partial_program.py 21.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# 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.

from __future__ import print_function
import numpy as np
17
import six
18

19
import paddle
20
from paddle.fluid import framework, backward, core, program_guard
21
from paddle.fluid.dygraph import layers
22
from paddle.fluid.dygraph.base import switch_to_static_graph
23
from paddle.fluid.dygraph.dygraph_to_static import logging_utils
24
from paddle.fluid.dygraph.dygraph_to_static.return_transformer import RETURN_NO_VALUE_MAGIC_NUM
25 26
from paddle.fluid.layers.utils import flatten
from paddle.fluid.layers.utils import pack_sequence_as
27 28
from paddle.fluid.layers.utils import _hash_with_id
from paddle.fluid.compiler import BuildStrategy
29
from paddle.fluid.contrib.mixed_precision.decorator import AutoMixedPrecisionLists
30 31
from paddle.fluid.contrib.mixed_precision.fp16_utils import rewrite_program, cast_model_to_fp16
from paddle.fluid.dygraph.amp.auto_cast import _in_amp_guard, _in_pure_fp16_guard
32
import paddle.compat as cpt
W
wanghuancoder 已提交
33
from paddle import _C_ops
34

35 36 37 38 39 40 41 42 43

class NestSequence(object):
    """
    A wrapper class that easily to flatten and restore the nest structure of
    given sequence.
    """

    def __init__(self, raw_input, need_check=False):
        self.__raw_input = raw_input
44
        self.__input_list = self.tolist()
45 46 47 48 49 50 51 52 53 54 55 56 57
        self.__var_ids = self._get_var_ids()
        self._check_non_variable(need_check)

    def tolist(self):
        """
        Flattens the nested sequences into single list.
        """
        return flatten(self.__raw_input)

    def restore(self, value_list):
        """
        Restores the nested sequence from value list.
        """
58
        assert len(self.__input_list) == len(value_list)
59 60 61 62
        return pack_sequence_as(self.__raw_input, value_list)

    def _get_var_ids(self):
        var_ids = []
63
        for idx, var in enumerate(self.__input_list):
64 65 66 67 68 69 70 71 72 73 74
            if isinstance(var, (framework.Variable, core.VarBase)):
                var_ids.append(idx)

        return var_ids

    def _check_non_variable(self, need_check):
        """
        Raises warning if output of traced function contains non-tensor type values.
        """
        if need_check:
            warning_types = set()
75
            for var in self.__input_list:
76 77 78
                if not isinstance(var, (framework.Variable, core.VarBase)):
                    warning_types.add(type(var))
            if warning_types:
79
                logging_utils.warn(
80 81 82 83 84 85 86 87 88 89
                    "Output of traced function contains non-tensor type values: {}. "
                    "Currently, We don't support to update them while training and will return "
                    "what we first saw. Please try to return them as tensor.".
                    format(list(warning_types)))

    @property
    def var_ids(self):
        return self.__var_ids

    def __getitem__(self, item):
90
        return self.__input_list[item]
91

92

93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
class LazyInitialized(object):
    """
    Descriptor to implement lazy initialization of property.
    """

    def __init__(self, function):
        self.function = function

    def __get__(self, instance, cls):
        val = self.function(instance)
        setattr(instance, self.function.__name__, val)
        return val


def _change_is_test_status(program, is_test):
    # change all `is_test` attributes
    for block in program.blocks:
        for op in block.ops:
            if op.has_attr('is_test'):
                op._set_attr('is_test', is_test)
    return program


116
class PartialProgramLayer:
117 118 119 120 121
    """
    PartialProgramLayer wraps all the ops from layers decorated by `@declarative`
    and execute them as a static subgraph.

    .. note::
122 123 124
        **1. This is a very low level API. Users should not use this API
             directly. Please use `partial_program_from(concrete_program)`
             to create it.
125 126 127 128 129 130 131 132 133 134 135 136
        **2. LoDTensorArray is not currently supported in the output.

    Args:
        main_program(Program): The main program that contains ops need to be executed.
        inputs(list[Variable]): The input list of the decorated function by `@declarative`.
        outputs(list[Variable]): The output list of the decorated function by `@declarative`.
        parameters(list[VarBase]|None): All trainable parameters included in the program. Default None.

    Returns:
        Layer: A Layer object that run all ops internally in static mode.
    """

137 138
    def __init__(self, main_program, inputs, outputs, parameters=None,
                 **kwargs):
139
        super(PartialProgramLayer, self).__init__()
140 141
        self._inputs = NestSequence(inputs)
        self._outputs = NestSequence(outputs, need_check=True)
142
        self._params = parameters if parameters is not None else []
143

144 145 146
        self._build_strategy = kwargs.get('build_strategy', BuildStrategy())
        assert isinstance(self._build_strategy, BuildStrategy)

147
        self._origin_main_program = self._verify_program(main_program)
148 149 150
        self._tmp_scope_vec = self._create_scope_vec()
        # A fake_var to handle empty input or output
        self.__fake_vars = _create_fake_var()
151
        # Set default mode to train
152
        self._double_grads = self._get_double_grads(self._origin_main_program)
153
        self.training = True
154

155 156 157 158
        custom_white_list, custom_black_list = None, None
        tracer = framework._dygraph_tracer()
        if tracer:
            custom_white_list, custom_black_list = tracer._get_amp_op_list()
159
        # For AMP training
160 161 162
        self._amp_list = AutoMixedPrecisionLists(
            custom_white_list=custom_white_list,
            custom_black_list=custom_black_list)
163

164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
    @LazyInitialized
    def _infer_program(self):
        """
        Lazy initialized property of infer_program.
        """
        return self._clone_for_test(self._origin_main_program)

    @LazyInitialized
    def _train_program(self):
        """
        Lazy initialized property of train_program.
        """
        train_program = self._append_backward_desc(self._origin_main_program)
        # Note: Only set grad type once after initializing train program. So we
        # put it here.
        self._set_grad_type(self._params, train_program)

        return train_program

183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201
    @LazyInitialized
    @switch_to_static_graph
    def _infer_amp_program(self):
        """
        Lazy initialized property of infer_amp_program.
        """
        infer_amp_program = self._origin_main_program.clone()
        with program_guard(infer_amp_program):
            rewrite_program(infer_amp_program, self._amp_list)

        return infer_amp_program

    @LazyInitialized
    def _train_amp_program(self):
        """
        Lazy initialized property of train_amp_program.
        """
        return self._append_backward_desc(self._infer_amp_program)

202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
    @LazyInitialized
    @switch_to_static_graph
    def _infer_pure_fp16_program(self):
        """
        Lazy initialized property of _infer_pure_fp16_program.
        """
        infer_pure_fp16_program = self._origin_main_program.clone()
        with program_guard(infer_pure_fp16_program):
            cast_model_to_fp16(
                infer_pure_fp16_program, self._amp_list, use_fp16_guard=False)

        return infer_pure_fp16_program

    @LazyInitialized
    def _train_pure_fp16_program(self):
        """
        Lazy initialized property of _train_pure_fp16_program.
        """
        return self._append_backward_desc(self._infer_pure_fp16_program)

222 223 224 225 226 227
    @LazyInitialized
    def _infer_program_id(self):
        return _hash_with_id(self._infer_program, self)

    @LazyInitialized
    def _train_program_id(self):
228 229 230 231 232
        program_id = _hash_with_id(self._train_program, self)
        core._set_cached_executor_build_strategy(program_id,
                                                 self._build_strategy)

        return program_id
233

234 235 236 237 238 239 240 241
    @LazyInitialized
    def _train_amp_program_id(self):
        program_id = _hash_with_id(self._train_amp_program, self)
        core._set_cached_executor_build_strategy(program_id,
                                                 self._build_strategy)

        return program_id

242 243 244 245 246 247 248 249
    @LazyInitialized
    def _train_pure_fp16_program_id(self):
        program_id = _hash_with_id(self._train_pure_fp16_program, self)
        core._set_cached_executor_build_strategy(program_id,
                                                 self._build_strategy)

        return program_id

250 251 252 253 254 255 256 257 258 259 260 261
    def _verify_program(self, main_program):
        """
        Verify that the program parameter is initialized, prune some unused params,
        and remove redundant op callstack.
        """
        # 1. Check all params from main program can be found in self._params
        self._check_params_all_inited(main_program)
        # 2. Prune the parameters not used anywhere in the program.
        self._prune_unused_params(main_program)

        return main_program

262
    @switch_to_static_graph
263
    def _append_backward_desc(self, main_program):
264 265
        # make sure all status of is_test are False in train mode.
        program = _change_is_test_status(main_program.clone(), is_test=False)
266
        targets = []
267
        for out in self._outputs.tolist():
268 269 270 271 272 273 274 275
            if isinstance(out, framework.Variable):
                targets.append(program.global_block().var(out.name))

        if targets and self._params:
            backward.gradients(targets=targets, inputs=[])

        return program

276 277 278 279 280 281 282 283 284 285
    def _prune_unused_params(self, program):
        """
        Prune the parameters not used anywhere in the program.
        The `@declarative` may only decorated a sub function which
        contains some unused parameters created in `__init__`.
        So prune these parameters to avoid unnecessary operations in
        `run_program_op`.
        """
        required_params = []
        for param in self._params:
286
            found_param = False
287
            for block in program.blocks:
288 289 290 291 292 293
                for op in block.ops:
                    if param.name in op.input_arg_names or param.name in op.output_arg_names:
                        required_params.append(param)
                        found_param = True
                        break
                if found_param:
294 295 296 297
                    break

        self._params = required_params

298 299 300 301 302 303 304 305 306 307 308
    def _get_double_grads(self, program):
        double_grads = []
        for block in program.blocks:
            for name in block.vars:
                if "@GRAD" in name:
                    var_desc = block.vars[name].desc
                    var_base = core.VarBase(var_desc.dtype(),
                                            var_desc.shape(),
                                            var_desc.name(),
                                            var_desc.type(), False)
                    double_grads.append(var_base)
309
        return self._valid_vars(double_grads)
310

311
    def _get_end_op_index(self):
312 313 314 315 316 317
        if _in_amp_guard():
            infer_program = self._infer_amp_program
        elif _in_pure_fp16_guard():
            infer_program = self._infer_pure_fp16_program
        else:
            infer_program = self._infer_program
318 319
        return infer_program.desc.block(0).op_size()

320 321
    def __call__(self, inputs):
        in_vars, out_vars = self._prepare(inputs)
322 323

        attrs = ('global_block', self.program.desc.block(0), 'start_op_index',
324 325
                 0, 'end_op_index', self._get_end_op_index(), 'is_test',
                 not self.training, 'program_id', self.program_id)
326 327 328

        self._cast_fp16_if_pure_fp16(in_vars)

W
wanghuancoder 已提交
329
        _C_ops.run_program(
330 331 332 333
            self._valid_vars(in_vars),
            self._valid_vars(self._params),
            self._valid_vars(out_vars), self._tmp_scope_vec, self._double_grads,
            *attrs)
334
        self.drop_scope_if_no_grad()
335 336
        restored_nest_out = self._restore_out(out_vars)
        return self._remove_no_value(restored_nest_out)
337

338 339 340 341 342 343 344 345 346 347
    def _cast_fp16_if_pure_fp16(self, in_vars):
        if _in_pure_fp16_guard():
            for i, var in enumerate(in_vars):
                name = var.name
                if (self.program.global_block().has_var(name) and
                        self.program.global_block().var(name).dtype ==
                        paddle.float16):
                    in_vars[i] = var.astype('float16')
                    in_vars[i].name = name

348 349 350 351 352
    def drop_scope_if_no_grad(self):
        tracer = framework._dygraph_tracer()
        if self.training and not tracer._has_grad:
            self._tmp_scope_vec.value().get_scope().drop_kids()

353 354
    @property
    def program(self):
355
        if self.training:
356 357 358 359 360 361
            if _in_amp_guard():
                return self._train_amp_program
            elif _in_pure_fp16_guard():
                return self._train_pure_fp16_program
            else:
                return self._train_program
362 363
        else:
            return self._infer_program
364

365 366
    @property
    def program_id(self):
367
        if self.training:
368 369 370 371 372 373
            if _in_amp_guard():
                return self._train_amp_program_id
            elif _in_pure_fp16_guard():
                return self._train_pure_fp16_program_id
            else:
                return self._train_program_id
374 375
        else:
            return self._infer_program_id
376

377 378 379 380 381
    def _prepare(self, inputs):
        """
        Prepare inputs, outputs, attrs.
        """
        assert isinstance(inputs, (tuple, list))
382 383
        # Flatten inputs with nested structure into single list.
        flatten_inputs = flatten(inputs)
384 385
        # Convert variable into VarBase and feed in training data.
        input_vars = []
386
        expected_place = framework._current_expected_place()
387
        for i, value in enumerate(flatten_inputs):
388 389 390
            if isinstance(value, np.ndarray):
                var = core.VarBase(
                    value=value,
391
                    name=self._inputs[i].desc.name(),
392
                    persistable=False,
393
                    place=expected_place,
394 395
                    zero_copy=True)
            elif isinstance(value, core.VarBase):
396 397 398 399 400 401 402
                # NOTE(Aurelius84): If var is on CPUPlace, it will be transformed multi times
                # into CUDAPlace when it's as input of multi Ops. so we move it in advance
                # to avoid this problem.
                if value.stop_gradient and not value.place._equals(
                        expected_place):
                    var = value._copy_to(expected_place, False)
                    var.stop_gradient = True
403 404
                else:
                    var = value
405
                var.name = self._inputs[i].desc.name()
406 407 408
            else:
                continue
            input_vars.append(var)
409

410 411
        def create_out(var_id):
            var = self._outputs[var_id]
412
            assert isinstance(var, framework.Variable)
413 414 415 416
            var_desc = var.desc
            var_base = core.VarBase(var_desc.dtype(),
                                    var_desc.shape(),
                                    var_desc.name(), var_desc.type(), False)
417 418 419 420 421 422
            return var_base

        # Create VarBase to receive output data.
        out_vars = list(map(create_out, self._outputs.var_ids))

        return input_vars, out_vars
423

424
    def _create_scope_vec(self):
425 426 427 428 429
        # Hold forward variables
        tmp_scope_vec = core.VarBase(core.VarDesc.VarType.FP32, [],
                                     "program_out_scope",
                                     core.VarDesc.VarType.STEP_SCOPES, True)

430 431 432
        inner_scope = core.Scope()
        tmp_scope_vec.value().set_scope(inner_scope)
        return tmp_scope_vec
433

434 435 436 437 438 439 440 441 442
    def _restore_out(self, out_vars):
        """
        Restores same nested outputs by only replacing the Variable with VarBase.
        """

        flatten_outputs = self._outputs.tolist()
        for i, idx in enumerate(self._outputs.var_ids):
            flatten_outputs[idx] = out_vars[i]
        outs = self._outputs.restore(flatten_outputs)
443
        if outs is not None and len(outs) == 1:
444 445 446 447
            outs = outs[0]

        return outs

448 449 450 451
    @switch_to_static_graph
    def _clone_for_test(self, main_program):
        return main_program.clone(for_test=True)

452
    def _is_no_value(self, var):
453 454 455
        if isinstance(var, core.VarBase) and var.shape == [1]:
            # NOTE: .numpy() will insert MemcpySync operation, it hits performance.
            if var.numpy()[0] == RETURN_NO_VALUE_MAGIC_NUM:
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
                return True
        return False

    def _remove_no_value(self, out_vars):
        """
        Removes invalid value for various-length return statement
        """
        if isinstance(out_vars, core.VarBase):
            if self._is_no_value(out_vars):
                return None
            return out_vars
        elif isinstance(out_vars, (tuple, list)):
            if isinstance(out_vars, tuple):
                res = tuple(
                    var for var in out_vars if not self._is_no_value(var))
            else:
                # isinstance(out_vars, list)
                res = [var for var in out_vars if not self._is_no_value(var)]

            has_removed = (len(out_vars) > len(res))
            # len(out_vars) > len(res) means we have removed var. This is
            # preventing out_vars is empty or just one element at the beginning
            if len(res) == 0 and has_removed:
                return None
            elif len(res) == 1 and has_removed:
                return res[0]
            return res

        return out_vars

486
    def _set_grad_type(self, params, train_program):
487 488 489 490 491 492 493 494
        # NOTE: if user set sparse gradient mode, the param's gradient
        # will be SelectedRows, not LoDTensor. But tracer will just
        # set param grad VarBase by forward VarBase(LoDTensor)
        # If we don't change grad_var type here, RunProgramOp need
        # transform SelectedRows to LoDTensor forcibly, it may not
        # be user wanted result.
        for param in params:
            grad_name = param.name + core.grad_var_suffix()
495
            grad_var = train_program.desc.block(0).find_var(
496 497 498 499 500 501
                cpt.to_bytes(grad_name))
            # NOTE: cannot find var desc maybe no problem, such as in batch_norm
            if grad_var is None:
                continue
            param._set_grad_type(grad_var.type())

502 503 504 505 506 507 508 509 510 511 512 513 514
    def _remove_op_call_stack(self, main_program):
        """
        Remove op's python call stack with redundant low-level error messages related to
        transforamtions to avoid confusing users.
        """
        assert isinstance(main_program, framework.Program)
        for block in main_program.blocks:
            for op in block.ops:
                if op.has_attr("op_callstack"):
                    op._remove_attr("op_callstack")

        return main_program

515 516 517 518 519 520 521 522 523 524 525 526
    def _check_params_all_inited(self, main_program):
        """
        Check all params from main program are already initialized, see details as follows:
            1. all parameters in self._params should be type `framework.ParamBase` which are created in dygraph.
            2. all parameters from transformed program can be found in self._params.
               Because they share same data with ParamBase of original dygraph.
        """
        if not isinstance(self._params, (list, tuple)):
            raise TypeError(
                "Type of self._params in PartialProgramLayer should be list or tuple, but received %s."
                % type(self._params))

527 528 529 530
        param_and_buffer_names_set = set()
        for i, var in enumerate(self._params):
            # self._params constains parameters and buffers with persistable=True.
            if not isinstance(var, core.VarBase):
531
                raise TypeError(
532 533 534
                    'Type of self._params[{}] in PartialProgramLayer should be Parameter or Variable, but received {}.'.
                    format(i, type(var)))
            param_and_buffer_names_set.add(var.name)
535 536

        for block in main_program.blocks:
537
            for name, var in six.iteritems(block.vars):
538
                if isinstance(var, framework.Parameter):
539
                    if name not in param_and_buffer_names_set:
540
                        raise ValueError(
541 542 543 544 545 546
                            "\n\tWe don't support to define layer with parameters in the function decorated by `@to_static`."
                            "\n\tBut we found parameter(%s) was created in the decorated function."
                            "\n"
                            "\n\tRevise suggestion: "
                            "\n\t\t1. Please ensure all your sublayers are inheritted from nn.Layer."
                            "\n\t\t2. Please use nn.ParameterList and nn.LayerList as container instead of using a native Python container such as List"
547 548
                            % name)

549 550 551 552 553 554 555 556
    def _valid_vars(self, vars):
        """
        Note: run_program_op.InferShape requires `X`/'Out' not be null.
        But it's common in dy2static, fake varBase is created to handle the
        problem.
        """
        return vars if vars else self.__fake_vars

557

558
def _create_fake_var():
559
    """
560
    Create a fake_var (force on CPU) to handle empty input or output
561 562
    """
    return [
563 564
        core.VarBase(core.VarDesc.VarType.FP32, [], "Fake_var",
                     core.VarDesc.VarType.RAW, False)
565 566 567 568 569 570 571 572
    ]


def partial_program_from(concrete_program):
    inputs = concrete_program.inputs
    if inputs and isinstance(inputs[0], layers.Layer):
        inputs = inputs[1:]

573 574 575
    return PartialProgramLayer(
        concrete_program.main_program, inputs, concrete_program.outputs,
        concrete_program.parameters, **concrete_program.kwargs)