jit.py 49.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
# Copyright (c) 2019 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.

15 16
from __future__ import print_function

17 18
import os
import pickle
19
import warnings
20
import functools
21
from collections import OrderedDict
22 23

import six
24
import paddle
25
from paddle.fluid import core
26 27
from paddle.fluid.compiler import BuildStrategy, CompiledProgram, ExecutionStrategy
from paddle.fluid.data_feeder import check_type
28
from paddle.fluid.layers.utils import flatten
29
from paddle.fluid.dygraph.base import program_desc_tracing_guard, switch_to_static_graph
30
from paddle.fluid.dygraph.dygraph_to_static import logging_utils
31
from paddle.fluid.dygraph.dygraph_to_static.logging_utils import set_code_level, set_verbosity
32
from paddle.fluid.dygraph.dygraph_to_static.program_translator import ProgramTranslator, StaticFunction, unwrap_decorators
33
from paddle.fluid.dygraph.io import TranslatedLayer, INFER_MODEL_SUFFIX, INFER_PARAMS_SUFFIX, INFER_PARAMS_INFO_SUFFIX
34 35
from paddle.fluid.dygraph.layers import Layer
from paddle.fluid.executor import Executor, scope_guard
36 37 38
from paddle.fluid.framework import Block, ParamBase, Program, Variable
from paddle.fluid.framework import _current_expected_place, _dygraph_guard, _dygraph_tracer
from paddle.fluid.framework import dygraph_only, in_dygraph_mode
39
from paddle.fluid.wrapped_decorator import wrap_decorator
40

41 42
__all__ = [
    'TracedLayer', 'declarative', 'dygraph_to_static_func', 'set_code_level',
43
    'set_verbosity', 'save', 'load'
44
]
45 46 47 48 49 50 51 52 53 54 55 56


def create_program_from_desc(program_desc):
    program = Program()
    program.desc = program_desc
    program.blocks = [Block(program, 0)]
    program._sync_with_cpp()
    return program


def _extract_vars(inputs, result_list):
    if isinstance(inputs, Variable):
57
        result_list.append(inputs)
58
    elif isinstance(inputs, (list, tuple)):
59 60
        for var in inputs:
            _extract_vars(var, result_list)
61 62 63 64
    else:
        raise TypeError(
            "The type of 'each element of inputs' in fluid.dygraph.jit.TracedLayer.trace must be fluid.Variable, but received {}.".
            format(type(inputs)))
65 66 67 68 69 70 71 72


def extract_vars(inputs):
    result_list = []
    _extract_vars(inputs, result_list)
    return result_list


73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121
def _dygraph_to_static_func_(dygraph_func):
    """
    Converts imperative dygraph APIs into declarative function APIs. Decorator
    @dygraph_to_static_func only converts imperative dygraph APIs into
    declarative net-building APIs, which means it doesn't return immediate
    digital result as imperative mode. Users should handle Program and Executor
    by themselves.

    Note:
    This decorator is NOT our recommended way to transform imperative function
    to declarative function. We will remove this decorator after we finalize
    cleaning up code.

    Args:
        dygraph_func (callable): callable imperative function.

    Returns:
        Callable: converting imperative dygraph APIs into declarative
        net-building APIs.

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          import numpy as np
          from paddle.fluid.dygraph.jit import dygraph_to_static_func

          @dygraph_to_static_func
          def func(x):
              if fluid.layers.mean(x) < 0:
                  x_v = x - 1
              else:
                  x_v = x + 1

               return x_v

          x = fluid.layers.fill_constant(shape=[3, 3], value=0, dtype='float64')

          x_v = func(x)
          exe = fluid.Executor(fluid.CPUPlace())
          out = exe.run(fetch_list=[x_v])
          print(out[0])
          # [[1. 1. 1.]
          #  [1. 1. 1.]
          #  [1. 1. 1.]]

    """

    # TODO: remove this decorator after we finalize training API
122 123
    def __impl__(*args, **kwargs):
        program_translator = ProgramTranslator()
124
        if in_dygraph_mode() or not program_translator.enable_to_static:
125
            logging_utils.warn(
126
                "The decorator 'dygraph_to_static_func' doesn't work in "
127
                "dygraph mode or set ProgramTranslator.enable to False. "
128 129 130 131
                "We will just return dygraph output.")
            return dygraph_func(*args, **kwargs)
        static_func = program_translator.get_func(dygraph_func)
        return static_func(*args, **kwargs)
132 133 134 135

    return __impl__


136
dygraph_to_static_func = wrap_decorator(_dygraph_to_static_func_)
137

138

139 140 141 142 143 144
def copy_decorator_attrs(original_func, decorated_obj):
    """
    Copies some necessary attributes from original function into decorated function.

    Args:
        original_func(callable): the original decorated function.
145
        decorated_obj(StaticFunction): the target decorated StaticFunction object.
146 147 148 149 150 151 152 153 154 155 156 157 158 159
    """
    decorator_name = "declarative"

    decorated_obj.__name__ = original_func.__name__
    decorated_obj._decorator_name = decorator_name
    decorated_obj.__wrapped__ = original_func
    decorated_obj.__doc__ = original_func.__doc__
    if hasattr(original_func, "__module__"):
        decorated_obj.__module__ = original_func.__module__

    return decorated_obj


def declarative(function=None, input_spec=None):
160 161 162
    """
    Converts imperative dygraph APIs into declarative function APIs. Decorator
    @declarative handles the Program and Executor of static mode and returns
163 164 165 166
    the result as dygraph Tensor(s). Users could use the returned dygraph
    Tensor(s) to do imperative training, inference, or other operations. If the
    decorated function calls other imperative function, the called one will be
    converted into declarative function as well.
167

168
    Args:
169 170 171
        function (callable): callable imperative function.
        input_spec(list[InputSpec]): list of InputSpec to specific the shape/dtype/name
            information of each input Tensor.
172

173
    Returns:
174
        Tensor(s): containing the numerical result.
175

176 177
    Examples:
        .. code-block:: python
178

179 180 181 182 183 184 185 186 187 188 189 190 191 192
            import paddle
            from paddle.jit import to_static

            @to_static
            def func(x):
                if paddle.mean(x) < 0:
                    x_v = x - 1
                else:
                    x_v = x + 1
                return x_v

            x = paddle.ones([1, 2], dtype='float32')
            x_v = func(x)
            print(x_v) # [[2. 2.]]
193

194
    """
195

196 197
    def decorated(python_func):
        """
198
        Decorates a python function into a StaticFunction object.
199 200 201
        """
        # Step 1. unwrap the function if it is already decorated.
        _, python_func = unwrap_decorators(python_func)
202

203 204 205
        # Step 2. copy some attributes from original python function.
        static_layer = copy_decorator_attrs(
            original_func=python_func,
206
            decorated_obj=StaticFunction(
207 208 209
                function=python_func, input_spec=input_spec))

        return static_layer
210

211 212
    # for usage: `declarative(foo, ...)`
    if function is not None:
213
        if isinstance(function, Layer):
214
            if isinstance(function.forward, StaticFunction):
215
                class_name = function.__class__.__name__
216
                logging_utils.warn(
217 218 219 220 221 222
                    "`{}.forward` has already been decorated somewhere. It will be redecorated to replace previous one.".
                    format(class_name))
            function.forward = decorated(function.forward)
            return function
        else:
            return decorated(function)
223

224 225
    # for usage: `@declarative`
    return decorated
226 227


228
class _SaveLoadConfig(object):
229 230 231 232 233
    def __init__(self):
        self._output_spec = None
        self._model_filename = None
        self._params_filename = None
        self._separate_params = False
234 235
        # used for `paddle.load`
        self._keep_name_table = False
236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253

        # NOTE: Users rarely use following configs, so these configs are not open to users,
        # reducing user learning costs, but we retain the configuration capabilities

        # If True, programs are modified to only support direct inference deployment. 
        # Otherwise,more information will be stored for flexible optimization and re-training. 
        # Currently, only True is supported
        self._export_for_deployment = True

        # If True, It will save inference program only, and do not save params of Program
        self._program_only = False

    @property
    def output_spec(self):
        return self._output_spec

    @output_spec.setter
    def output_spec(self, spec):
254 255
        if spec is None:
            return
256 257
        if not isinstance(spec, list):
            raise TypeError(
258
                "The config `output_spec` should be 'list', but received input type is %s."
259 260 261 262
                % type(input))
            for var in spec:
                if not isinstance(var, core.VarBase):
                    raise TypeError(
263
                        "The element in config `output_spec` list should be 'Variable', but received element's type is %s."
264 265 266 267 268 269 270 271 272
                        % type(var))
        self._output_spec = spec

    @property
    def model_filename(self):
        return self._model_filename

    @model_filename.setter
    def model_filename(self, filename):
273 274
        if filename is None:
            return
275 276
        if not isinstance(filename, six.string_types):
            raise TypeError(
277
                "The config `model_filename` should be str, but received input's type is %s."
278 279
                % type(filename))
        if len(filename) == 0:
280
            raise ValueError("The config `model_filename` is empty string.")
281 282 283 284 285 286 287 288
        self._model_filename = filename

    @property
    def params_filename(self):
        return self._params_filename

    @params_filename.setter
    def params_filename(self, filename):
289 290
        if filename is None:
            return
291 292
        if not isinstance(filename, six.string_types):
            raise TypeError(
293
                "The config `params_filename` should be str, but received input's type is %s."
294 295
                % type(filename))
        if len(filename) == 0:
296
            raise ValueError("The config `params_filename` is empty string.")
297 298
        self._params_filename = filename

299 300 301 302 303 304
    @property
    def keep_name_table(self):
        return self._keep_name_table

    @keep_name_table.setter
    def keep_name_table(self, value):
305 306
        if value is None:
            return
307 308
        if not isinstance(value, bool):
            raise TypeError(
309
                "The config `keep_name_table` should be bool value, but received input's type is %s."
310 311 312
                % type(value))
        self._keep_name_table = value

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
def _parse_save_configs(configs):
    supported_configs = ['output_spec']

    # input check
    for key in configs:
        if key not in supported_configs:
            raise ValueError(
                "The additional config (%s) of `paddle.jit.save` is not supported."
                % (key))

    # construct inner config
    inner_config = _SaveLoadConfig()
    inner_config.output_spec = configs.get('output_spec', None)

    return inner_config


def _parse_load_config(configs):
    supported_configs = ['model_filename', 'params_filename']

    # input check
    for key in configs:
        if key not in supported_configs:
            raise ValueError(
                "The additional config (%s) of `paddle.jit.load` is not supported."
                % (key))

    # construct inner config
    inner_config = _SaveLoadConfig()
    inner_config.model_filename = configs.get('model_filename', None)
    inner_config.params_filename = configs.get('params_filename', None)

    return inner_config


349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396
def _get_input_var_names(inputs, input_spec):
    name_none_error = "The %s's name is None. " \
        "When using jit.save, please set InputSepc's name in " \
        "to_static(input_spec=[]) and jit.save(input_spec=[]) " \
        "and make sure they are consistent."
    name_no_exists_error = "The tensor `%s` does not exists. " \
        "Please make sure the name of InputSpec or example Tensor " \
        "in input_spec is the same as the name of InputSpec in " \
        "`to_static` decorated on the Layer.forward method."
    result_list = []
    input_var_names = [var.name for var in inputs if isinstance(var, Variable)]
    if input_spec is None:
        # no prune
        result_list = input_var_names
    elif input_spec is not None and len(input_spec) == len(input_var_names):
        # no prune
        result_list = input_var_names
        # if input spec name not in input_var_names, only raise warning 
        for spec in input_spec:
            if spec.name is None:
                warnings.warn(name_none_error % spec)
            elif spec.name not in input_var_names:
                warnings.warn(name_no_exists_error % spec.name)
            else:
                # do nothing
                pass
    else:
        # prune
        for spec in input_spec:
            if spec.name is None:
                # name is None, the input_spec only can be InputSpec
                raise ValueError(name_none_error % spec)
            elif spec.name not in input_var_names:
                # the input_spec can be `InputSpec` or `VarBase`
                raise ValueError(name_no_exists_error % spec.name)
            else:
                result_list.append(spec.name)

    return result_list


def _get_output_vars(outputs, output_spec):
    name_no_exists_error = "The tensor `%s` does not exists. " \
        "Please make sure the name of example Tensor " \
        "in configs.output_spec is the output tensor of " \
        "Layer.forward method."
    result_list = []
    output_vars_dict = OrderedDict()
397
    for var in flatten(outputs):
398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415
        if isinstance(var, Variable):
            output_vars_dict[var.name] = var
    if output_spec is None:
        result_list = output_vars_dict.values()
    elif output_spec is not None and len(output_spec) == len(output_vars_dict):
        result_list = output_vars_dict.values()
        for var in output_spec:
            if var.name not in output_vars_dict:
                warnings.warn(name_no_exists_error % var.name)
    else:
        for var in output_spec:
            if var.name not in output_vars_dict:
                raise ValueError(name_no_exists_error % var.name)
            else:
                result_list.append(output_vars_dict[var.name])
    return result_list


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
# NOTE(chenweihang): [ Handling of use cases of API paddle.jit.load ]
# `paddle.jit.load` may be used to load saved results of:
# 1. Expected cases:
#   - paddle.jit.save
#   - paddle.static.save_inference_model
#   - paddle.fluid.io.save_inference_model
# 2. Error cases:
#   - paddle.save: no .pdmodel for prefix
#   - paddle.static.save: no .pdiparams but .pdparams exists
#   - paddle.fluid.io.save_params/save_persistables: no __model__
# TODO(chenweihang): polish error message in above error cases
def _build_load_path_and_config(path, config):
    # NOTE(chenweihang): If both [prefix save format] and [directory save format] exist,
    # raise error, avoid confusing behavior
    prefix_format_path = path + INFER_MODEL_SUFFIX
    prefix_format_exist = os.path.exists(prefix_format_path)
    directory_format_exist = os.path.isdir(path)
    if prefix_format_exist and directory_format_exist:
        raise ValueError(
            "The %s.pdmodel and %s directory exist at the same time, "
            "don't know which one to load, please make sure that the specified target "
            "of ``path`` is unique." % (path, path))
    elif not prefix_format_exist and not directory_format_exist:
        raise ValueError("The ``path`` (%s) to load model not exists." % path)
    else:
        if prefix_format_exist:
            file_prefix = os.path.basename(path)
            model_path = os.path.dirname(path)
            if config.model_filename is not None:
                warnings.warn(
                    "When loading the result saved with the "
                    "specified file prefix, the ``model_filename`` config does "
                    "not take effect.")
            config.model_filename = file_prefix + INFER_MODEL_SUFFIX
            if config.params_filename is not None:
                warnings.warn(
                    "When loading the result saved with the "
                    "specified file prefix, the ``params_filename`` config does "
                    "not take effect.")
            config.params_filename = file_prefix + INFER_PARAMS_SUFFIX
        else:
            # Compatible with the old save_inference_model format
            model_path = path
459

460
    return model_path, config
461 462


463
@switch_to_static_graph
464
def save(layer, path, input_spec=None, **configs):
465
    """
466
    Saves input Layer as ``paddle.jit.TranslatedLayer``
467 468 469
    format model, which can be used for inference or fine-tuning after loading.

    It will save the translated program and all related persistable 
470
    variables of input Layer to given ``path`` .
471
    
472
    ``path`` is the prefix of saved objects, and the saved translated program file 
473
    suffix is ``.pdmodel`` , the saved persistable variables file suffix is ``.pdiparams`` ,
474 475
    and here also saved some additional variable description information to a file,  
    its suffix is ``.pdiparams.info``, these additional information is used in fine-tuning.
476 477

    The saved model can be loaded by follow APIs:
478 479
      - ``paddle.jit.load`` 
      - ``paddle.static.load_inference_model`` 
480 481 482
      - Other C++ inference APIs

    Args:
483
        layer (Layer): The Layer to be saved.
484
        path (str): The path prefix to save model. The format is ``dirname/file_prefix`` or ``file_prefix``.
485 486 487 488
        input_spec (list[InputSpec|Tensor], optional): Describes the input of the saved model's forward 
            method, which can be described by InputSpec or example Tensor. If None, all input variables of 
            the original Layer's forward method would be the inputs of the saved model. Default None.
        **configs (dict, optional): Other save configuration options for compatibility. We do not 
489 490 491 492
            recommend using these configurations, they may be removed in the future. If not necessary, 
            DO NOT use them. Default None.
            The following options are currently supported:
            (1) output_spec (list[Tensor]): Selects the output targets of the saved model.
493
            By default, all return variables of original Layer's forward method are kept as the 
494 495 496
            output of the saved model. If the provided ``output_spec`` list is not all output variables, 
            the saved model will be pruned according to the given ``output_spec`` list. 

497 498 499 500 501 502 503
    Returns:
        None

    Examples:
        .. code-block:: python

            import numpy as np
504 505 506
            import paddle
            import paddle.nn as nn
            import paddle.optimizer as opt
507

508 509 510
            BATCH_SIZE = 16
            BATCH_NUM = 4
            EPOCH_NUM = 4
511

512 513 514 515 516 517 518
            IMAGE_SIZE = 784
            CLASS_NUM = 10

            # define a random dataset
            class RandomDataset(paddle.io.Dataset):
                def __init__(self, num_samples):
                    self.num_samples = num_samples
519

520 521 522 523
                def __getitem__(self, idx):
                    image = np.random.random([IMAGE_SIZE]).astype('float32')
                    label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
                    return image, label
524

525 526
                def __len__(self):
                    return self.num_samples
527

528 529
            class LinearNet(nn.Layer):
                def __init__(self):
530
                    super(LinearNet, self).__init__()
531
                    self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)
532

533
                @paddle.jit.to_static
534 535 536
                def forward(self, x):
                    return self._linear(x)

537 538 539 540 541 542 543 544 545 546 547 548
            def train(layer, loader, loss_fn, opt):
                for epoch_id in range(EPOCH_NUM):
                    for batch_id, (image, label) in enumerate(loader()):
                        out = layer(image)
                        loss = loss_fn(out, label)
                        loss.backward()
                        opt.step()
                        opt.clear_grad()
                        print("Epoch {} batch {}: loss = {}".format(
                            epoch_id, batch_id, np.mean(loss.numpy())))

            # 1. train & save model.
549

550 551 552 553
            # create network
            layer = LinearNet()
            loss_fn = nn.CrossEntropyLoss()
            adam = opt.Adam(learning_rate=0.001, parameters=layer.parameters())
554

555 556 557 558 559 560 561
            # create data loader
            dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
            loader = paddle.io.DataLoader(dataset,
                batch_size=BATCH_SIZE,
                shuffle=True,
                drop_last=True,
                num_workers=2)
562

563 564
            # train
            train(layer, loader, loss_fn, adam)
565

566
            # save
567 568
            path = "example_model/linear"
            paddle.jit.save(layer, path)
569 570
    """

571
    # 1. input build & check
572
    prog_translator = ProgramTranslator()
573
    if not prog_translator.enable_to_static:
574
        raise RuntimeError(
575
            "The paddle.jit.save doesn't work when setting ProgramTranslator.enable to False."
576 577 578
        )
    if not isinstance(layer, Layer):
        raise TypeError(
579
            "The input layer of paddle.jit.save should be 'Layer', but received layer type is %s."
580 581
            % type(layer))

582 583 584 585 586 587 588 589 590 591 592
    # path check
    file_prefix = os.path.basename(path)
    if file_prefix == "":
        raise ValueError(
            "The input path MUST be format of dirname/file_prefix "
            "[dirname\\file_prefix in Windows system], but received "
            "file_prefix is empty string.")

    dirname = os.path.dirname(path)
    if dirname and not os.path.exists(dirname):
        os.makedirs(dirname)
593

594 595
    # avoid change user given input_spec
    inner_input_spec = None
596
    if input_spec is not None:
597 598 599 600 601 602 603
        for attr_func in dir(layer):
            static_func = getattr(layer, attr_func, None)
            if isinstance(static_func,
                          StaticFunction) and 'forward' != attr_func:
                raise ValueError(
                    "If there are static functions other than 'forward' that need to be saved, the input 'input_spec' should be None, but received the type of 'input_spec' is %s."
                    % type(input_spec))
604 605 606 607
        if not isinstance(input_spec, list):
            raise TypeError(
                "The input input_spec should be 'list', but received input_spec's type is %s."
                % type(input_spec))
608
        inner_input_spec = []
609
        for var in input_spec:
610 611 612 613 614 615
            if isinstance(var, paddle.static.InputSpec):
                inner_input_spec.append(var)
            elif isinstance(var, (core.VarBase, Variable)):
                inner_input_spec.append(
                    paddle.static.InputSpec.from_tensor(var))
            else:
616
                raise TypeError(
617
                    "The element in input_spec list should be 'Variable' or `paddle.static.InputSpec`, but received element's type is %s."
618 619
                    % type(var))

620 621
    # parse configs
    configs = _parse_save_configs(configs)
622 623
    scope = core.Scope()
    extra_var_info = dict()
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
    for attr_func in dir(layer):
        static_func = getattr(layer, attr_func, None)
        if isinstance(static_func, StaticFunction):
            concrete_program = static_func.concrete_program
        elif 'forward' == attr_func:
            # transform in jit.save, if input_spec is incomplete, declarative will throw error
            static_forward = declarative(
                layer.forward, input_spec=inner_input_spec)
            concrete_program = static_forward.concrete_program
            # the input_spec has been used in declarative, which is equal to 
            # @declarative with input_spec and jit.save without input_spec,
            # avoid needless warning
            inner_input_spec = None
        else:
            continue

        # 3. build input & output of save_infernece_model
        # NOTE(chenweihang): [ Get input variables name ]
        # There are two cases, whether to prune the inputs or not
        # - not prune inputs (recommend):
        #   - the len(input_spec) == len((concrete_program.inputs) - 1
        #   - here can use concrete_program.inputs directly
        # - prune inputs:
        #   - the input_spec length < len((concrete_program.inputs) - 1
        #   - the input_spec's name should be in concrete_program.inputs
        input_var_names = _get_input_var_names(concrete_program.inputs,
                                               inner_input_spec)

        # NOTE(chenweihang): [ Get output variables ]
        # the rule is like [ Get input variables name ]. For output var, 
        # we only support VarBase spec, and actually, we only need the 
        # var name of output, and we don't recommended to use output_spec
        output_vars = _get_output_vars(concrete_program.outputs,
                                       configs.output_spec)

        # NOTE(chenweihang): we maintain the mapping of variable name to
        # structured name, the buffer variable (non-persistable)
        # saved to inference program may not need by dygraph Layer, 
        # we only record the state_dict variable's structured name
        state_names_dict = dict()
        for structured_name, var in six.iteritems(layer.state_dict()):
            state_names_dict[var.name] = structured_name

        # 4. share parameters from Layer to scope & record var info        
        for param_or_buffer in concrete_program.parameters:
            # share to scope
            param_or_buffer_tensor = scope.var(param_or_buffer.name).get_tensor(
            )
            src_tensor = param_or_buffer.value().get_tensor()
            param_or_buffer_tensor._share_data_with(src_tensor)
            # record var info
            if param_or_buffer.name not in extra_var_info:
                extra_info_dict = dict()
                if param_or_buffer.name in state_names_dict:
                    extra_info_dict['structured_name'] = state_names_dict[
                        param_or_buffer.name]
                extra_info_dict['stop_gradient'] = param_or_buffer.stop_gradient
                if isinstance(param_or_buffer, ParamBase):
                    extra_info_dict['trainable'] = param_or_buffer.trainable
                extra_var_info[param_or_buffer.name] = extra_info_dict

        # 5. save inference model
        from paddle.fluid.io import save_inference_model

        # construct new save_inference_model arguments
        model_path = dirname
        # NOTE(chenweihang): because prefix contains model and params filename,
        # so we don't support set model_filename & params_filename 
        if 'forward' == attr_func:
            model_filename = file_prefix + INFER_MODEL_SUFFIX
            params_filename = file_prefix + INFER_PARAMS_SUFFIX
        else:
            model_filename = file_prefix + '.' + attr_func + INFER_MODEL_SUFFIX
            params_filename = file_prefix + '.' + attr_func + INFER_PARAMS_SUFFIX

        with scope_guard(scope):
            save_inference_model(
                dirname=model_path,
                feeded_var_names=input_var_names,
                target_vars=output_vars,
                executor=Executor(_current_expected_place()),
                main_program=concrete_program.main_program.clone(),
                model_filename=model_filename,
                params_filename=params_filename,
                export_for_deployment=configs._export_for_deployment,
                program_only=configs._program_only)

    # NOTE(chenweihang): [ Save extra variable info ]
    # save_inference_model will lose some important variable information, including:
    #   - Variable name and correspondence (when saved variables as one file)
    #   - Variable.stop_gradient information
    #   - Which persistent variable are parameter and which are not
    #   - Parameter.trainable information
    #
    # The lost information cannot be recovered when it is loaded again, 
    # so if we want to perform fine-tune after loading, we may need to 
    # configure redundant information to proceed.
    #
    # Due to compatibility issues, we cannot change the original storage structure, 
    # but we can save these information in `jit.save` without changing the original 
    # storage to improve user experience. So we save extra information into
    # file `***.pdiparams.info`
726
    with scope_guard(scope):
727
        extra_var_info_path = path + INFER_PARAMS_INFO_SUFFIX
728 729 730 731 732
        with open(extra_var_info_path, 'wb') as f:
            pickle.dump(extra_var_info, f, protocol=2)


@dygraph_only
733
def load(path, **configs):
734 735 736
    """
    :api_attr: imperative

737 738 739
    Load model saved by ``paddle.jit.save`` or ``paddle.static.save_inference_model`` or 
    paddle 1.x API ``paddle.fluid.io.save_inference_model`` as ``paddle.jit.TranslatedLayer``, 
    then performing inference or fine-tune training.
740 741

    .. note::
742
        If you load model saved by ``paddle.static.save_inference_model`` ,
743 744
        there will be the following limitations when using it in fine-tuning:
        1. Imperative mode do not support LoDTensor. All original model's feed targets or parametars that depend on LoD are temporarily unavailable.
745
        2. All saved model's feed targets need to be passed into TranslatedLayer's forward function.
746 747 748 749
        3. The variable's ``stop_gradient`` information is lost and can not be recovered.
        4. The parameter's ``trainable`` information is lost and can not be recovered.

    Args:
750 751
        path (str): The path prefix to load model. The format is ``dirname/file_prefix`` or ``file_prefix`` .
        **configs (dict, optional): Other load configuration options for compatibility. We do not 
752 753 754
            recommend using these configurations, they may be removed in the future. If not necessary, 
            DO NOT use them. Default None.
            The following options are currently supported:
755
            (1) model_filename (str): The inference model file name of the paddle 1.x 
756
            ``save_inference_model`` save format. Default file name is :code:`__model__` . 
757
            (2) params_filename (str): The persistable variables file name of the paddle 1.x 
758 759 760
            ``save_inference_model`` save format. No default file name, save variables separately 
            by default.

761 762 763 764 765

    Returns:
        TranslatedLayer: A Layer object can run saved translated model.

    Examples:
766
        1. Load model saved by ``paddle.jit.save`` then performing inference and fine-tune training.
767 768 769 770

        .. code-block:: python

            import numpy as np
771 772 773
            import paddle
            import paddle.nn as nn
            import paddle.optimizer as opt
774

775 776 777
            BATCH_SIZE = 16
            BATCH_NUM = 4
            EPOCH_NUM = 4
778

779 780
            IMAGE_SIZE = 784
            CLASS_NUM = 10
781

782 783 784 785
            # define a random dataset
            class RandomDataset(paddle.io.Dataset):
                def __init__(self, num_samples):
                    self.num_samples = num_samples
786

787 788 789 790
                def __getitem__(self, idx):
                    image = np.random.random([IMAGE_SIZE]).astype('float32')
                    label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
                    return image, label
791

792 793 794 795 796
                def __len__(self):
                    return self.num_samples

            class LinearNet(nn.Layer):
                def __init__(self):
797
                    super(LinearNet, self).__init__()
798
                    self._linear = nn.Linear(IMAGE_SIZE, CLASS_NUM)
799

800
                @paddle.jit.to_static
801 802 803
                def forward(self, x):
                    return self._linear(x)

804 805 806 807 808 809 810 811 812 813 814
            def train(layer, loader, loss_fn, opt):
                for epoch_id in range(EPOCH_NUM):
                    for batch_id, (image, label) in enumerate(loader()):
                        out = layer(image)
                        loss = loss_fn(out, label)
                        loss.backward()
                        opt.step()
                        opt.clear_grad()
                        print("Epoch {} batch {}: loss = {}".format(
                            epoch_id, batch_id, np.mean(loss.numpy())))

815
            # 1. train & save model.
816

817
            # create network
818 819 820 821
            layer = LinearNet()
            loss_fn = nn.CrossEntropyLoss()
            adam = opt.Adam(learning_rate=0.001, parameters=layer.parameters())

822
            # create data loader
823 824 825 826 827 828
            dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
            loader = paddle.io.DataLoader(dataset,
                batch_size=BATCH_SIZE,
                shuffle=True,
                drop_last=True,
                num_workers=2)
829

830 831
            # train
            train(layer, loader, loss_fn, adam)
832

833
            # save
834 835
            path = "example_model/linear"
            paddle.jit.save(layer, path)
836

837
            # 2. load model
838

839
            # load
840
            loaded_layer = paddle.jit.load(path)
841 842

            # inference
843 844 845
            loaded_layer.eval()
            x = paddle.randn([1, IMAGE_SIZE], 'float32')
            pred = loaded_layer(x)
846 847

            # fine-tune
848 849 850
            loaded_layer.train()
            adam = opt.Adam(learning_rate=0.001, parameters=loaded_layer.parameters())
            train(loaded_layer, loader, loss_fn, adam)
851 852


853
        2. Load model saved by ``paddle.fluid.io.save_inference_model`` then performing and fine-tune training.
854 855 856 857

        .. code-block:: python

            import numpy as np
858
            import paddle
859
            import paddle.static as static
860 861
            import paddle.nn as nn
            import paddle.optimizer as opt
862
            import paddle.nn.functional as F
863

864 865 866
            BATCH_SIZE = 16
            BATCH_NUM = 4
            EPOCH_NUM = 4
867

868 869 870 871 872 873 874
            IMAGE_SIZE = 784
            CLASS_NUM = 10

            # define a random dataset
            class RandomDataset(paddle.io.Dataset):
                def __init__(self, num_samples):
                    self.num_samples = num_samples
875

876 877 878 879
                def __getitem__(self, idx):
                    image = np.random.random([IMAGE_SIZE]).astype('float32')
                    label = np.random.randint(0, CLASS_NUM - 1, (1, )).astype('int64')
                    return image, label
880

881 882
                def __len__(self):
                    return self.num_samples
883

884 885
            paddle.enable_static()

886 887
            image = static.data(name='image', shape=[None, 784], dtype='float32')
            label = static.data(name='label', shape=[None, 1], dtype='int64')
888
            pred = static.nn.fc(x=image, size=10, activation='softmax')
889 890
            loss = F.cross_entropy(input=pred, label=label)
            avg_loss = paddle.mean(loss)
891

892
            optimizer = paddle.optimizer.SGD(learning_rate=0.001)
893 894
            optimizer.minimize(avg_loss)

895 896 897
            place = paddle.CPUPlace()
            exe = static.Executor(place)
            exe.run(static.default_startup_program())
898

899 900 901 902 903 904 905 906 907
            # create data loader
            dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
            loader = paddle.io.DataLoader(dataset,
                feed_list=[image, label],
                places=place,
                batch_size=BATCH_SIZE, 
                shuffle=True,
                drop_last=True,
                num_workers=2)
908 909 910 911

            # 1. train and save inference model
            for data in loader():
                exe.run(
912
                    static.default_main_program(),
913 914 915 916
                    feed=data, 
                    fetch_list=[avg_loss])

            model_path = "fc.example.model"
917
            paddle.fluid.io.save_inference_model(
918 919 920
                model_path, ["image"], [pred], exe)

            # 2. load model
921 922

            # enable dygraph mode
923 924 925 926
            paddle.disable_static(place)

            # load
            fc = paddle.jit.load(model_path)
927

928 929 930
            # inference
            fc.eval()
            x = paddle.randn([1, IMAGE_SIZE], 'float32')
931 932
            pred = fc(x)

933
            # fine-tune
934
            fc.train()
935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951
            loss_fn = nn.CrossEntropyLoss()
            adam = opt.Adam(learning_rate=0.001, parameters=fc.parameters())
            loader = paddle.io.DataLoader(dataset,
                places=place,
                batch_size=BATCH_SIZE,
                shuffle=True,
                drop_last=True,
                num_workers=2)
            for epoch_id in range(EPOCH_NUM):
                for batch_id, (image, label) in enumerate(loader()):
                    out = fc(image)
                    loss = loss_fn(out, label)
                    loss.backward()
                    adam.step()
                    adam.clear_grad()
                    print("Epoch {} batch {}: loss = {}".format(
                        epoch_id, batch_id, np.mean(loss.numpy())))
952
    """
953 954 955 956
    # 1. construct correct config
    config = _parse_load_config(configs)
    model_path, config = _build_load_path_and_config(path, config)

957
    return TranslatedLayer._construct(model_path, config)
958 959


960
@dygraph_only
Z
Zeng Jinle 已提交
961 962 963 964 965
def _trace(layer,
           inputs,
           feed_prefix='feed_',
           fetch_prefix='fetch_',
           tmp_prefix='t_'):
966
    assert isinstance(layer, Layer)
967 968 969 970 971 972 973 974 975

    if not isinstance(inputs, (list, tuple)):
        inputs = [inputs]

    tracer = _dygraph_tracer()._get_program_desc_tracer()

    var_list = extract_vars(inputs)

    with program_desc_tracing_guard(True):
976
        original_outputs = layer(*inputs)
977 978 979 980
        if not isinstance(original_outputs, (list, tuple)):
            outputs = [original_outputs]
        else:
            outputs = original_outputs
981
        out_vars = [var for var in outputs]
982

983
        program_desc, feed_names, fetch_names, parameters = tracer.create_program_desc(
Z
Zeng Jinle 已提交
984
            var_list, feed_prefix, out_vars, fetch_prefix, tmp_prefix)
985 986 987 988 989
        tracer.reset()

    with _dygraph_guard(None):
        program = create_program_from_desc(program_desc)

990
    return original_outputs, program, feed_names, fetch_names, parameters
991 992 993 994


class TracedLayer(object):
    """
995 996
    :api_attr: imperative
    
997 998 999 1000 1001
    TracedLayer is used to convert a forward dygraph model to a static
    graph model. This is mainly used to save the dygraph model for online
    inference using C++. Besides, users can also do inference in Python
    using the converted static graph model, which usually has better
    performance than the original dygraph model.
1002 1003 1004 1005

    TracedLayer would run the static graph model using :code:`Executor`
    and :code:`CompiledProgram` . The static graph model would share
    parameters with the dygraph model.
1006 1007

    All TracedLayer objects should not be created by constructor and should
1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
    be created by static method :code:`TracedLayer.trace(layer, inputs)` .

    The TracedLayer can only be used to convert the data-independent dygraph
    model into the static graph model, which means the dygraph model should
    be independent with the tensor data and shape.
    """

    def __init__(self, program, parameters, feed_names, fetch_names):
        self._program = program
        self._feed_names = feed_names
        self._fetch_names = fetch_names
1019
        self._params = parameters
1020 1021 1022 1023 1024

        self._place = _current_expected_place()

        self._scope = core.Scope()
        for p in parameters:
1025
            src_tensor = p.value().get_tensor()
1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048
            dst_tensor = self._scope.var(p.name).get_tensor()
            dst_tensor._share_data_with(src_tensor)

        self._exe = Executor(self._place)
        self._compiled_program = None
        self._build_strategy = None
        self._exec_strategy = None

    @property
    def program(self):
        return self._program

    def _switch(self, is_test=True):
        for block_id in range(self._program.num_blocks):
            block = self._program.block(block_id)
            for op in block.ops:
                if op.has_attr("is_test"):
                    op._set_attr("is_test", is_test)

    @staticmethod
    @dygraph_only
    def trace(layer, inputs):
        """
1049
        This method is the only allowed method to create TracedLayer object.
1050 1051 1052 1053
        It would call the :code:`layer(*inputs)` method to run the dygraph
        model and convert it into a static graph model.

        Args:
1054
            layer (dygraph.Layer): the layer object to be traced.
1055 1056
            inputs (list(Tensor)|tuple(Tensor)|Tensor): the input tensors of
                the layer object.
1057 1058

        Returns:
1059
            tuple: A tuple of 2 items, whose the first item is the output of
1060 1061
                :code:`layer(*inputs)` , and the second item is the created
                TracedLayer object.
1062

1063
        Examples:
1064 1065 1066
            .. code-block:: python:

                import paddle.fluid as fluid
1067
                from paddle.fluid.dygraph import Linear, to_variable, TracedLayer
1068 1069 1070
                import numpy as np

                class ExampleLayer(fluid.dygraph.Layer):
1071 1072 1073
                    def __init__(self):
                        super(ExampleLayer, self).__init__()
                        self._fc = Linear(3, 10)
1074 1075 1076 1077 1078

                    def forward(self, input):
                        return self._fc(input)

                with fluid.dygraph.guard():
1079
                    layer = ExampleLayer()
1080 1081 1082
                    in_np = np.random.random([2, 3]).astype('float32')
                    in_var = to_variable(in_np)
                    out_dygraph, static_layer = TracedLayer.trace(layer, inputs=[in_var])
1083 1084 1085 1086 1087 1088 1089 1090 1091

                    # run the static graph model using Executor inside
                    out_static_graph = static_layer([in_var])

                    print(len(out_static_graph)) # 1
                    print(out_static_graph[0].shape) # (2, 10)

                    # save the static graph model for inference
                    static_layer.save_inference_model(dirname='./saved_infer_model')
1092
        """
1093 1094 1095 1096
        assert isinstance(
            layer, Layer
        ), "The type of 'layer' in fluid.dygraph.jit.TracedLayer.trace must be fluid.dygraph.Layer, but received {}.".format(
            type(layer))
1097 1098
        outs, prog, feed, fetch, parameters = _trace(layer, inputs)
        traced = TracedLayer(prog, parameters, feed, fetch)
1099 1100 1101 1102 1103 1104 1105
        return outs, traced

    def set_strategy(self, build_strategy=None, exec_strategy=None):
        """
        Set the strategies when running static graph model.

        Args:
1106
            build_strategy (BuildStrategy, optional): build strategy of
1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117
                :code:`CompiledProgram` inside TracedLayer. Default None.
            exec_strategy (ExecutionStrategy, optional): execution strategy of
                :code:`CompiledProgram` inside TracedLayer. Default None.

        Returns:
            None

        Examples:
            .. code-block:: python:

                import paddle.fluid as fluid
1118
                from paddle.fluid.dygraph import Linear, to_variable, TracedLayer
1119 1120 1121
                import numpy as np

                class ExampleLayer(fluid.dygraph.Layer):
1122 1123 1124
                    def __init__(self):
                        super(ExampleLayer, self).__init__()
                        self._fc = Linear(3, 10)
1125 1126 1127 1128 1129

                    def forward(self, input):
                        return self._fc(input)

                with fluid.dygraph.guard():
1130
                    layer = ExampleLayer()
1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145
                    in_np = np.random.random([2, 3]).astype('float32')
                    in_var = to_variable(in_np)

                    out_dygraph, static_layer = TracedLayer.trace(layer, inputs=[in_var])

                    build_strategy = fluid.BuildStrategy()
                    build_strategy.enable_inplace = True

                    exec_strategy = fluid.ExecutionStrategy()
                    exec_strategy.num_threads = 2

                    static_layer.set_strategy(build_strategy=build_strategy, exec_strategy=exec_strategy)
                    out_static_graph = static_layer([in_var])
        """
        assert self._compiled_program is None, "Cannot set strategy after run"
1146 1147 1148 1149 1150 1151 1152 1153
        assert isinstance(
            build_strategy, (type(None), BuildStrategy)
        ), "The type of 'build_strategy' in fluid.dygraph.jit.TracedLayer.set_strategy must be fluid.BuildStrategy, but received {}.".format(
            type(build_strategy))
        assert isinstance(
            exec_strategy, (type(None), ExecutionStrategy)
        ), "The type of 'exec_strategy' in fluid.dygraph.jit.TracedLayer.set_strategy must be fluid.ExecutionStrategy, but received {}.".format(
            type(exec_strategy))
1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171
        self._build_strategy = build_strategy
        self._exec_strategy = exec_strategy

    @switch_to_static_graph
    def _compile(self):
        self._compiled_program = CompiledProgram(
            self._program).with_data_parallel(
                build_strategy=self._build_strategy,
                exec_strategy=self._exec_strategy,
                places=self._place)

    def _build_feed(self, inputs):
        assert isinstance(inputs, (list, tuple)), \
            "Inputs should be a list or tuple of variables"
        assert len(inputs) == len(self._feed_names)
        feed_dict = {}
        if in_dygraph_mode():
            for x, name in zip(inputs, self._feed_names):
1172
                feed_dict[name] = x.value().get_tensor()
1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194
        else:
            for x, name in zip(inputs, self._feed_names):
                feed_dict[name] = x

        return feed_dict

    @switch_to_static_graph
    def _run(self, feed):
        return self._exe.run(self._compiled_program,
                             feed=feed,
                             fetch_list=self._fetch_names)

    def __call__(self, inputs):
        with scope_guard(self._scope):
            if self._compiled_program is None:
                self._compile()

            return self._run(self._build_feed(inputs))

    @switch_to_static_graph
    def save_inference_model(self, dirname, feed=None, fetch=None):
        """
1195 1196
        Save the TracedLayer to a model for inference. The saved
        inference model can be loaded by C++ inference APIs.
1197 1198

        Args:
1199
            dirname (str): the directory to save the inference model.
1200
            feed (list[int], optional): the input variable indices of the saved
1201
                inference model. If None, all input variables of the
1202 1203 1204 1205 1206 1207 1208 1209
                TracedLayer object would be the inputs of the saved inference
                model. Default None.
            fetch (list[int], optional): the output variable indices of the
                saved inference model. If None, all output variables of the
                TracedLayer object would be the outputs of the saved inference
                model. Default None.

        Returns:
1210
            None
1211 1212 1213 1214 1215

        Examples:
            .. code-block:: python:

                import paddle.fluid as fluid
1216
                from paddle.fluid.dygraph import Linear, to_variable, TracedLayer
1217 1218 1219
                import numpy as np

                class ExampleLayer(fluid.dygraph.Layer):
1220 1221 1222
                    def __init__(self):
                        super(ExampleLayer, self).__init__()
                        self._fc = Linear(3, 10)
1223 1224 1225 1226

                    def forward(self, input):
                        return self._fc(input)

1227 1228 1229
                save_dirname = './saved_infer_model'
                in_np = np.random.random([2, 3]).astype('float32')

1230
                with fluid.dygraph.guard():
1231
                    layer = ExampleLayer()
1232 1233
                    in_var = to_variable(in_np)
                    out_dygraph, static_layer = TracedLayer.trace(layer, inputs=[in_var])
1234
                    static_layer.save_inference_model(save_dirname, feed=[0], fetch=[0])
1235 1236

                place = fluid.CPUPlace()
1237 1238
                exe = fluid.Executor(place)
                program, feed_vars, fetch_vars = fluid.io.load_inference_model(save_dirname,
1239
                                                    exe)
1240 1241 1242

                fetch, = exe.run(program, feed={feed_vars[0]: in_np}, fetch_list=fetch_vars)
                print(fetch.shape) # (2, 10)
1243
        """
1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258
        check_type(dirname, "dirname", str,
                   "fluid.dygraph.jit.TracedLayer.save_inference_model")
        check_type(feed, "feed", (type(None), list),
                   "fluid.dygraph.jit.TracedLayer.save_inference_model")
        if isinstance(feed, list):
            for f in feed:
                check_type(f, "each element of feed", int,
                           "fluid.dygraph.jit.TracedLayer.save_inference_model")
        check_type(fetch, "fetch", (type(None), list),
                   "fluid.dygraph.jit.TracedLayer.save_inference_model")
        if isinstance(fetch, list):
            for f in fetch:
                check_type(f, "each element of fetch", int,
                           "fluid.dygraph.jit.TracedLayer.save_inference_model")

1259
        from paddle.fluid.io import save_inference_model
1260 1261 1262 1263 1264

        def get_feed_fetch(all_vars, partial_vars):
            if partial_vars is None:
                return all_vars

1265
            return [all_vars[idx] for idx in partial_vars]
1266 1267 1268 1269 1270 1271 1272 1273 1274 1275

        with scope_guard(self._scope):
            feeded_var_names = get_feed_fetch(self._feed_names, feed)
            target_var_names = get_feed_fetch(self._fetch_names, fetch)
            target_vars = []
            for name in target_var_names:
                target_var = self._program.global_block().vars.get(name, None)
                assert target_var is not None, "{} cannot be found".format(name)
                target_vars.append(target_var)

1276
            save_inference_model(
1277 1278 1279 1280 1281
                dirname=dirname,
                feeded_var_names=feeded_var_names,
                target_vars=target_vars,
                executor=self._exe,
                main_program=self._program.clone())