked by the optimizer's `minimize` function.
F
fengjiayi 已提交
570 571

    Args:
F
fengjiayi 已提交
572
        loss(Variable): The loss variable of the network.
573 574 575
        parameter_list(list[string]|None): Names of parameters that need
                                           to be updated by optimizers.
                                           If it is None, all parameters
F
fengjiayi 已提交
576 577
                                           will be updated.
                                           Default: None
578 579
        no_grad_set(set|None): Variables in the Block 0 whose gradients
                               should be ignored. All variables with
580
                               `stop_gradient=True` from all blocks will
F
fengjiayi 已提交
581 582
                               be automatically added into this set.
                               Default: None
583 584 585 586 587 588 589 590 591 592 593 594 595 596
        callbacks(list[callable object]|None): The callbacks are used for
                                               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'.
                                               The 'block' is the block which
                                               the new gradient operator will
                                               be added to. The 'context' is a
                                               map, whose keys are gradient
                                               variable names and values are
F
fengjiayi 已提交
597
                                               corresponding original variables.
598 599 600 601 602 603
                                               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.
F
fengjiayi 已提交
604 605

    Returns:
606 607
        list[(Variable,Variable)]: Pairs of parameter and its
        corresponding gradients. The key is the parameter and the
F
fengjiayi 已提交
608 609 610 611 612 613 614 615
        value is gradient variable.

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

    Examples:
        .. code-block:: python

F
fengjiayi 已提交
616
            # network configuration code
L
lujun 已提交
617
            # loss from ...
618
            import paddle.fluid as fluid
L
lujun 已提交
619 620 621 622 623 624
            x = fluid.layers.data(name='x', shape=[13], dtype='float32')
            y = fluid.layers.data(name='y', shape=[1], dtype='float32')

            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
625 626
            avg_loss = fluid.layers.mean(loss)
            param_grad_list = fluid.backward.append_backward(loss=avg_loss)
627 628
    """
    assert isinstance(loss, framework.Variable)
Y
yuyang18 已提交
629

Y
Fix bug  
yuyang18 已提交
630 631 632 633 634 635 636 637 638 639 640
    if loss.op is None:
        # the loss is from a cloned program. Find loss op manually.
        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")

W
Wu Yi 已提交
641 642 643
    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
644

Y
Yang Yang 已提交
645 646
    if callbacks is not None:
        isinstance(callbacks, list)
Y
Yu Yang 已提交
647

F
fengjiayi 已提交
648
    program = loss.block.program
649 650
    program._appending_grad_times += 1

F
fengjiayi 已提交
651
    if no_grad_set is None:
652 653 654
        no_grad_set = set()
    no_grad_set = copy.copy(no_grad_set)
    no_grad_dict = _get_stop_gradients_(program)
655
    no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set)))
Y
Yu Yang 已提交
656

F
update  
fengjiayi 已提交
657
    grad_info_map = dict()
F
fengjiayi 已提交
658
    root_block = program.block(0)
F
fengjiayi 已提交
659

F
fengjiayi 已提交
660 661
    fwd_op_num = root_block.desc.op_size()
    current_block_idx = program.current_block_idx
F
fengjiayi 已提交
662 663
    grad_to_var = dict()

Y
yuyang18 已提交
664
    op_desc = _create_op_desc_(
X
Xin Pan 已提交
665 666 667 668 669
        "fill_constant",
        {},
        {"Out": [_append_grad_suffix_(loss.name)]},
        {
            "shape": [1],  # TODO(panyx0718): This can be loss.shape.
Y
yuyang18 已提交
670 671 672 673 674 675 676
            "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),
        })
677 678 679 680
    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)
681 682 683
    no_grad_vars = _find_no_grad_vars(root_block, op_path, [loss],
                                      block_no_grad_set)
    block_no_grad_set.update(no_grad_vars)
684
    no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set)))
685

686 687 688 689 690 691 692 693 694 695 696 697 698 699
    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)])

    _append_backward_ops_(
        root_block,
        op_path,
        root_block,
        no_grad_dict,
        grad_to_var,
        callbacks,
        input_grad_names_set=input_grad_names_set)
700 701 702 703 704 705

    # 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
706
    _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map)
F
fengjiayi 已提交
707

F
fengjiayi 已提交
708
    program.current_block_idx = current_block_idx
W
Wu Yi 已提交
709
    program._sync_with_cpp()
F
fengjiayi 已提交
710

711 712 713
    if parameter_list is not None:
        parameters = parameter_list
    else:
F
fengjiayi 已提交
714
        params = program.global_block().all_parameters()
C
chengduo 已提交
715
        parameters = [param.name for param in params if param.trainable]
716

717 718
    params_and_grads = []
    for param in parameters:
M
minqiyang 已提交
719
        if cpt.to_text(param) not in grad_info_map:
F
fengjiayi 已提交
720
            continue
F
update  
fengjiayi 已提交
721
        grad_info = grad_info_map[param]
F
fengjiayi 已提交
722
        grad_block = grad_info[1]
723 724 725 726
        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
727
        param_var = program.global_block().var(param)
728 729 730 731 732
        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
733 734 735 736 737 738 739 740 741 742 743 744 745

    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
746
        attr_val = [p.name, g.name]
Y
yuyang18 已提交
747 748
        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
749
        g.op._set_attr(op_role_var_attr_name, attr_val)
Y
yuyang18 已提交
750

751
    return params_and_grads
752 753 754 755 756 757 758 759


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


760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779
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)


780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813
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():
814
                if name not in input_names and block.vars[name].stop_gradient:
815 816 817 818 819 820 821
                    no_grad_set.add(name)

    return op_path


def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
    """
822
    Backpropagate the gradients of targets to inputs.
823 824 825 826

    Args:
        targets(Variable|list[Variable]): The target variables
        inputs(Variable|list[Variable]): The input variables
827 828 829
        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.
830 831 832 833 834
        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:
835
        (list[Variable]): A list of gradients for inputs
836 837 838 839 840 841 842 843 844
        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
845 846
    # increase appending gradients times
    prog._appending_grad_times += 1
847 848 849 850 851 852 853 854 855 856 857 858 859
    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)
860
    no_grad_dict[0].update(list(map(_append_grad_suffix_, no_grad_set)))
861 862 863

    fwd_op_num = block.desc.op_size()

864 865
    input_grad_names_set = set()

866 867 868 869 870 871 872 873 874 875 876 877 878 879 880
    target_grad_map = {}
    for i, grad in enumerate(target_gradients):
        target = targets[i]
        if grad is None:
            grad_name = _append_grad_suffix_(target.name)
            op_desc = _create_op_desc_("fill_constant_batch_size_like",
                                       {"Input": [target.name]},
                                       {"Out": [grad_name]}, {
                                           "shape": target.shape,
                                           "value": 1.0,
                                           "dtype": target.dtype,
                                           'input_dim_idx': 0,
                                           'output_dim_idx': 0
                                       })
            block.desc.append_op().copy_from(op_desc)
881
            input_grad_names_set.add(grad_name)
882 883 884 885 886 887 888 889
        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
890 891 892 893 894 895
            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
896 897 898 899 900 901 902

    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)
903
    no_grad_dict[0].update(list(map(_append_grad_suffix_, block_no_grad_set)))
904 905
    grad_to_var = dict()
    grad_info_map = dict()
906 907 908 909 910 911 912
    _append_backward_ops_(
        block,
        op_path,
        block,
        no_grad_dict,
        grad_to_var,
        input_grad_names_set=input_grad_names_set)
913 914 915 916 917 918 919

    # 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
920
    prog._sync_with_cpp()
921 922 923 924 925 926 927 928 929 930 931 932 933 934 935

    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
936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972


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)