backward.py 28.9 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
from paddle.fluid import framework as framework
F
update  
fengjiayi 已提交
16
from . import core
F
update  
fengjiayi 已提交
17
import collections
18
import copy
Y
Yu Yang 已提交
19
import unique_name
20

21 22 23 24
__all__ = [
    'append_backward',
    'calc_gradient',
]
25 26


27 28
def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None):
    """
29
    Traverse all ops in op_descs[begin_idx : end_idx],
30 31
    if any op has inputs/outputs named "old_name", rename it as 'new_name'
    """
F
update  
fengjiayi 已提交
32 33 34
    if begin_idx is None:
        begin_idx = 0
    if end_idx is None:
35
        end_idx = len(op_descs)
F
update  
fengjiayi 已提交
36
    for i in range(begin_idx, end_idx):
37
        op_desc = op_descs[i]
F
fengjiayi 已提交
38 39 40 41
        if isinstance(op_desc, tuple):
            op_desc = op_desc[0]
        op_desc.rename_input(old_name, new_name)
        op_desc.rename_output(old_name, new_name)
F
update  
fengjiayi 已提交
42 43


F
fengjiayi 已提交
44
def _create_op_desc_(op_type, inputs, outputs, attrs):
45 46 47
    """
    Create a C++ OpDesc object with specified inputs, outputs and attributes.
    """
F
fengjiayi 已提交
48 49 50 51 52 53
    op_desc = core.OpDesc()
    op_desc.set_type(op_type)
    for para, args in inputs.iteritems():
        op_desc.set_input(para, args)
    for para, args in outputs.iteritems():
        op_desc.set_output(para, args)
Y
yuyang18 已提交
54 55 56 57 58 59

    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
F
fengjiayi 已提交
60 61 62 63 64 65 66 67
    for name, val in attrs.iteritems():
        if isinstance(val, framework.Block):
            op_desc.set_block_attr(name, val.desc)
        else:
            op_desc.set_attr(name, val)
    return op_desc


68 69 70 71 72 73
def _infer_var_data_type_(grad_var_name, block):
    """
    Infer the data type of given grad variable
    """
    grad_var = block.desc.find_var(grad_var_name.encode("ascii"))
    fwd_name = _strip_grad_suffix_(grad_var_name.encode("ascii"))
F
fengjiayi 已提交
74 75 76 77
    if block.desc.has_var_recursive(fwd_name):
        fwd_var = block.desc.find_var_recursive(fwd_name.encode("ascii"))
        grad_var.set_dtype(fwd_var.dtype())
    else:
78
        grad_var.set_dtype(core.VarDesc.VarType.FP32)
F
fengjiayi 已提交
79 80


F
fengjiayi 已提交
81
def _all_in_set_(cands, s):
82 83 84
    """
    Test if all elements of 'cands' are in set 's'
    """
F
fengjiayi 已提交
85 86
    if len(cands) == 0:
        return False
F
fengjiayi 已提交
87 88 89 90 91 92
    for c in cands:
        if not c in s:
            return False
    return True


93 94 95 96 97 98 99 100 101 102 103 104
def _some_in_set_(cands, s):
    """
    Test if some elements of 'cands' are in set 's'
    """
    if len(cands) == 0:
        return False
    for c in cands:
        if c in s:
            return True
    return False


F
fengjiayi 已提交
105
def _strip_grad_suffix_(name):
106 107 108 109 110
    """
    Strip the grad suffix from the given varibale name
    e.g. x@GRAD ==> x
         y@GRAD@RENAME@1 ==> y
    """
F
fengjiayi 已提交
111 112
    pos = name.find(core.grad_var_suffix())
    return name[:pos] if pos != -1 else name
F
fengjiayi 已提交
113 114 115


def _append_grad_suffix_(name):
116 117 118 119
    """
    Append grad suffix to the given variable name
    e.g. x ==> x@GRAD
    """
F
fengjiayi 已提交
120 121 122
    return name + core.grad_var_suffix()


F
fengjiayi 已提交
123
def _addup_repetitive_outputs_(op_descs):
124 125 126 127 128
    """
    In backward part, an variable may be the output of more than one ops.
    In this case, the variable should be the accumulation of all the outputs.
    `sum_op`s are added to implement the accumulate.
    """
F
update  
fengjiayi 已提交
129 130
    pending_sum_ops = []
    var_rename_count = collections.defaultdict(int)
F
fengjiayi 已提交
131 132
    renamed_vars = collections.defaultdict(list)
    for idx, op_desc in enumerate(op_descs):
F
update  
fengjiayi 已提交
133
        for var_name in op_desc.input_arg_names():
F
fengjiayi 已提交
134
            if len(renamed_vars[var_name]) > 1:
135 136 137
                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
138
                renamed_vars[var_name] = [var_name]
F
fengjiayi 已提交
139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
        for param_name in op_desc.output_names():
            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]
                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
                        _rename_arg_(op_descs, var_name, new_name, 0, idx)
                        _rename_arg_(pending_sum_ops, var_name, new_name)

F
update  
fengjiayi 已提交
159
                    new_name = var_name + "@RENAME@" + \
F
fengjiayi 已提交
160
                        str(var_rename_count[var_name])
F
fengjiayi 已提交
161
                    var_rename_count[var_name] += 1
F
fengjiayi 已提交
162 163 164
                    arg_names[arg_idx] = new_name
                    op_desc.set_output(param_name, arg_names)
                    renamed_vars[var_name].append(new_name)
F
fengjiayi 已提交
165
    for var_name, inputs in renamed_vars.iteritems():
F
update  
fengjiayi 已提交
166
        if len(inputs) > 1:
167 168 169
            pending_sum_ops.append(
                (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]},
                                  {"use_mkldnn": False}), len(op_descs)))
F
fengjiayi 已提交
170
    # sum_op descs are sorted according to their insert position
F
update  
fengjiayi 已提交
171
    for p in reversed(pending_sum_ops):
F
fengjiayi 已提交
172 173 174 175 176 177
        op_descs.insert(p[1], p[0])

    return op_descs


def _remove_no_grad_branch_(op_descs, no_grad_set):
178 179 180 181
    """
    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
182
        2. all grad inputs of the grad op are in 'no_grad_set'
183
    """
F
fengjiayi 已提交
184 185

    def _op_can_be_removed_(op_desc, no_grad_set):
F
fengjiayi 已提交
186 187
        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
188 189 190 191
            return True
        if _all_in_set_(
                filter(lambda name: name.find(core.grad_var_suffix()) != -1,
                       op_desc.input_arg_names()), no_grad_set):
F
fengjiayi 已提交
192
            no_grad_set.update(out_arg_names)
F
fengjiayi 已提交
193 194 195
            return True
        return False

F
fengjiayi 已提交
196 197
    # Remove ops whose outputs are all in no_grad_dict
    op_descs = filter(
F
fengjiayi 已提交
198
        lambda op_desc: not _op_can_be_removed_(op_desc, no_grad_set), op_descs)
F
fengjiayi 已提交
199 200
    # Insert fill_zeros_like_op
    to_insert = []
F
fengjiayi 已提交
201
    for idx, op_desc in enumerate(op_descs):
F
fengjiayi 已提交
202
        for arg in op_desc.input_arg_names():
F
fengjiayi 已提交
203 204 205
            if core.grad_var_suffix() in arg and arg in no_grad_set:
                to_insert.append((_create_op_desc_("fill_zeros_like", {
                    "X": [_strip_grad_suffix_(arg)]
206
                }, {"Out": [arg]}, {}), idx))
F
fengjiayi 已提交
207 208 209 210 211 212

    map(lambda p: op_descs.insert(p[1], p[0]), reversed(to_insert))

    return op_descs


Y
Yang Yang 已提交
213 214 215 216 217 218 219 220 221
import proto.framework_pb2 as framework_pb2


def serialize_op_decs(op_desc):
    protostr = op_desc.serialize_to_string()
    proto = framework_pb2.OpDesc.FromString(str(protostr))
    return proto.__str__()


222 223 224 225 226 227 228 229 230 231
def _callback_lookup_(op):
    """
    Only used in _append_backward_ops_
    Build and returns a callback function for certain op. For example

    parallel_do:           AllReduce

    :param op:
    :return: callback function
    """
Y
Yang Yang 已提交
232
    if op.type == 'parallel_do' and op.attr('use_nccl'):
Q
qiaolongfei 已提交
233
        all_vars = op.block.vars
234
        param_names = set(op.input('parameters'))
Q
qiaolongfei 已提交
235 236
        param_names = filter(lambda name: all_vars[name].stop_gradient is False,
                             param_names)
237 238 239
        param_grad_names = [n + "@GRAD" for n in param_names]

        class ParallelDoCallBack(object):
Y
Yang Yang 已提交
240
            def __init__(self, param_grad_names, parallel_scopes_name):
241 242
                self.has_inserted_nccl_init = False
                self.param_grad_names = param_grad_names
Y
Yang Yang 已提交
243
                self.parallel_scopes_name = parallel_scopes_name
244 245

            def __call__(self, block, context):
Y
Yang Yang 已提交
246
                if not self.has_inserted_nccl_init:
Y
Yang Yang 已提交
247
                    op_desc = _create_op_desc_(
Y
Yang Yang 已提交
248 249
                        "ncclInit",
                        {"parallel_scopes": self.parallel_scopes_name},
Y
Yang Yang 已提交
250 251 252
                        {"Communicator": ['nccl_com__do_not_change_']}, {})
                    block.program.global_block().desc.append_op().copy_from(
                        op_desc)
Y
Yang Yang 已提交
253 254 255 256 257
                    self.has_inserted_nccl_init = True

                current_op_desc = context["__current_op_desc__"]
                for o_param in current_op_desc.output_names():
                    for o_argu in current_op_desc.output(o_param):
258
                        if o_argu in self.param_grad_names:
Y
Yang Yang 已提交
259 260
                            allreduce_out_name = o_argu + "__nccl_all_reduce__"
                            op_desc = _create_op_desc_(
C
chengduoZH 已提交
261 262
                                "ncclReduce",
                                {
Y
Yang Yang 已提交
263
                                    "X": [o_argu],
Y
Yang Yang 已提交
264 265
                                    "Communicator":
                                    ['nccl_com__do_not_change_']
C
chengduoZH 已提交
266 267 268 269
                                },
                                {"Out": [allreduce_out_name]},
                                {"reduction": "ncclSum",
                                 "root": 0}, )
Y
Yang Yang 已提交
270 271 272 273 274 275
                            block.desc.append_op().copy_from(op_desc)

                            op_desc = _create_op_desc_(
                                "assign", {"X": [allreduce_out_name]},
                                {"Out": [o_argu]}, {})
                            block.desc.append_op().copy_from(op_desc)
276

Y
Yang Yang 已提交
277 278
        return ParallelDoCallBack(param_grad_names,
                                  op.output("parallel_scopes"))
279 280 281 282
    else:
        return None


283 284
def _append_backward_ops_(block,
                          ops,
F
fengjiayi 已提交
285 286 287
                          target_block,
                          no_grad_dict,
                          grad_to_var,
Y
Yang Yang 已提交
288
                          callbacks=None):
289 290 291 292 293
    """
    Create all grad ops, and insert them into given block

    Args:
        block(Block): the block where forward ops are
294
        ops(Op): the forward operators whose backward ops need to be added
295
        target_block(Block): the block which is going to hold new generated grad ops
296
        no_grad_dict(dict):
297 298 299 300 301
            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
F
fengjiayi 已提交
302
        callback(callable object): a callable object used to decorate new generated grad ops
303
    """
Y
Yang Yang 已提交
304
    if callbacks is not None:
Y
Yang Yang 已提交
305 306 307 308
        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
309

F
fengjiayi 已提交
310
    # grad_op_descs holds created grad_op, and will be appended to target_block
F
fengjiayi 已提交
311 312
    grad_op_descs = []
    program = block.program
313
    for op in reversed(ops):
F
fengjiayi 已提交
314 315 316 317
        grad_sub_block_list = []
        # If the op has its own sub-block, deal with the sub-block first
        if op.has_attr("sub_block"):
            sub_block = program.block(op.block_attr("sub_block"))
Y
Yu Yang 已提交
318 319
            grad_sub_block = program.create_block()
            grad_sub_block.set_forward_block_idx(sub_block.idx)
Y
Yang Yang 已提交
320 321 322 323 324 325 326 327
            cb = _callback_lookup_(op)
            if cb is not None:
                if callbacks is None:
                    new_callbacks = [cb]
                else:
                    new_callbacks = callbacks + [_callback_lookup_(op)]
                _append_backward_ops_(sub_block, sub_block.ops, grad_sub_block,
                                      no_grad_dict, grad_to_var, new_callbacks)
Y
Yang Yang 已提交
328
            else:
Y
Yang Yang 已提交
329 330
                _append_backward_ops_(sub_block, sub_block.ops, grad_sub_block,
                                      no_grad_dict, grad_to_var, callbacks)
Y
Yu Yang 已提交
331 332

            program.rollback()
F
fengjiayi 已提交
333 334
            grad_sub_block_list.append(grad_sub_block.desc)

F
fengjiayi 已提交
335
        # Getting op's corresponding grad_op
F
fengjiayi 已提交
336 337
        grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
            op.desc, no_grad_dict[block.idx], grad_sub_block_list)
Y
Yang Yu 已提交
338

F
fengjiayi 已提交
339 340 341 342 343 344 345
        grad_op_descs.extend(grad_op_desc)
        grad_to_var.update(op_grad_to_var)

    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
346

F
fengjiayi 已提交
347
    # append op_desc in grad_op_descs to target_block
Y
yuyang18 已提交
348 349
    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
350
    for op_desc in grad_op_descs:
F
fengjiayi 已提交
351 352
        new_op_desc = target_block.desc.append_op()
        new_op_desc.copy_from(op_desc)
Y
yuyang18 已提交
353
        new_op_desc.set_attr(op_role_attr_name, backward)
Y
Yang Yang 已提交
354
        grad_to_var["__current_op_desc__"] = new_op_desc
Y
Yang Yang 已提交
355 356 357 358
        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
359

F
fengjiayi 已提交
360 361

def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
362 363 364 365 366 367 368 369 370 371 372 373
    """
    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
374
            val(tuple): a tuple of (str, Block), str is the corresponding grad name, Block is the block containing grad variable
375
    """
F
fengjiayi 已提交
376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395
    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"):
            sub_block = block.program.block(op_desc.block_attr("sub_block"))
            _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():
            grad_var_name = grad_var_name.encode("ascii")
            if block.desc.has_var_recursive(
                    grad_var_name) or grad_var_name == core.empty_var_name():
                continue
            block.desc.var(grad_var_name)
            new_vars.add(grad_var_name)
            if not grad_to_var.has_key(grad_var_name):
                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)
Y
Yang Yang 已提交
396 397 398
        # ncclInit dones't need to set data_type
        if op_desc.type() == 'ncclInit':
            continue
F
fengjiayi 已提交
399 400 401
        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
402 403


404 405 406 407 408 409 410 411 412 413
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:
                op_desc.rename_input(name, var_map[name])

        for name in op_desc.output_arg_names():
            if block.desc.find_var(name.encode("ascii")):
Y
Yu Yang 已提交
414
                new_name = unique_name.generate(name)
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437
                op_desc.rename_output(name, new_name)
                var_map[name] = new_name

    for g, ng in var_map.iteritems():
        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()
        for var in block.vars.itervalues():
            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
438 439
def append_backward(loss, parameter_list=None, no_grad_set=None,
                    callbacks=None):
440
    """
F
fengjiayi 已提交
441 442 443 444 445 446 447 448 449
    Append backward part to main_program.

    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 
    according to the forward part by this function.

    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
450 451

    Args:
F
fengjiayi 已提交
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
        loss(Variable): The loss variable of the network.
        parameter_list(list[string]|None): Names of parameters that need 
                                           to be updated by optimizers. 
                                           If it is None, all parameters 
                                           will be updated.
                                           Default: None
        no_grad_set(set|None): Variables in the Block 0 whose gradients 
                               should be ignored. All variables with 
                               `step_gradient=True` from all blocks will 
                               be automatically added into this set.
                               Default: None
        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 
                                               corresponding original variables.
                                               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. 

    Returns:
        list[(Variable,Variable)]: Pairs of parameter and its 
        corresponding gradients. The key is the parameter and the 
        value is gradient variable.

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

    Examples:
        .. code-block:: python

F
fengjiayi 已提交
496 497 498 499
            # network configuration code
            # ...
            avg_loss = fluid.layers.mean(loss)
            param_grad_list = fluid.backward.append_backward(loss=avg_loss)
500 501
    """
    assert isinstance(loss, framework.Variable)
Y
yuyang18 已提交
502

Y
Fix bug  
yuyang18 已提交
503 504 505 506 507 508 509 510 511 512 513
    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")

Y
yuyang18 已提交
514 515 516 517
    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
Yang Yang 已提交
518 519
    if callbacks is not None:
        isinstance(callbacks, list)
Y
Yu Yang 已提交
520

F
fengjiayi 已提交
521
    program = loss.block.program
F
fengjiayi 已提交
522
    if no_grad_set is None:
523 524 525 526
        no_grad_set = set()
    no_grad_set = copy.copy(no_grad_set)
    no_grad_dict = _get_stop_gradients_(program)
    no_grad_dict[0].update(map(_append_grad_suffix_, no_grad_set))
Y
Yu Yang 已提交
527

F
update  
fengjiayi 已提交
528
    grad_info_map = dict()
F
fengjiayi 已提交
529
    root_block = program.block(0)
F
fengjiayi 已提交
530

F
fengjiayi 已提交
531 532
    fwd_op_num = root_block.desc.op_size()
    current_block_idx = program.current_block_idx
F
fengjiayi 已提交
533 534
    grad_to_var = dict()

Y
yuyang18 已提交
535 536 537 538 539 540 541 542 543 544
    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),
        })
545 546 547 548 549 550 551
    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)
    no_grad_dict[0].update(map(_append_grad_suffix_, block_no_grad_set))

    _append_backward_ops_(root_block, op_path, root_block, no_grad_dict,
Y
Yang Yang 已提交
552
                          grad_to_var, callbacks)
553 554 555 556 557 558

    # 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
559
    _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map)
F
fengjiayi 已提交
560

F
fengjiayi 已提交
561 562
    program.current_block_idx = current_block_idx
    program.sync_with_cpp()
C
chengduoZH 已提交
563 564
    # FIXME(zcd): prevent loss.grad optimized by mem_opt.
    loss.block.var(_append_grad_suffix_(loss.name)).persistable = True
F
fengjiayi 已提交
565

566 567 568
    if parameter_list is not None:
        parameters = parameter_list
    else:
F
fengjiayi 已提交
569
        params = program.global_block().all_parameters()
570
        parameters = [param.name for param in params]
571

572 573
    params_and_grads = []
    for param in parameters:
F
update  
fengjiayi 已提交
574
        if param not in grad_info_map:
F
fengjiayi 已提交
575
            continue
F
update  
fengjiayi 已提交
576
        grad_info = grad_info_map[param]
F
fengjiayi 已提交
577
        grad_block = grad_info[1]
578 579 580 581
        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
582
        param_var = program.global_block().var(param)
583 584 585 586 587
        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
588 589 590 591 592 593 594 595 596 597 598 599 600

    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
601
        attr_val = [p.name, g.name]
Y
yuyang18 已提交
602 603 604
        if g.op.has_attr(op_role_var_attr_name):
            attr_val.extend(g.op.attr(op_role_var_attr_name))
        g.op.set_attr(op_role_var_attr_name, attr_val)
Y
yuyang18 已提交
605

606
    return params_and_grads
607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750


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


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():
                if name not in input_names:
                    no_grad_set.add(name)

    return op_path


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

    Args:
        targets(Variable|list[Variable]): The target variables
        inputs(Variable|list[Variable]): The input variables
        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]): list of gradients for inputs
        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
    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)
    no_grad_dict[0].update(map(_append_grad_suffix_, no_grad_set))

    fwd_op_num = block.desc.op_size()

    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)
        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

    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)
    no_grad_dict[0].update(map(_append_grad_suffix_, block_no_grad_set))
    grad_to_var = dict()
    grad_info_map = dict()
    _append_backward_ops_(block, op_path, block, no_grad_dict, grad_to_var)

    # 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)
    prog.sync_with_cpp()

    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