backward.py 20.2 KB
Newer Older
D
dzhwinter 已提交
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
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.v2.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
19

20 21 22 23
__all__ = [
    'append_backward',
    'calc_gradient',
]
24 25


26 27
def _rename_arg_(op_descs, old_name, new_name, begin_idx=None, end_idx=None):
    """
28
    Traverse all ops in op_descs[begin_idx : end_idx],
29 30
    if any op has inputs/outputs named "old_name", rename it as 'new_name'
    """
F
update  
fengjiayi 已提交
31 32 33
    if begin_idx is None:
        begin_idx = 0
    if end_idx is None:
34
        end_idx = len(op_descs)
F
update  
fengjiayi 已提交
35
    for i in range(begin_idx, end_idx):
36
        op_desc = op_descs[i]
F
fengjiayi 已提交
37 38 39 40
        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
41 42


F
fengjiayi 已提交
43
def _create_op_desc_(op_type, inputs, outputs, attrs):
44 45 46
    """
    Create a C++ OpDesc object with specified inputs, outputs and attributes.
    """
F
fengjiayi 已提交
47 48 49 50 51 52 53 54 55 56 57 58 59 60
    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)
    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


61 62 63 64 65 66
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
67 68 69 70 71 72 73
    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:
        grad_var.set_dtype(core.DataType.FP32)


F
fengjiayi 已提交
74
def _all_in_set_(cands, s):
75 76 77
    """
    Test if all elements of 'cands' are in set 's'
    """
78 79
    if len(cands) == 0:
        return False
80 81 82 83 84 85
    for c in cands:
        if not c in s:
            return False
    return True


86 87 88 89 90 91 92 93 94 95 96 97
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


98
def _strip_grad_suffix_(name):
99 100 101 102 103
    """
    Strip the grad suffix from the given varibale name
    e.g. x@GRAD ==> x
         y@GRAD@RENAME@1 ==> y
    """
F
fengjiayi 已提交
104 105
    pos = name.find(core.grad_var_suffix())
    return name[:pos] if pos != -1 else name
106 107 108


def _append_grad_suffix_(name):
109 110 111 112
    """
    Append grad suffix to the given variable name
    e.g. x ==> x@GRAD
    """
113 114 115
    return name + core.grad_var_suffix()


F
fengjiayi 已提交
116
def _addup_repetitive_outputs_(op_descs):
117 118 119 120 121
    """
    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
122 123
    pending_sum_ops = []
    var_rename_count = collections.defaultdict(int)
F
fengjiayi 已提交
124 125
    renamed_vars = collections.defaultdict(list)
    for idx, op_desc in enumerate(op_descs):
F
update  
fengjiayi 已提交
126
        for var_name in op_desc.input_arg_names():
F
fengjiayi 已提交
127 128 129 130 131
            if len(renamed_vars[var_name]) > 1:
                pending_sum_ops.append(
                    (_create_op_desc_("sum", {"X": renamed_vars[var_name]},
                                      {"Out": [var_name]}, {}), idx))
                renamed_vars[var_name] = [var_name]
F
update  
fengjiayi 已提交
132
        for var_name in op_desc.output_arg_names():
F
fengjiayi 已提交
133 134 135
            if var_name == core.empty_var_name(
            ) or var_name in op_desc.input_arg_names():
                # empty variable or inplace op
F
fengjiayi 已提交
136
                continue
F
fengjiayi 已提交
137
            if len(renamed_vars[var_name]) == 0:
F
update  
fengjiayi 已提交
138
                # it's the first time we get the variable
F
fengjiayi 已提交
139
                renamed_vars[var_name] = [var_name]
F
update  
fengjiayi 已提交
140
            else:
F
fengjiayi 已提交
141
                if len(renamed_vars[var_name]) == 1:
F
update  
fengjiayi 已提交
142 143
                    new_name = var_name + "@RENAME@" + \
                        str(var_rename_count[var_name])
F
fengjiayi 已提交
144
                    var_rename_count[var_name] += 1
F
update  
fengjiayi 已提交
145
                    # rename original var_name
F
fengjiayi 已提交
146 147
                    renamed_vars[var_name][0] = new_name
                    _rename_arg_(op_descs, var_name, new_name, 0, idx)
F
fengjiayi 已提交
148
                    _rename_arg_(pending_sum_ops, var_name, new_name)
F
update  
fengjiayi 已提交
149 150 151

                new_name = var_name + "@RENAME@" + \
                    str(var_rename_count[var_name])
F
fengjiayi 已提交
152
                var_rename_count[var_name] += 1
F
update  
fengjiayi 已提交
153
                op_desc.rename_output(var_name, new_name)
F
fengjiayi 已提交
154 155
                renamed_vars[var_name].append(new_name)
    for var_name, inputs in renamed_vars.iteritems():
F
update  
fengjiayi 已提交
156
        if len(inputs) > 1:
F
fengjiayi 已提交
157
            pending_sum_ops.append((_create_op_desc_(
F
fengjiayi 已提交
158
                "sum", {"X": inputs}, {"Out": [var_name]}, {}), len(op_descs)))
159
    # sum_op descs are sorted according to their insert position
F
update  
fengjiayi 已提交
160
    for p in reversed(pending_sum_ops):
F
fengjiayi 已提交
161 162 163 164 165 166
        op_descs.insert(p[1], p[0])

    return op_descs


def _remove_no_grad_branch_(op_descs, no_grad_set):
167 168 169 170
    """
    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
171
        2. all grad inputs of the grad op are in 'no_grad_set'
172
    """
173 174

    def _op_can_be_removed_(op_desc, no_grad_set):
F
fengjiayi 已提交
175 176
        out_arg_names = op_desc.output_arg_names()
        if len(out_arg_names) == 0 or _all_in_set_(out_arg_names, no_grad_set):
177 178 179 180
            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
181
            no_grad_set.union(out_arg_names)
182 183 184
            return True
        return False

F
fengjiayi 已提交
185 186
    # Remove ops whose outputs are all in no_grad_dict
    op_descs = filter(
187
        lambda op_desc: not _op_can_be_removed_(op_desc, no_grad_set), op_descs)
188 189
    # Insert fill_zeros_like_op
    to_insert = []
F
fengjiayi 已提交
190
    for idx, op_desc in enumerate(op_descs):
191
        for arg in op_desc.input_arg_names():
F
fengjiayi 已提交
192 193 194
            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)]
195
                }, {"Out": [arg]}, {}), idx))
F
fengjiayi 已提交
196 197 198 199 200 201

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

    return op_descs


202 203
def _append_backward_ops_(block,
                          ops,
F
fengjiayi 已提交
204 205 206 207
                          target_block,
                          no_grad_dict,
                          grad_to_var,
                          callback=None):
208 209 210 211 212
    """
    Create all grad ops, and insert them into given block

    Args:
        block(Block): the block where forward ops are
213
        ops(Op): the forward operators whose backward ops need to be added
214
        target_block(Block): the block which is going to hold new generated grad ops
215
        no_grad_dict(dict):
216 217 218 219 220
            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
221
        callback(callable object): a callable object used to decorate new generated grad ops
222
    """
F
fengjiayi 已提交
223 224
    if callback is None:

F
fix bug  
fengjiayi 已提交
225
        def empty_callback(block, context):
F
fengjiayi 已提交
226 227 228 229
            pass

        callback = empty_callback
    elif not hasattr(callback, '__call__'):
F
fengjiayi 已提交
230
        raise ValueError("'callback' must be a callable object.")
F
fengjiayi 已提交
231

F
fengjiayi 已提交
232
    # grad_op_descs holds created grad_op, and will be appended to target_block
F
fengjiayi 已提交
233 234
    grad_op_descs = []
    program = block.program
235
    for op in reversed(ops):
F
fengjiayi 已提交
236 237 238 239 240
        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"))
            grad_sub_block = program.create_block(parent_idx=sub_block.idx)
241 242
            _append_backward_ops_(sub_block, sub_block.ops, grad_sub_block,
                                  no_grad_dict, grad_to_var)
F
fengjiayi 已提交
243 244
            grad_sub_block_list.append(grad_sub_block.desc)

F
fengjiayi 已提交
245
        # Getting op's corresponding grad_op
F
fengjiayi 已提交
246 247
        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
248

F
fengjiayi 已提交
249 250 251 252 253 254 255
        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])
256

F
fengjiayi 已提交
257
    # append op_desc in grad_op_descs to target_block
F
update  
fengjiayi 已提交
258
    for op_desc in grad_op_descs:
F
fengjiayi 已提交
259 260
        new_op_desc = target_block.desc.append_op()
        new_op_desc.copy_from(op_desc)
F
fengjiayi 已提交
261
        callback(block=target_block, context=grad_to_var)
F
update  
fengjiayi 已提交
262

263 264

def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
265 266 267 268 269 270 271 272 273 274 275 276
    """
    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
277
            val(tuple): a tuple of (str, Block), str is the corresponding grad name, Block is the block containing grad variable
278
    """
279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301
    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)
        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
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
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")):
                new_name = "%s_%s" % (name, core.unique_integer(name))
                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


F
fengjiayi 已提交
338
def append_backward(loss, parameter_list=None, no_grad_set=None, callback=None):
339
    """
F
fengjiayi 已提交
340 341 342 343
    Append backward part to main_program

    Args:
        loss(Variable): The variable generated by cost function.
344 345
        parameter_list(list[string]): Parameters that need to be updated by
            optimizer. If None, it means all parameters need to be updated.
346
        no_grad_set(set): Variables that have no gradients in Block 0.
347 348
            All variables with `step_gradient=True` from all blocks will be
            automatically added.
F
fengjiayi 已提交
349 350

    Return:
351
        (list[(Variable,Variable)]): list of (parameter, gradient) pair.
352 353
    """
    assert isinstance(loss, framework.Variable)
354

355
    program = loss.block.program
F
fengjiayi 已提交
356
    if no_grad_set is None:
357 358 359 360
        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))
361

F
update  
fengjiayi 已提交
362
    grad_info_map = dict()
363
    root_block = program.block(0)
364

365 366
    fwd_op_num = root_block.desc.op_size()
    current_block_idx = program.current_block_idx
F
fengjiayi 已提交
367 368
    grad_to_var = dict()

369 370 371 372 373 374 375 376 377 378 379 380
    op_desc = _create_op_desc_("fill_constant", {}, {
        "Out": [_append_grad_suffix_(loss.name)]
    }, {"shape": [1],
        "value": 1.0,
        "dtype": loss.dtype})
    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,
F
fengjiayi 已提交
381
                          grad_to_var, callback)
382 383 384 385 386 387

    # 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, {})

388
    _append_backward_vars_(root_block, fwd_op_num, grad_to_var, grad_info_map)
F
fengjiayi 已提交
389

390 391
    program.current_block_idx = current_block_idx
    program.sync_with_cpp()
392

393 394 395
    if parameter_list is not None:
        parameters = parameter_list
    else:
396
        params = program.global_block().all_parameters()
397
        parameters = [param.name for param in params]
398

399 400
    params_and_grads = []
    for param in parameters:
F
update  
fengjiayi 已提交
401
        if param not in grad_info_map:
402
            raise ValueError("param %s is not in map" % param)
F
update  
fengjiayi 已提交
403
        grad_info = grad_info_map[param]
F
fengjiayi 已提交
404
        grad_block = grad_info[1]
405 406 407 408
        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
409
        param_var = program.global_block().var(param)
410 411 412 413 414 415
        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))
    return params_and_grads
416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559


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