jit.py 49.6 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 19
import os
import six
import pickle
20

21
import warnings
22
from paddle.fluid import core
23 24
from paddle.fluid.compiler import BuildStrategy, CompiledProgram, ExecutionStrategy
from paddle.fluid.data_feeder import check_type
25
from paddle.fluid.dygraph.base import program_desc_tracing_guard, switch_to_static_graph
26
from paddle.fluid.dygraph.dygraph_to_static.program_translator import ProgramTranslator, FunctionSpec
27 28
from paddle.fluid.dygraph.layers import Layer
from paddle.fluid.executor import Executor, scope_guard
29
from paddle.fluid.framework import Program, Block, Variable, ParamBase, _dygraph_tracer, dygraph_only, _dygraph_guard, _current_expected_place, in_dygraph_mode
30
from paddle.fluid.wrapped_decorator import wrap_decorator
31 32 33
from paddle.fluid.dygraph.io import TranslatedLayer, VARIABLE_FILENAME, EXTRA_VAR_INFO_FILENAME

__all__ = ['TracedLayer', 'declarative', 'dygraph_to_static_func']
34 35 36 37 38 39 40 41 42 43 44 45


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):
46
        result_list.append(inputs)
47
    elif isinstance(inputs, (list, tuple)):
48 49
        for var in inputs:
            _extract_vars(var, result_list)
50 51 52 53
    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)))
54 55 56 57 58 59 60 61


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


62 63 64 65 66 67 68 69 70 71 72 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
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
111 112
    def __impl__(*args, **kwargs):
        program_translator = ProgramTranslator()
113
        if in_dygraph_mode() or not program_translator.enable_declarative:
114
            warnings.warn(
115
                "The decorator 'dygraph_to_static_func' doesn't work in "
116
                "dygraph mode or set ProgramTranslator.enable to False. "
117 118 119 120
                "We will just return dygraph output.")
            return dygraph_func(*args, **kwargs)
        static_func = program_translator.get_func(dygraph_func)
        return static_func(*args, **kwargs)
121 122 123 124

    return __impl__


125
dygraph_to_static_func = wrap_decorator(_dygraph_to_static_func_)
126

127

128 129 130 131 132
def _declarative_(dygraph_func):
    """
    Converts imperative dygraph APIs into declarative function APIs. Decorator
    @declarative handles the Program and Executor of static mode and returns
    the result as a dygraph VarBase.
133

134 135
    Args:
        dygraph_func (callable): callable imperative function.
136

137 138
    Returns:
        VarBase: containing the numerical result.
139

140 141
    Examples:
        .. code-block:: python
142

143 144 145
          import paddle.fluid as fluid
          import numpy as np
          from paddle.fluid.dygraph.jit import declarative
146

147

148 149 150 151 152 153 154 155
          @declarative
          def func(x):
              x = fluid.dygraph.to_variable(x)
              if fluid.layers.mean(x) < 0:
                  x_v = x - 1
              else:
                  x_v = x + 1
              return x_v
156

157 158 159
          x = np.ones([1, 2])
          x_v = func(x)
          print(x_v.numpy()) # [[2. 2.]]
160

161
    """
162 163

    def __impl__(*args, **kwargs):
164
        program_translator = ProgramTranslator()
165
        if not program_translator.enable_declarative:
166
            warnings.warn(
167 168
                "The decorator 'declarative' doesn't work when setting ProgramTranslator.enable=False. "
                "We will just return dygraph output.")
169
            return dygraph_func(*args, **kwargs)
170
        return program_translator.get_output(dygraph_func, *args, **kwargs)
171

172
    return __impl__
173 174


175
declarative = wrap_decorator(_declarative_)
176 177


178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659
class SaveLoadConfig(object):
    """
    The additional configuration options may be used in function 
    :ref:`api_imperative_jit_save` that save :ref:`api_imperative_TranslatedLayer` 
    or used in function :ref:`api_imperative_jit_load` that 
    load :ref:`api_imperative_TranslatedLayer` .
    
    Examples:
        1. Using ``SaveLoadConfig`` when saving model

        .. code-block:: python

            import numpy as np
            import paddle.fluid as fluid
            from paddle.fluid.dygraph import Linear
            from paddle.fluid.dygraph import declarative

            class SimpleNet(fluid.dygraph.Layer):
                def __init__(self, in_size, out_size):
                    super(SimpleNet, self).__init__()
                    self._linear = Linear(in_size, out_size)

                @declarative
                def forward(self, x):
                    y = self._linear(x)
                    z = self._linear(y)
                    return z

            # enable dygraph mode
            fluid.enable_dygraph() 

            # train model
            net = SimpleNet(8, 8)
            adam = fluid.optimizer.AdamOptimizer(learning_rate=0.1, parameter_list=net.parameters())
            x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
            for i in range(10):
                out = net(x)
                loss = fluid.layers.mean(out)
                loss.backward()
                adam.minimize(loss)
                net.clear_gradients()

            # use SaveLoadconfig when saving model
            model_path = "simplenet.example.model"
            configs = fluid.dygraph.jit.SaveLoadConfig()
            configs.model_filename = "__simplenet__"
            fluid.dygraph.jit.save(
                layer=net,
                model_path=model_path,
                input_spec=[x],
                configs=configs)

        2. Using ``SaveLoadConfig`` when loading model

        .. code-block:: python

            import numpy as np
            import paddle.fluid as fluid

            # enable dygraph mode
            fluid.enable_dygraph() 

            # use SaveLoadconfig when loading model
            model_path = "simplenet.example.model"
            configs = fluid.dygraph.jit.SaveLoadConfig()
            configs.model_filename = "__simplenet__"
            infer_net = fluid.dygraph.jit.load(model_path, configs=configs)
            # inference
            x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
            pred = infer_net(x)
    """

    def __init__(self):
        self._output_spec = None
        self._model_filename = None
        self._params_filename = None
        self._separate_params = False

        # 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):
        """
        Selects the output targets of the saved model ( :ref:`api_imperative_TranslatedLayer` ).
        By default, all return variables of original Layer's forward function
        are kept as the output of the saved TranslatedLayer.

        The ``output_spec`` type should be list[Variable]. If the provided ``output_spec``
        list is not all output variables, the saved model will be pruned according to the
        given ``output_spec`` list.

        .. note::
            The ``output_spec`` is only used when saving model.

        Examples:
            .. code-block:: python

                import numpy as np
                import paddle.fluid as fluid
                from paddle.fluid.dygraph import Linear
                from paddle.fluid.dygraph import declarative

                class SimpleNet(fluid.dygraph.Layer):
                    def __init__(self, in_size, out_size):
                        super(SimpleNet, self).__init__()
                        self._linear = Linear(in_size, out_size)

                    @declarative
                    def forward(self, x):
                        y = self._linear(x)
                        z = self._linear(y)
                        loss = fluid.layers.mean(z)
                        return z, loss

                # enable dygraph mode
                fluid.enable_dygraph() 

                # train model
                net = SimpleNet(8, 8)
                adam = fluid.optimizer.AdamOptimizer(learning_rate=0.1, parameter_list=net.parameters())
                x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
                for i in range(10):
                    out, loss = net(x)
                    loss.backward()
                    adam.minimize(loss)
                    net.clear_gradients()

                # use SaveLoadconfig.output_spec
                model_path = "simplenet.example.model.output_spec"
                configs = fluid.dygraph.jit.SaveLoadConfig()
                # only keep the predicted output in saved model, diccard loss
                configs.output_spec = [out]

                fluid.dygraph.jit.save(
                    layer=net,
                    model_path=model_path,
                    input_spec=[x],
                    configs=configs)

                infer_net = fluid.dygraph.jit.load(model_path, configs=configs)
                x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
                # only have the predicted output
                pred = infer_net(x)
        """
        return self._output_spec

    @output_spec.setter
    def output_spec(self, spec):
        if not isinstance(spec, list):
            raise TypeError(
                "The SaveLoadConfig.output_spec should be 'list', but received input type is %s."
                % type(input))
            for var in spec:
                if not isinstance(var, core.VarBase):
                    raise TypeError(
                        "The element in SaveLoadConfig.output_spec list should be 'Variable', but received element's type is %s."
                        % type(var))
        self._output_spec = spec

    @property
    def model_filename(self):
        """
        The name of file to save the translated program of target Layer.
        Default filename is :code:`__model__` .

        Exampels:
            .. code-block:: python

                import numpy as np
                import paddle.fluid as fluid
                from paddle.fluid.dygraph import Linear
                from paddle.fluid.dygraph import declarative

                class SimpleNet(fluid.dygraph.Layer):
                    def __init__(self, in_size, out_size):
                        super(SimpleNet, self).__init__()
                        self._linear = Linear(in_size, out_size)

                    @declarative
                    def forward(self, x):
                        y = self._linear(x)
                        z = self._linear(y)
                        return z

                # enable dygraph mode
                fluid.enable_dygraph() 

                # train model
                net = SimpleNet(8, 8)
                adam = fluid.optimizer.AdamOptimizer(learning_rate=0.1, parameter_list=net.parameters())
                x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
                for i in range(10):
                    out = net(x)
                    loss = fluid.layers.mean(out)
                    loss.backward()
                    adam.minimize(loss)
                    net.clear_gradients()

                model_path = "simplenet.example.model.model_filename"
                configs = fluid.dygraph.jit.SaveLoadConfig()
                configs.model_filename = "__simplenet__"

                # saving with configs.model_filename
                fluid.dygraph.jit.save(
                    layer=net,
                    model_path=model_path,
                    input_spec=[x],
                    configs=configs)
                # [result] the saved model directory contains:
                # __simplenet__  __variables__  __variables.info__

                # loading with configs.model_filename
                infer_net = fluid.dygraph.jit.load(model_path, configs=configs)
                x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
                pred = infer_net(x)
        """
        return self._model_filename

    @model_filename.setter
    def model_filename(self, filename):
        if not isinstance(filename, six.string_types):
            raise TypeError(
                "The SaveLoadConfig.model_filename should be str, but received input's type is %s."
                % type(filename))
        if len(filename) == 0:
            raise ValueError(
                "The SaveLoadConfig.model_filename is empty string.")
        self._model_filename = filename

    @property
    def params_filename(self):
        """
        The name of file to save all persistable variables in target Layer. 
        Default file name is :code:`__variables__` .
        
        Exampels:
            .. code-block:: python

                import numpy as np
                import paddle.fluid as fluid
                from paddle.fluid.dygraph import Linear
                from paddle.fluid.dygraph import declarative

                class SimpleNet(fluid.dygraph.Layer):
                    def __init__(self, in_size, out_size):
                        super(SimpleNet, self).__init__()
                        self._linear = Linear(in_size, out_size)

                    @declarative
                    def forward(self, x):
                        y = self._linear(x)
                        z = self._linear(y)
                        return z

                # enable dygraph mode
                fluid.enable_dygraph() 

                # train model
                net = SimpleNet(8, 8)
                adam = fluid.optimizer.AdamOptimizer(learning_rate=0.1, parameter_list=net.parameters())
                x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
                for i in range(10):
                    out = net(x)
                    loss = fluid.layers.mean(out)
                    loss.backward()
                    adam.minimize(loss)
                    net.clear_gradients()

                model_path = "simplenet.example.model.params_filename"
                configs = fluid.dygraph.jit.SaveLoadConfig()
                configs.params_filename = "__params__"

                # saving with configs.params_filename
                fluid.dygraph.jit.save(
                    layer=net,
                    model_path=model_path,
                    input_spec=[x],
                    configs=configs)
                # [result] the saved model directory contains:
                # __model__  __params__  __variables.info__

                # loading with configs.params_filename
                infer_net = fluid.dygraph.jit.load(model_path, configs=configs)
                x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
                pred = infer_net(x)
        """
        return self._params_filename

    @params_filename.setter
    def params_filename(self, filename):
        if not isinstance(filename, six.string_types):
            raise TypeError(
                "The SaveLoadConfig.params_filename should be str, but received input's type is %s."
                % type(filename))
        if len(filename) == 0:
            raise ValueError(
                "The SaveLoadConfig.params_filename is empty string.")
        self._params_filename = filename

    # NOTE: [why not use params_filename=None control params saved separately]
    # The new save interface does not recommend parameters to be saved separately. 
    # Here, the concept should be separated as clearly as possible. 
    # Setting params_filename=None only means that the saved file name is set 
    # and without any other meaning. New separate_params control for file saved
    # separately can makes the concept clearer.
    @property
    def separate_params(self):
        """
        Configure whether to save the Layer parameters as separete files.
        (In order to be compatible with the behavior of :ref:`api_fluid_io_save_inference_model` )

        If True, each parameter will be saved to a file separately, the file name is the parameter name,
        and the SaveLoadConfig.params_filename configuration will not take effect. Default False.

        Examples:
            .. code-block:: python

                import numpy as np
                import paddle.fluid as fluid
                from paddle.fluid.dygraph import Linear
                from paddle.fluid.dygraph import declarative

                class SimpleNet(fluid.dygraph.Layer):
                    def __init__(self, in_size, out_size):
                        super(SimpleNet, self).__init__()
                        self._linear = Linear(in_size, out_size)

                    @declarative
                    def forward(self, x):
                        y = self._linear(x)
                        z = self._linear(y)
                        return z

                # enable dygraph mode
                fluid.enable_dygraph() 

                # train model
                net = SimpleNet(8, 8)
                adam = fluid.optimizer.AdamOptimizer(learning_rate=0.1, parameter_list=net.parameters())
                x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
                for i in range(10):
                    out = net(x)
                    loss = fluid.layers.mean(out)
                    loss.backward()
                    adam.minimize(loss)
                    net.clear_gradients()

                model_path = "simplenet.example.model.separate_params"
                configs = fluid.dygraph.jit.SaveLoadConfig()
                configs.separate_params = True

                # saving with configs.separate_params
                fluid.dygraph.jit.save(
                    layer=net,
                    model_path=model_path,
                    input_spec=[x],
                    configs=configs)
                # [result] the saved model directory contains:
                # linear_0.b_0  linear_0.w_0  __model__  __variables.info__

                # loading with configs.params_filename
                infer_net = fluid.dygraph.jit.load(model_path, configs=configs)
                x = fluid.dygraph.to_variable(np.random.random((4, 8)).astype('float32'))
                pred = infer_net(x)
        """
        return self._separate_params

    @separate_params.setter
    def separate_params(self, value):
        if not isinstance(value, bool):
            raise TypeError(
                "The SaveLoadConfig.separate_params should be bool value, but received input's type is %s."
                % type(value))
        self._separate_params = value


@switch_to_static_graph
def save(layer, model_path, input_spec=None, configs=None):
    """
    Saves input declarative Layer as :ref:`api_imperative_TranslatedLayer` 
    format model, which can be used for inference or fine-tuning after loading.

    It will save the translated program and all related persistable 
    variables of input declarative Layer to given ``model_path``.
    
    The default saved translated program file name is ``__model__``,
    and the default saved persistable variables file name is ``__variables__``,
    and it also saved some additional variable description information to file 
    ``__varibales.info__``, these additional information is used in fine-tuning.

    The saved model can be loaded by follow APIs:
      - :ref:`api_imperative_jit_load`
      - :ref:`api_fluid_io_load_inference_model` (need pass ``params_filename='__variables__'``)
      - Other C++ inference APIs

    Args:
        layer (Layer): the Layer to be saved. The Layer should be decorated by `@declarative`.
        model_path (str): the directory to save the model.
        input_spec (list[Varibale], optional): Describes the input of the saved model. 
            It is the example inputs that will be passed to saved TranslatedLayer's forward
            function. If None, all input variables of the original Layer's forward function
            would be the inputs of the saved model. Default None.
        configs (SaveLoadConfig, optional): :ref:`api_imperative_jit_saveLoadConfig` object
            that specifies additional configuration options. Default None.
    Returns:
        None

    Examples:
        .. code-block:: python

            import numpy as np
            import paddle.fluid as fluid
            from paddle.fluid.dygraph import Linear
            from paddle.fluid.dygraph import declarative

            BATCH_SIZE = 32
            BATCH_NUM = 20

            def random_batch_reader():
                def _get_random_images_and_labels(image_shape, label_shape):
                    image = np.random.random(size=image_shape).astype('float32')
                    label = np.random.random(size=label_shape).astype('int64')
                    return image, label

                def __reader__():
                    for _ in range(BATCH_NUM):
                        batch_image, batch_label = _get_random_images_and_labels(
                            [BATCH_SIZE, 784], [BATCH_SIZE, 1])
                        yield batch_image, batch_label

                return __reader__

            class LinearNet(fluid.dygraph.Layer):
                def __init__(self, in_size, out_size):
                    super(LinearNet, self).__init__()
                    self._linear = Linear(in_size, out_size)

                @declarative
                def forward(self, x):
                    return self._linear(x)

            # enable dygraph mode
            fluid.enable_dygraph() 

            # create network
            net = LinearNet(784, 1)
            adam = fluid.optimizer.AdamOptimizer(learning_rate=0.1, parameter_list=net.parameters())
            # create data loader
            train_loader = fluid.io.DataLoader.from_generator(capacity=5)
            train_loader.set_batch_generator(random_batch_reader())
            # train
            for data in train_loader():
                img, label = data
                label.stop_gradient = True

                cost = net(img)

                loss = fluid.layers.cross_entropy(cost, label)
                avg_loss = fluid.layers.mean(loss)

                avg_loss.backward()
                adam.minimize(avg_loss)
                net.clear_gradients()

            # save model
            model_path = "linear.example.model"
            fluid.dygraph.jit.save(
                layer=net,
                model_path=model_path,
                input_spec=[img])
    """

    def get_inout_spec(all_vars, target_vars, return_name=False):
660
        result_list = []
661
        valid_var_dict = {}
662
        valid_vars = [var for var in all_vars if isinstance(var, Variable)]
663 664 665 666 667 668 669 670
        for var in valid_vars:
            valid_var_dict[var.name] = var
        if target_vars:
            for i, var in enumerate(target_vars):
                # check target var whether exists
                if var.name not in valid_var_dict:
                    raise RuntimeError(
                        "The variable to feed/fetch are not exist.")
671
                result_list.append(valid_var_dict[var.name])
672
        else:
673
            result_list = valid_vars
674
        if return_name:
675
            result_list = [var.name for var in target_vars]
676

677
        return result_list
678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974

    # 1. input check
    prog_translator = ProgramTranslator()
    if not prog_translator.enable:
        raise RuntimeError(
            "The paddle.imperative.jit.save doesn't work when setting ProgramTranslator.enable=False."
        )
    if not isinstance(layer, Layer):
        raise TypeError(
            "The input layer of paddle.imperative.jit.save should be 'Layer', but received layer type is %s."
            % type(layer))

    if configs is None:
        configs = SaveLoadConfig()

    if input_spec is not None:
        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))
        for var in input_spec:
            if not isinstance(var, core.VarBase):
                raise TypeError(
                    "The element in input_spec list should be 'Variable', but received element's type is %s."
                    % type(var))

    # 2. get program of declarative Layer.forward
    prog_cache = prog_translator.get_program_cache()
    # make dummy args & kwargs, to get excepted FunctionSpec
    layer_func = FunctionSpec(type(layer).forward, [layer], {})
    concrete_program, _ = prog_cache.get_program(layer_func)

    # 3. share parameters from Layer to scope & record var info
    scope = core.Scope()
    state_dict = layer.state_dict()
    extra_var_info = dict()
    for structured_name, param_or_buffer in state_dict.items():
        # 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
        extra_info_dict = dict()
        extra_info_dict['structured_name'] = structured_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

    # 4. build input & output spec
    input_var_names = get_inout_spec(concrete_program.inputs, input_spec, True)
    output_vars = get_inout_spec(concrete_program.outputs, configs.output_spec)

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

    # VARIABLE_FILENAME keep nameing style consistent with '__model__'
    if configs.params_filename is None:
        configs.params_filename = VARIABLE_FILENAME

    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=configs.model_filename,
            params_filename=None
            if configs.separate_params else configs.params_filename,
            export_for_deployment=configs._export_for_deployment,
            program_only=configs._program_only)

        # NOTE: [ 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 `__variables.info__`
        extra_var_info_path = os.path.join(model_path, EXTRA_VAR_INFO_FILENAME)
        with open(extra_var_info_path, 'wb') as f:
            pickle.dump(extra_var_info, f, protocol=2)


@dygraph_only
def load(model_path, configs=None):
    """
    :api_attr: imperative

    Load model saved by :ref:`api_imperative_jit_save` or :ref:`api_fluid_io_save_inference_model`
    as :ref:`api_imperative_TranslatedLayer`, then performing inference or fine-tune training.

    .. note::
        For some historical reasons, if you load model saved by :ref:`api_fluid_io_save_inference_model`,
        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.
        2. All saved model's feed targets need to be passed into TranslatedLayer's forwrad function.
        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:
        model_path (str): The directory path where the model is saved.
        configs (SaveLoadConfig, optional): :ref:`api_imperative_jit_saveLoadConfig` object that specifies 
            additional configuration options. Default None.

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

    Examples:
        1. Load model saved by :ref:`api_imperative_jit_save` then performing inference and fine-tune training.

        .. code-block:: python

            import numpy as np
            import paddle.fluid as fluid
            from paddle.fluid.dygraph import Linear
            from paddle.fluid.dygraph import declarative

            BATCH_SIZE = 32
            BATCH_NUM = 20

            def random_batch_reader():
                def _get_random_images_and_labels(image_shape, label_shape):
                    image = np.random.random(size=image_shape).astype('float32')
                    label = np.random.random(size=label_shape).astype('int64')
                    return image, label

                def __reader__():
                    for _ in range(BATCH_NUM):
                        batch_image, batch_label = _get_random_images_and_labels(
                            [BATCH_SIZE, 784], [BATCH_SIZE, 1])
                        yield batch_image, batch_label

                return __reader__

            class LinearNet(fluid.dygraph.Layer):
                def __init__(self, in_size, out_size):
                    super(LinearNet, self).__init__()
                    self._linear = Linear(in_size, out_size)

                @declarative
                def forward(self, x):
                    return self._linear(x)

            # enable dygraph mode
            fluid.enable_dygraph() 

            # 1. train & save model.
            # create network
            net = LinearNet(784, 1)
            adam = fluid.optimizer.AdamOptimizer(learning_rate=0.1, parameter_list=net.parameters())
            # create data loader
            train_loader = fluid.io.DataLoader.from_generator(capacity=5)
            train_loader.set_batch_generator(random_batch_reader())
            # train
            for data in train_loader():
                img, label = data
                label.stop_gradient = True

                cost = net(img)

                loss = fluid.layers.cross_entropy(cost, label)
                avg_loss = fluid.layers.mean(loss)

                avg_loss.backward()
                adam.minimize(avg_loss)
                net.clear_gradients()

            model_path = "linear.example.model"
            fluid.dygraph.jit.save(
                layer=net,
                model_path=model_path,
                input_spec=[img])

            # 2. load model & inference
            # load model
            infer_net = fluid.dygraph.jit.load(model_path)
            # inference
            x = fluid.dygraph.to_variable(np.random.random((1, 784)).astype('float32'))
            pred = infer_net(x)

            # 3. load model & fine-tune
            # load model
            train_net = fluid.dygraph.jit.load(model_path)
            train_net.train()
            adam = fluid.optimizer.AdamOptimizer(learning_rate=0.1, parameter_list=train_net.parameters())
            # create data loader
            train_loader = fluid.io.DataLoader.from_generator(capacity=5)
            train_loader.set_batch_generator(random_batch_reader())
            # fine-tune
            for data in train_loader():
                img, label = data
                label.stop_gradient = True

                cost = train_net(img)

                loss = fluid.layers.cross_entropy(cost, label)
                avg_loss = fluid.layers.mean(loss)

                avg_loss.backward()
                adam.minimize(avg_loss)
                train_net.clear_gradients()

        2. Load model saved by :ref:`api_fluid_io_save_inference_model` then performing and fine-tune training.

        .. code-block:: python

            import numpy as np
            import paddle.fluid as fluid

            BATCH_SIZE = 32
            BATCH_NUM = 20

            def random_batch_reader():
                def _get_random_images_and_labels(image_shape, label_shape):
                    image = np.random.random(size=image_shape).astype('float32')
                    label = np.random.random(size=label_shape).astype('int64')
                    return image, label

                def __reader__():
                    for _ in range(BATCH_NUM):
                        batch_image, batch_label = _get_random_images_and_labels(
                            [BATCH_SIZE, 784], [BATCH_SIZE, 1])
                        yield batch_image, batch_label

                return __reader__

            img = fluid.data(name='img', shape=[None, 784], dtype='float32')
            label = fluid.data(name='label', shape=[None, 1], dtype='int64')
            pred = fluid.layers.fc(input=img, size=10, act='softmax')
            loss = fluid.layers.cross_entropy(input=pred, label=label)
            avg_loss = fluid.layers.mean(loss)

            optimizer = fluid.optimizer.SGD(learning_rate=0.001)
            optimizer.minimize(avg_loss)

            place = fluid.CPUPlace()
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())

            loader = fluid.io.DataLoader.from_generator(
                feed_list=[img, label], capacity=5, iterable=True)
            loader.set_batch_generator(random_batch_reader(), places=place)

            # 1. train and save inference model
            for data in loader():
                exe.run(
                    fluid.default_main_program(),
                    feed=data, 
                    fetch_list=[avg_loss])

            model_path = "fc.example.model"
            fluid.io.save_inference_model(
                model_path, ["img"], [pred], exe)

            # enable dygraph mode
            fluid.enable_dygraph() 

            # 2. load model & inference
            fc = fluid.dygraph.jit.load(model_path)
            x = fluid.dygraph.to_variable(np.random.random((1, 784)).astype('float32'))
            pred = fc(x)

            # 3. load model & fine-tune
            fc = fluid.dygraph.jit.load(model_path)
            fc.train()
            sgd = fluid.optimizer.SGD(learning_rate=0.001,
                                        parameter_list=fc.parameters())

            train_loader = fluid.io.DataLoader.from_generator(capacity=5)
            train_loader.set_batch_generator(
                random_batch_reader(), places=place)

            for data in train_loader():
                img, label = data
                label.stop_gradient = True

                cost = fc(img)

                loss = fluid.layers.cross_entropy(cost, label)
                avg_loss = fluid.layers.mean(loss)

                avg_loss.backward()
                sgd.minimize(avg_loss)
    """
    return TranslatedLayer._construct(model_path, configs)


975
@dygraph_only
Z
Zeng Jinle 已提交
976 977 978 979 980
def _trace(layer,
           inputs,
           feed_prefix='feed_',
           fetch_prefix='fetch_',
           tmp_prefix='t_'):
981
    assert isinstance(layer, Layer)
982 983 984 985 986 987 988 989 990

    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):
991
        original_outputs = layer(*inputs)
992 993 994 995
        if not isinstance(original_outputs, (list, tuple)):
            outputs = [original_outputs]
        else:
            outputs = original_outputs
996
        out_vars = [var for var in outputs]
997

998
        program_desc, feed_names, fetch_names, parameters = tracer.create_program_desc(
Z
Zeng Jinle 已提交
999
            var_list, feed_prefix, out_vars, fetch_prefix, tmp_prefix)
1000 1001 1002 1003 1004
        tracer.reset()

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

1005
    return original_outputs, program, feed_names, fetch_names, parameters
1006 1007 1008 1009


class TracedLayer(object):
    """
1010 1011
    :api_attr: imperative
    
1012 1013 1014 1015 1016
    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.
1017 1018 1019 1020

    TracedLayer would run the static graph model using :code:`Executor`
    and :code:`CompiledProgram` . The static graph model would share
    parameters with the dygraph model.
1021 1022

    All TracedLayer objects should not be created by constructor and should
1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033
    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
1034
        self._params = parameters
1035 1036 1037 1038 1039

        self._place = _current_expected_place()

        self._scope = core.Scope()
        for p in parameters:
1040
            src_tensor = p.value().get_tensor()
1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063
            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):
        """
1064
        This method is the only allowed method to create TracedLayer object.
1065 1066 1067 1068
        It would call the :code:`layer(*inputs)` method to run the dygraph
        model and convert it into a static graph model.

        Args:
1069
            layer (dygraph.Layer): the layer object to be traced.
1070 1071
            inputs (list(Tensor)|tuple(Tensor)|Tensor): the input tensors of
                the layer object.
1072 1073

        Returns:
1074
            tuple: A tuple of 2 items, whose the first item is the output of
1075 1076
                :code:`layer(*inputs)` , and the second item is the created
                TracedLayer object.
1077

1078
        Examples:
1079 1080 1081
            .. code-block:: python:

                import paddle.fluid as fluid
1082
                from paddle.fluid.dygraph import Linear, to_variable, TracedLayer
1083 1084 1085
                import numpy as np

                class ExampleLayer(fluid.dygraph.Layer):
1086 1087 1088
                    def __init__(self):
                        super(ExampleLayer, self).__init__()
                        self._fc = Linear(3, 10)
1089 1090 1091 1092 1093

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

                with fluid.dygraph.guard():
1094
                    layer = ExampleLayer()
1095 1096 1097
                    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])
1098 1099 1100 1101 1102 1103 1104 1105 1106

                    # 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')
1107
        """
1108 1109 1110 1111
        assert isinstance(
            layer, Layer
        ), "The type of 'layer' in fluid.dygraph.jit.TracedLayer.trace must be fluid.dygraph.Layer, but received {}.".format(
            type(layer))
1112 1113
        outs, prog, feed, fetch, parameters = _trace(layer, inputs)
        traced = TracedLayer(prog, parameters, feed, fetch)
1114 1115 1116 1117 1118 1119 1120
        return outs, traced

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

        Args:
1121
            build_strategy (BuildStrategy, optional): build strategy of
1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132
                :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
1133
                from paddle.fluid.dygraph import Linear, to_variable, TracedLayer
1134 1135 1136
                import numpy as np

                class ExampleLayer(fluid.dygraph.Layer):
1137 1138 1139
                    def __init__(self):
                        super(ExampleLayer, self).__init__()
                        self._fc = Linear(3, 10)
1140 1141 1142 1143 1144

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

                with fluid.dygraph.guard():
1145
                    layer = ExampleLayer()
1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160
                    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"
1161 1162 1163 1164 1165 1166 1167 1168
        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))
1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186
        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):
1187
                feed_dict[name] = x.value().get_tensor()
1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209
        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):
        """
1210 1211
        Save the TracedLayer to a model for inference. The saved
        inference model can be loaded by C++ inference APIs.
1212 1213

        Args:
1214
            dirname (str): the directory to save the inference model.
1215
            feed (list[int], optional): the input variable indices of the saved
1216
                inference model. If None, all input variables of the
1217 1218 1219 1220 1221 1222 1223 1224
                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:
1225
            None
1226 1227 1228 1229 1230

        Examples:
            .. code-block:: python:

                import paddle.fluid as fluid
1231
                from paddle.fluid.dygraph import Linear, to_variable, TracedLayer
1232 1233 1234
                import numpy as np

                class ExampleLayer(fluid.dygraph.Layer):
1235 1236 1237
                    def __init__(self):
                        super(ExampleLayer, self).__init__()
                        self._fc = Linear(3, 10)
1238 1239 1240 1241

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

1242 1243 1244
                save_dirname = './saved_infer_model'
                in_np = np.random.random([2, 3]).astype('float32')

1245
                with fluid.dygraph.guard():
1246
                    layer = ExampleLayer()
1247 1248
                    in_var = to_variable(in_np)
                    out_dygraph, static_layer = TracedLayer.trace(layer, inputs=[in_var])
1249
                    static_layer.save_inference_model(save_dirname, feed=[0], fetch=[0])
1250 1251

                place = fluid.CPUPlace()
1252 1253
                exe = fluid.Executor(place)
                program, feed_vars, fetch_vars = fluid.io.load_inference_model(save_dirname,
1254
                                                    exe)
1255 1256 1257

                fetch, = exe.run(program, feed={feed_vars[0]: in_np}, fetch_list=fetch_vars)
                print(fetch.shape) # (2, 10)
1258
        """
1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273
        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")

1274
        from paddle.fluid.io import save_inference_model
1275 1276 1277 1278 1279

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

1280
            return [all_vars[idx] for idx in partial_vars]
1281 1282 1283 1284 1285 1286 1287 1288 1289 1290

        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)

1291
            save_inference_model(
1292 1293 1294 1295 1296
                dirname=dirname,
                feeded_var_names=feeded_var_names,
                target_vars=target_vars,
                executor=self._exe,
                main_program=self._program.clone())