translated_layer.py

# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import os
import pickle

import numpy as np

import paddle
from paddle import _legacy_C_ops
from paddle.fluid import backward, core, framework, unique_name
from paddle.fluid.dygraph.base import switch_to_static_graph
from paddle.fluid.executor import (
    _is_dy2st_enable_standalone_executor,
    _is_enable_standalone_executor,
)
from paddle.fluid.framework import OpProtoHolder, _non_static_mode
from paddle.jit.dy2static.partial_program import (
    LazyInitialized,
    add_build_strategy_for,
)
from paddle.nn.layer import layers

from .dy2static.utils import _out_grad_names, _param_grad_names

__all__ = []

INFER_MODEL_SUFFIX = ".pdmodel"
INFER_PARAMS_SUFFIX = ".pdiparams"
INFER_PARAMS_INFO_SUFFIX = ".pdiparams.info"
INFER_PROPERTY_SUFFIX = '.meta'

LOADED_VAR_SUFFIX = "load"
PARAMETER_NAME_PREFIX = "param"
BUFFER_NAME_PREFIX = "buffer"


def _load_program_desc(model_file_path):
    # 1. parse program desc
    with open(model_file_path, "rb") as f:
        program_desc_str = f.read()

    program_desc = core.ProgramDesc(program_desc_str)
    if not core._is_program_version_supported(program_desc._version()):
        raise ValueError(
            "Unsupported program version: %d\n" % program_desc._version()
        )
    return program_desc


def _is_persistable(var_desc):
    if (
        var_desc.type() == core.VarDesc.VarType.FEED_MINIBATCH
        or var_desc.type() == core.VarDesc.VarType.FETCH_LIST
        or var_desc.type() == core.VarDesc.VarType.READER
        or var_desc.type() == core.VarDesc.VarType.RAW
    ):
        return False
    return var_desc.persistable()


def _is_parameter(persistable_var_desc, program_desc):
    # 1. firstly, param should be input of op
    input_ops = []  # op can be repeated
    for block_idx in range(program_desc.num_blocks()):
        block = program_desc.block(block_idx)
        for op_idx in range(block.op_size()):
            op = block.op(op_idx)
            # NOTE: parameter is the input of a certain op
            if persistable_var_desc.name() in op.input_arg_names():
                input_ops.append(op)
    # 2. secondly, param should not be output of op or be same op's output
    for block_idx in range(program_desc.num_blocks()):
        block = program_desc.block(block_idx)
        for op_idx in range(block.op_size()):
            op = block.op(op_idx)
            if persistable_var_desc.name() in op.output_arg_names():
                # such as batch_norm_op
                if op in input_ops:
                    continue
                else:
                    return False
    return True


def _get_persistable_vars(program_desc):
    persistable_vars = []
    for i in range(program_desc.num_blocks()):
        block = program_desc.block(i)
        persistable_vars.extend(list(filter(_is_persistable, block.all_vars())))
    return persistable_vars


def _get_persistable_var_names(program_desc):
    """
    Get all persistable variable names in ProgramDesc.
    """
    var_names = []
    persistable_vars = _get_persistable_vars(program_desc)
    for var in persistable_vars:
        var_names.append(var.name())
    return var_names


def _get_all_var_names(program_desc):
    all_var_names = set()
    for i in range(program_desc.num_blocks()):
        block = program_desc.block(i)
        for var in block.all_vars():
            all_var_names.add(var.name())
    return all_var_names


@switch_to_static_graph
def _append_loaded_suffix(name):
    """
    Append loaded suffix to the given variable name
    e.g. x ==> x.load_0, x.load_0 ==> x.load_0.load_0
    """
    suffix = LOADED_VAR_SUFFIX
    new_name = unique_name.generate_with_ignorable_key('.'.join((name, suffix)))
    return new_name


@switch_to_static_graph
def _generate_unique_var_name(prefix):
    return unique_name.generate_with_ignorable_key(prefix)


def _append_loaded_suffix_to_var(program_desc):
    suffix_varname_dict = {}
    persistable_vars = _get_persistable_vars(program_desc)
    for var_desc in persistable_vars:
        old_name = var_desc.name()
        new_name = _append_loaded_suffix(var_desc.name())
        suffix_varname_dict[new_name] = old_name
        var_desc.set_name(new_name)
        for block_idx in range(program_desc.num_blocks()):
            block = program_desc.block(block_idx)
            block._rename_var(old_name.encode(), new_name.encode())
            for op_idx in range(block.op_size()):
                op = block.op(op_idx)
                op._rename_input(old_name, new_name)
                op._rename_output(old_name, new_name)
    return suffix_varname_dict


@switch_to_static_graph
def _generate_unique_var_name_sync_with_main_program(prefix):
    return unique_name.generate(prefix)


def _get_loaded_var_new_old(program_desc, all_new_old_dict_all):
    new_old_dict = {}
    persistable_vars = _get_persistable_vars(program_desc)
    for var_desc in persistable_vars:
        name_new = var_desc.name()
        new_old_dict[name_new] = all_new_old_dict_all[name_new]
    return new_old_dict


def _rename_var_program_desc(program_desc, include=None, exclude=None):
    """
    Change the name of the loaded variables.Use 'unique_name.generate' to avoid duplication.
    It is used when loading multiple program during inference.

    e.g. linear_0.tmp_3 ==> linear_0.tmp_1, x ==> x_0. For double grad, x@GRAD ==> x_0@GRAD
    If 'include' is not `None`,variables in include and the corresponding
      double grad variables (if exist) are renamed.
    If 'exclude' is not `None`,variables that are in exclude and the
      corresponding double grad variables (if exist) are not renamed.

    Args:
        program_desc(ProgramDesc):the variables in it will be modified.
        include(List):list of names of variables.
        exclude(List):list of names of variables.

    Returns:
        tuple of (dict_rename_var_new_old, dict_rename_var_old_new)
        dict_rename_var_new_old is a dict mapping from new name to old name
        dict_rename_var_old_new is a dict mapping from old name to new name
    """
    dict_rename_var_old_new = {}
    dict_rename_var_new_old = {}
    old_names = []
    # Store all old names
    for b_idx in range(program_desc.num_blocks()):
        cur_block = program_desc.block(b_idx)
        for var in cur_block.all_vars():
            old_names.append(var.name())

    # Create dict_rename_var_new_old and dict_rename_var_old_new for non double
    # grad variables
    has_double_grad = False
    for b_idx in range(program_desc.num_blocks()):
        cur_block = program_desc.block(b_idx)
        for var_idx, var in enumerate(cur_block.all_vars()):
            name_old = var.name()
            is_double_grad_var = "@GRAD" in name_old
            has_double_grad = has_double_grad or is_double_grad_var
            should_rename = (
                (include is None or name_old in include)
                and (exclude is None or name_old not in exclude)
                and not is_double_grad_var
            )
            if should_rename:
                temp_name = name_old.split('_')
                if len(temp_name) > 1 and temp_name[-1].isnumeric():
                    temp_name = "_".join(temp_name[:-1])
                else:
                    temp_name = name_old
                while True:
                    name_new = _generate_unique_var_name_sync_with_main_program(
                        temp_name
                    )
                    if (
                        name_new
                        not in old_names[:var_idx] + old_names[var_idx + 1 :]
                    ):
                        break
            else:
                name_new = name_old
            if name_old != name_new:
                cur_block._rename_var(name_old.encode(), name_new.encode())
            if not is_double_grad_var:
                dict_rename_var_old_new[name_old] = name_new
                dict_rename_var_new_old[name_new] = name_old

    # Handle double grad names
    if has_double_grad:
        double_grad_rename_dict = {}
        for name_old in dict_rename_var_old_new:
            for b_idx in range(program_desc.num_blocks()):
                cur_block = program_desc.block(b_idx)
                for var_idx, var in enumerate(cur_block.all_vars()):
                    var_name = var.name()
                    if "@GRAD" in var_name and name_old in var_name:
                        new_var_name = var_name.replace(
                            name_old, dict_rename_var_old_new[name_old]
                        )
                        double_grad_rename_dict[var_name] = new_var_name
        for var_name in double_grad_rename_dict:
            dict_rename_var_old_new[var_name] = double_grad_rename_dict[
                var_name
            ]
            dict_rename_var_new_old[
                double_grad_rename_dict[var_name]
            ] = var_name

    # Rename on program desc
    for b_idx in range(program_desc.num_blocks()):
        cur_block = program_desc.block(b_idx)
        for op_idx in range(cur_block.op_size()):
            op = cur_block.op(op_idx)
            for input_arg_name in op.input_arg_names():
                if input_arg_name in dict_rename_var_old_new:
                    if (
                        input_arg_name
                        != dict_rename_var_old_new[input_arg_name]
                    ):
                        op._rename_input(
                            input_arg_name,
                            dict_rename_var_old_new[input_arg_name],
                        )
                        if cur_block.has_var(input_arg_name.encode()):
                            cur_block._rename_var(
                                input_arg_name.encode(),
                                dict_rename_var_old_new[
                                    input_arg_name
                                ].encode(),
                            )
            for output_arg_name in op.output_arg_names():
                if output_arg_name in dict_rename_var_old_new:
                    if (
                        output_arg_name
                        != dict_rename_var_old_new[output_arg_name]
                    ):
                        op._rename_output(
                            output_arg_name,
                            dict_rename_var_old_new[output_arg_name],
                        )
                        if cur_block.has_var(output_arg_name.encode()):
                            cur_block._rename_var(
                                output_arg_name.encode(),
                                dict_rename_var_old_new[
                                    output_arg_name
                                ].encode(),
                            )
    program_desc.flush()
    return dict_rename_var_new_old, dict_rename_var_old_new


@switch_to_static_graph
def _build_program_by_desc(program_desc):
    prog = framework.Program()
    prog.desc = program_desc
    prog.blocks = [
        framework.Block(prog, i) for i in range(prog.desc.num_blocks())
    ]
    prog._sync_with_cpp()
    return prog


def _change_is_test_status(program_desc, is_test):
    # change all `is_test` attributes
    for i in range(program_desc.num_blocks()):
        block = program_desc.block(i)
        for j in range(block.op_size()):
            op = block.op(j)
            if op.has_attr('is_test'):
                op._set_attr('is_test', is_test)


class _ProgramHolder:
    """
    Holds the execution information of a Program.

    _ProgramHolder is the execution unit of TranslatedLayer,
    if TranslatedLayer contains multiple _ProgramHolder,
    it can execute multiple methods

    _ProgramHolder is an internal concept.
    """

    def __init__(self, program_desc):
        super().__init__()

        # input, output, persistable, double_grads var info
        self._input_descs = []
        self._output_descs = []
        self._double_grad_descs = []
        self._persistable_names = []

        # execution scope
        self._inner_scope = core.Scope()

        # append suffix var name dict
        self._suffix_varname_dict = None
        # forward program
        self._infer_program_desc = self._preprocess(program_desc)
        # forward + backward program
        self._train_program_desc = self._append_backward_desc(
            self._infer_program_desc
        )

    # forward:
    @switch_to_static_graph
    def _create_forward_train_program(self):
        whole_program = _build_program_by_desc(self._train_program_desc)
        end_op_index = self._infer_program_desc.block(0).op_size()
        if end_op_index > 0:
            return add_build_strategy_for(whole_program, 0, end_op_index)
        else:
            return whole_program

    @LazyInitialized
    def _forward_program_desc(self):
        return self._create_forward_train_program().desc

    # backward
    @switch_to_static_graph
    def _create_backward_train_program(self):
        whole_program = _build_program_by_desc(self._train_program_desc)
        start_op_index = self._infer_program_desc.block(0).op_size() + len(
            self._output_descs
        )
        end_op_index = whole_program.desc.block(0).op_size()
        if start_op_index < end_op_index:
            return add_build_strategy_for(
                whole_program, start_op_index, end_op_index
            )
        else:
            return paddle.static.Program()

    @LazyInitialized
    def _backward_program_desc(self):
        return self._create_backward_train_program().desc

    @property
    def infer_program(self):
        return self._infer_program_desc

    @property
    def train_program(self):
        return self._train_program_desc

    @property
    def forward_program(self):
        return self._forward_program_desc

    @property
    def backward_program(self):
        return self._backward_program_desc

    @property
    def input_descs(self):
        return self._input_descs

    @property
    def output_descs(self):
        return self._output_descs

    @property
    def persistable_names(self):
        return self._persistable_names

    @property
    def double_grad_descs(self):
        return self._double_grad_descs

    @property
    def scope(self):
        return self._inner_scope

    def _preprocess(self, program_desc):
        # rename persistable variables of 'program_desc'
        list_persistable_var = _get_persistable_var_names(program_desc)
        rename_new_old_dict, _ = _rename_var_program_desc(
            program_desc, list_persistable_var
        )
        # 1. Prune original program
        # remove feed, fetch and scale-1 op, remove op_callstack attr
        ops_to_remove = []
        root_block = program_desc.block(0)
        for i in range(root_block.op_size()):
            op = root_block.op(i)
            if op.type() == 'feed':
                ops_to_remove.append(i)
                feed_var_name = op.input('X')[0].encode()
                root_block._remove_var(feed_var_name)
                self._input_descs.append(
                    root_block.find_var(op.output('Out')[0].encode())
                )
            elif op.type() == 'scale' and op.output('Out')[0].startswith(
                'save_infer_model/scale_'
            ):
                ops_to_remove.append(i)
                out_var_name = op.output('Out')[0].encode()
                root_block._remove_var(out_var_name)
                self._output_descs.append(
                    root_block.find_var(op.input('X')[0].encode())
                )
            elif op.type() == 'fetch':
                ops_to_remove.append(i)
                fetch_var_name = op.output('Out')[0].encode()
                root_block._remove_var(fetch_var_name)
                # NOTE: some old pre-train models have no extra scale_op
                if not op.input('X')[0].startswith('save_infer_model/scale_'):
                    self._output_descs.append(
                        root_block.find_var(op.input('X')[0].encode())
                    )
            else:
                if op.has_attr("op_callstack"):
                    op.remove_attr("op_callstack")

        for op_idx in reversed(ops_to_remove):
            root_block._remove_op(op_idx, op_idx + 1)

        for i in range(program_desc.num_blocks()):
            block_desc = program_desc.block(i)
            for var_desc in block_desc.all_vars():
                if "@GRAD" in var_desc.name():
                    self._double_grad_descs.append(var_desc)

        # 2. Input processing, reverse feed vars
        self._input_descs.reverse()

        # 3. Output processing, add scale for outputs
        tmp_program = _build_program_by_desc(program_desc)
        # NOTE: [why need append scale for outputs]
        # When dealing with some more complex pre-training models, there
        # will be situations where the pre-training model has multiple
        # fetch outputs. In the scenario of multiple fetch outputs,
        # there is a special case where multiple outputs of the model
        # may be on the same branch. According to the user's subsequent
        # use, multiple outputs may be associated with multiple branches.
        # These subsequent operations are added in TranslatedLayer is
        # agnostic during initialization, which results in subsequent
        # gradient accumulation operations that are required on the
        # output node in the middle of the branch will not be performed,
        # resulting in error, details see pull request:
        # [https://github.com/PaddlePaddle/Paddle/pull/24627]
        self._append_scale_to_output(tmp_program)

        # 4. Persistable vars processing
        # - append loaded suffix to persistable vars
        # NOTE: [why need to append suffix to persistable vars]
        # Dygraph and static graph mode use the same naming mechanism.
        # If users want to load the model fine-tune, it is possible
        # to add the existing Layer in the loaded model to enhance
        # the network. For example, the original saved model has linear,
        # and later after loading, a new linear is added. At this time,
        # there will be a problem of duplicate names, so here is unified
        # to add the LOADED suffix to the parameters of the model loaded
        self._suffix_varname_dict = _get_loaded_var_new_old(
            program_desc, rename_new_old_dict
        )

        # - get persistable var
        self._persistable_names = _get_persistable_var_names(program_desc)

        return program_desc

    @switch_to_static_graph
    def _append_scale_to_output(self, program):
        # 1. append scale & save var
        scale_output_vars = []
        with framework.program_guard(program):
            for i, out in enumerate(self._output_descs):
                var = program.global_block().var(out.name())
                var = paddle.scale(
                    var, 1.0, name="translated_layer/scale_{}".format(i)
                )
                scale_output_vars.append(var)
        # 2. update output names & descs
        for i, var in enumerate(scale_output_vars):
            self._output_descs[i] = var.desc

    @switch_to_static_graph
    def _get_train_forward_program(self, infer_program_desc):
        program_desc_copy = core.ProgramDesc(infer_program_desc)

        # 1. set all `is_test` attributes to False
        _change_is_test_status(program_desc_copy, False)

        # 2. prepare program and related var
        # NOTE: To reuse backward interfaces, build Program firstly.
        # Originally, there is no need to build a program, but need to almost
        # rewrite a series of methods for append_backward for program_desc.
        # Therefore, in order to reuse the method of backward.py, build the program here.
        program = _build_program_by_desc(program_desc_copy)
        # 3. Add the outputs which is only used for training and not saved in
        # inference program.
        for block_idx in range(program.num_blocks):
            block = program.block(block_idx)
            for op in block.ops:
                if op.type == "batch_norm":
                    if (
                        "ReserveSpace" not in op.output_names
                        or len(op.output("ReserveSpace")) == 0
                    ):
                        reserve_space = block.create_var(
                            name=unique_name.generate_with_ignorable_key(
                                ".".join(["reserve_space", 'tmp'])
                            ),
                            dtype=block.var(op.input("X")[0]).dtype,
                            type=core.VarDesc.VarType.LOD_TENSOR,
                            persistable=False,
                            stop_gradient=True,
                        )
                        op.desc.set_output("ReserveSpace", [reserve_space.name])
                    continue

                # There are some situations that users will add backward op in Forward
                # function of Layer. And because backward op doesn't have proto. So, we
                # should skip it when we meet it.
                if not OpProtoHolder.instance().has_op_proto(op.type):
                    continue
                proto = OpProtoHolder.instance().get_op_proto(op.type)
                has_create_intermediate_out = False
                for output_proto in proto.outputs:
                    if output_proto.intermediate:
                        intermediate_name = output_proto.name
                        if intermediate_name not in op.output_names:
                            has_create_intermediate_out = True
                            intermediate_var = block.create_var(
                                name=unique_name.generate_with_ignorable_key(
                                    ".".join(
                                        [
                                            op.type + '_' + intermediate_name,
                                            'tmp',
                                        ]
                                    )
                                ),
                                type=core.VarDesc.VarType.LOD_TENSOR,
                                persistable=False,
                                stop_gradient=True,
                            )
                            op.desc.set_output(
                                intermediate_name, [intermediate_var.name]
                            )
                if has_create_intermediate_out:
                    op.desc.infer_var_type(block.desc)
                    op.desc.infer_shape(block.desc)

        return program

    @switch_to_static_graph
    def _append_backward_desc(self, infer_program_desc):
        program = self._get_train_forward_program(infer_program_desc)

        targets = []
        for out in self._output_descs:
            targets.append(program.global_block().var(out.name()))

        # 3. append backward
        backward.gradients(targets=targets, inputs=[])
        return program.desc


# [ TranslatedLayer : Run program in imperative mode ]
#
# DESIGN IDEA: using an special operator `RunProgram`, execute program inside operator.
#
# Op's Inputs:
#   - the input variable of the user feed
#   - the necessary parameters of the network
# Op's Outputs:
#   - the output variable of fetch
#
# This op receives a complete program desc, internally creates scope
# and executor, executes this program. Key points:
#
# 1. Data Sharing:
#   The varBase of the dynamic graph is not in the scope, so before the op
#   executes the program internally, create persistent variables with the
#   same name as feed, parameters, and fetch in the scope, and share the
#   LoDTensor of the op input.
#
# 2. Forward and Backward Separation:
#   Because the dynamic graph op performs the forward and backward separately,
#   in the forward op RunProgram, we only execute the forward part of whole program,
#   and in the backward op RunProgramGrad, we execute the backward part of program.
#   We can not separate the program into forward and backward part, which will
#   make some control flow execution logic wrong.


# NOTE: [compatible] deal with model saved by save_inference_model,
# which need get var info from program desc
def _load_persistable_vars_by_program(
    model_path, program_holder, params_filename=None
):
    # make sure the path has been checked
    persistable_vars = _get_persistable_vars(program_holder.infer_program)
    load_var_dict = {}
    for each_var in persistable_vars:
        orig_each_name = program_holder._suffix_varname_dict[each_var.name()]
        if _is_parameter(each_var, program_holder.infer_program):
            # create output varbase
            new_var = framework.EagerParamBase(
                shape=each_var.shape(),
                dtype=each_var.dtype(),
                name=each_var.name(),
                type=each_var.type(),
                persistable=True,
            )
        else:
            new_var = framework._varbase_creator(
                type=each_var.type(),
                name=each_var.name(),
                shape=each_var.shape(),
                dtype=each_var.dtype(),
                persistable=True,
            )
        if params_filename is None:
            framework._dygraph_tracer().trace_op(
                type='load',
                inputs={},
                outputs={'Out': new_var},
                attrs={'file_path': os.path.join(model_path, orig_each_name)},
            )
        new_var.stop_gradient = False
        load_var_dict[each_var.name()] = new_var

    if params_filename is not None:
        load_var_list = []
        dict_name_old_new = {
            v: k for k, v in program_holder._suffix_varname_dict.items()
        }
        for name in sorted(dict_name_old_new.keys()):
            load_var_list.append(load_var_dict[dict_name_old_new[name]])

        framework._dygraph_tracer().trace_op(
            type='load_combine',
            inputs={},
            outputs={'Out': load_var_list},
            attrs={'file_path': os.path.join(model_path, params_filename)},
        )

        for each_var in persistable_vars:
            if not _is_parameter(each_var, program_holder.infer_program):
                continue
            param = load_var_dict[each_var.name()]
            param.stop_gradient = False

    # NOTE: [Recovery stop gradient information based on the program]
    # After loading the model, the stop_gradient information
    # of the original variable is lost, but if a parameter does not
    # have a corresponding @GRAD variable in the backward program,
    # it can be said that it is also stop_gradient
    all_var_names = _get_all_var_names(program_holder.train_program)
    for var_name in load_var_dict:
        grad_var_name = var_name + core.grad_var_suffix()
        if grad_var_name not in all_var_names:
            load_var_dict[var_name].stop_gradient = True

    return load_var_dict


def _load_persistable_vars(
    model_path, var_info_path, program_holder, params_filename
):
    # 1. load extra var info
    with open(var_info_path, 'rb') as f:
        extra_var_info = pickle.load(f)

    # 2. construct var dict
    load_var_dict = {}
    load_var_list = []
    inv_suffix_varname_dict = {
        value: key for key, value in program_holder._suffix_varname_dict.items()
    }

    # NOTE(chenweihang): we need load persistable vars based the program,
    # because the program may be pruned when `save_inference_model`, some
    # var in `extra_var_info` may have been pruned
    for name in sorted(inv_suffix_varname_dict):
        if name not in extra_var_info:
            raise RuntimeError(
                "The model to be loaded is not complete."
                "The variable `%s` of program cannot be found in loaded model.",
                name,
            )
        # get suffix var name, see [why need to append suffix to persistable vars]
        new_name = inv_suffix_varname_dict[name]
        # create output varbase
        if extra_var_info[name].get('trainable', None) is not None:
            # use default shape and dtype
            new_var = framework.EagerParamBase(
                shape=[1],  # only to pass check, this shape is not meaningful
                dtype=core.VarDesc.VarType.FP32,
                name=new_name,
                persistable=True,
            )
        else:
            new_var = framework._varbase_creator(
                name=new_name, persistable=True
            )

        new_var.stop_gradient = extra_var_info[name]['stop_gradient']
        load_var_dict[new_name] = new_var
        load_var_list.append(new_var)

    # 3. load all vars
    assert params_filename is not None, "params_filename should not be None."
    var_file_path = os.path.join(model_path, params_filename)
    if not os.path.exists(var_file_path):
        if len(extra_var_info) != 0:
            raise ValueError("The model to be loaded is incomplete.")
    else:
        framework._dygraph_tracer().trace_op(
            type='load_combine',
            inputs={},
            outputs={'Out': load_var_list},
            attrs={'file_path': var_file_path},
        )

    return load_var_dict


# NOTE(chenweihang): to adapt paddle.load to get state_dict
def _remove_varname_suffix(var_dict, program_holder):
    no_suffix_var_dict = {}
    for var_name in var_dict:
        no_suffix_name = program_holder._suffix_varname_dict[var_name]
        no_suffix_var_dict[no_suffix_name] = var_dict[var_name]
    return no_suffix_var_dict


def _construct_program_holders(model_path, model_filename=None):
    # make sure the path has been checked
    program_holder_dict = {}

    if model_filename is not None:
        # [compatible] if assign model_filename, only can load one program as Layer.forward
        model_filename = os.path.basename(model_filename)
        model_file_path = os.path.join(model_path, model_filename)
        model_name = model_filename[: -len(INFER_MODEL_SUFFIX)]
        # Load every file that meets the requirements in the directory model_path.
        for filename in os.listdir(model_path):
            if model_filename == filename:
                func_name = 'forward'
                model_file_path = os.path.join(model_path, model_filename)
            elif filename.endswith(INFER_MODEL_SUFFIX) and filename.startswith(
                model_name
            ):
                parsing_names = filename[
                    len(model_name) : -len(INFER_MODEL_SUFFIX) + 1
                ].split('.')
                if len(parsing_names) == 3 and len(parsing_names[1]) > 0:
                    func_name = parsing_names[1]
                    model_file_path = os.path.join(model_path, filename)
                else:
                    continue
            else:
                continue
            program_holder_dict[func_name] = _ProgramHolder(
                _load_program_desc(model_file_path)
            )
    else:
        for _, _, file_names in os.walk(model_path):
            for name in file_names:
                if 'model' in name:
                    model_file_path = os.path.join(model_path, name)
                    method_name = name.strip('_')
                    if method_name == 'model':
                        method_name = 'forward'
                    else:
                        method_name.replace('model', '')
                    program_holder_dict[method_name] = _ProgramHolder(
                        _load_program_desc(model_file_path)
                    )

    return program_holder_dict


def _construct_params_and_buffers(
    model_path, programs, params_filename=None, append_suffix=True
):
    var_info_filename = str(params_filename) + ".info"
    var_info_path = os.path.join(model_path, var_info_filename)
    params_path = os.path.join(model_path, str(params_filename))

    if os.path.exists(var_info_path):
        var_dict = _load_persistable_vars(
            model_path, var_info_path, programs['forward'], params_filename
        )
        model_name = params_filename[: -len(INFER_PARAMS_SUFFIX)]
        # Load every file that meets the requirements in the directory model_path.
        for file_name in os.listdir(model_path):
            if file_name.startswith(model_name) and file_name.endswith(
                INFER_PARAMS_SUFFIX
            ):
                parsing_names = file_name[
                    len(model_name) : -len(INFER_PARAMS_SUFFIX) + 1
                ].split('.')
                if len(parsing_names) == 3 and len(parsing_names[1]) > 0:
                    func_name = parsing_names[1]
                else:
                    continue
            else:
                continue
            var_info_path = os.path.join(model_path, var_info_filename)
            var_dict.update(
                _load_persistable_vars(
                    model_path, var_info_path, programs[func_name], file_name
                )
            )
    elif params_filename is not None and not os.path.exists(params_path):
        # When saving XX, there is only '*.pdmodel'
        return {}
    else:
        var_dict = _load_persistable_vars_by_program(
            model_path, programs['forward'], params_filename
        )

    if not append_suffix:
        var_dict = _remove_varname_suffix(var_dict, programs['forward'])

    return var_dict


def _valid_vars(vars):
    return vars if vars else None


def _run_dygraph(instance, input, program_holder):

    # 1. prepare inputs, outputs, attrs
    input_vars = []
    for i, value in enumerate(input):
        if not isinstance(value, (np.ndarray, core.eager.Tensor)):
            raise TypeError(
                "The type of input in TranslatedLayer must be numpy array or Variable(Tensor), but received %s."
                % type(value)
            )
        # NOTE: In order to unify the API, firstly convert the input to Tensor
        if isinstance(value, np.ndarray):
            var = core.eager.Tensor(
                value=value,
                name=program_holder.input_descs[i].name(),
                persistable=False,
                place=framework._current_expected_place(),
                zero_copy=True,
            )
        else:
            var = value
            # NOTE: we changed var name here,
            # but it may be an important name set by user
            var.name = program_holder.input_descs[i].name()
        input_vars.append(var)
    if instance._input_args_names is None:
        instance._input_args_names = [
            ins.name() for ins in program_holder.input_descs
        ]

    persistable_vars = []
    for var_name in program_holder.persistable_names:
        dy_var_name = instance._persistable_var_name_dict[var_name]
        if dy_var_name in instance._parameters:
            persistable_vars.append(instance._parameters[dy_var_name])
        elif dy_var_name in instance._buffers:
            persistable_vars.append(instance._buffers[dy_var_name])
        else:
            raise ValueError(
                "The persistable variable %s does not exist in current TranslatedLayer."
                % var_name
            )

    output_vars = []
    for var_desc in program_holder.output_descs:
        var = core.eager.Tensor(
            dtype=var_desc.dtype(),
            dims=var_desc.shape(),
            name=var_desc.name(),
            type=var_desc.type(),
            persistable=False,
        )
        output_vars.append(var)

    # hold forward variables
    tmp_scope_vec = [program_holder.scope]

    double_grad_vars = []
    for var_desc in program_holder.double_grad_descs:
        var = core.eager.Tensor(
            dtype=var_desc.dtype(),
            dims=var_desc.shape(),
            name=var_desc.name(),
            type=var_desc.type(),
            persistable=False,
        )
        double_grad_vars.append(var)

    # 2. run program by op
    trace_program = (
        program_holder.infer_program
        if instance._is_test
        else program_holder.train_program
    )
    forward_program = (
        program_holder._infer_program_desc
        if instance._is_test
        else program_holder.forward_program
    )
    end_op_index = program_holder.infer_program.block(0).op_size()

    attrs = [
        'global_block',
        trace_program.block(0),
        'start_op_index',
        0,
        'end_op_index',
        end_op_index,
        'is_test',
        instance._is_test,
        'program_id',
        paddle.utils._hash_with_id(trace_program, instance),
    ]
    if not instance._is_test:
        attrs.extend(
            (
                'param_grad_names',
                _param_grad_names(trace_program, persistable_vars),
                'out_grad_names',
                _out_grad_names(trace_program, end_op_index, len(output_vars)),
            )
        )

    use_interpretorcore = (
        _is_enable_standalone_executor()
        and _is_dy2st_enable_standalone_executor()
    )
    attrs.extend(('use_interpretorcore', use_interpretorcore))
    if use_interpretorcore:
        attrs.extend(
            (
                'forward_global_block',
                forward_program.block(0),
                'backward_global_block',
                program_holder.backward_program.block(0),
            )
        )

    _legacy_C_ops.run_program(
        _valid_vars(input_vars),
        _valid_vars(persistable_vars),
        _valid_vars(output_vars),
        tmp_scope_vec,
        _valid_vars(double_grad_vars),
        None,
        *attrs
    )

    # NOTE: [ why need set param's gradient type here ]
    # if user set sparse gradient mode, the param's gradient
    # will be SelectedRows, not LoDTensor. But tracer will just
    # set param grad Tensor by forward Tensor(LoDTensor)
    # If we don't change grad_var type here, RunProgramOp need
    # transform SelectedRows to LoDTensor forcibly, it may not
    # be user wanted result.
    for persistable_var in persistable_vars:
        grad_var_name = persistable_var.name + core.grad_var_suffix()
        grad_var = trace_program.block(0).find_var(grad_var_name.encode())
        # NOTE: cannot find var desc maybe not problem,
        # such as in batch_norm
        if grad_var is None:
            continue
        persistable_var._set_grad_type(grad_var.type())

    # 3. prepare output, keep same form with inputs
    outs = output_vars
    if len(output_vars) == 1:
        outs = output_vars[0]
    return outs


def _run_static_graph(input, program_holder, trace_program):
    main_program = framework.default_main_program()
    param_var_names = _get_persistable_var_names(trace_program)
    _, dict_rename_var_old_new = _rename_var_program_desc(
        trace_program, exclude=param_var_names
    )
    trace_program.flush()
    output_names = [var.name() for var in program_holder.output_descs]
    # append blocks from 'trace_program'
    _append_block(
        main_program,
        trace_program,
        program_holder,
        input,
        dict_rename_var_old_new,
    )
    main_program._sync_with_cpp()
    outs = _get_output_from_program(
        main_program, program_holder, dict_rename_var_old_new
    )
    if len(outs) == 1:
        outs = outs[0]
    return outs


def _collect_current_and_parent_var(program, block_idx):
    '''
    Get variables in current block and its parent block.

    Args:
        program(Program): The program containing the current block.
        block_idx(int): index of current block.

    Returns:
        List: list of variables.
    '''
    vars = []
    if block_idx < 0:
        return vars
    for var in program.block(block_idx).vars:
        vars.append(var)
    parent_idx = program.block(block_idx).parent_idx
    if parent_idx > -1:
        vars += _collect_current_and_parent_var(program, parent_idx)
    return vars


def _append_block(
    dest_program,
    src_program_desc,
    program_holder,
    input_variables,
    dict_rename_var_old_new=None,
):
    '''
    Append Variables and Operators in 'src_program_desc' to dest_program.

    Args:
        dest_program(Program): Variables and Operators are appended to it.
        src_program_desc(ProgramDesc): Variables in it will be appended to 'dest_program'.
        program_holder(_ProgramHolder): program_holder of TranslatedLayer
        input_variables(list): list of input variables
        dict_rename_var_old_new(None|dict): When using '_rename_var_program_desc',
        use it to map the name of the variable before it was modified and the new name.
    '''

    origin_block_idx = dest_program.current_block_idx
    param_var_names = _collect_current_and_parent_var(
        dest_program, origin_block_idx
    )
    append_var_from_block_desc_static(
        dest_program.block(origin_block_idx),
        src_program_desc.block(0),
        exclude=param_var_names,
    )

    name_inp_desc = [inp.name() for inp in program_holder.input_descs]
    input_names = [inp.name for inp in input_variables]
    if len(name_inp_desc) != len(input_names):
        raise ValueError(
            "The number of input is invalid, expected {}, but received {}.".format(
                len(name_inp_desc), len(input_names)
            )
        )
    for i, out_name in enumerate(name_inp_desc):
        if dict_rename_var_old_new:
            out_name = dict_rename_var_old_new[out_name]
        dest_program.block(origin_block_idx).append_op(
            type='assign',
            inputs={'X': [input_names[i]]},
            outputs={'Out': [out_name]},
        )

    append_ops = append_op_from_block_desc_static(
        dest_program.block(origin_block_idx), src_program_desc.block(0)
    )
    dest_program._sync_with_cpp()

    offset_block_idx = dest_program.num_blocks - 1
    parent_idx = 0
    if src_program_desc.num_blocks() > 1:
        for src_block_idx in range(1, src_program_desc.num_blocks()):
            src_block = src_program_desc.block(src_block_idx)
            src_parent_idx = src_block.parent
            if src_parent_idx > 0:
                parent_idx = offset_block_idx + parent_idx
            else:
                parent_idx = origin_block_idx
            dest_block = dest_program._create_block(parent_idx=parent_idx)
            append_var_from_block_desc_static(
                dest_block, src_block, exclude=param_var_names
            )
            append_ops += append_op_from_block_desc_static(
                dest_block, src_block
            )

    dest_program._sync_with_cpp()
    for op in append_ops:
        if op.has_attr('sub_block'):
            sub = op.attr('sub_block')
            if isinstance(sub, framework.core.BlockDesc):
                origin_id = sub.id
            if isinstance(sub, framework.Block):
                origin_id = sub.idx
            op._set_attr(
                'sub_block', dest_program.block(offset_block_idx + origin_id)
            )
    dest_program._sync_with_cpp()
    dest_program.current_block_idx = origin_block_idx


def _get_output_from_program(
    program, program_holder, dict_rename_var_old_new=None
):
    """
    Get output name of 'program' according to program_holder
    """
    outs = []
    for var in program_holder.output_descs:
        for idx in range(program.num_blocks):
            vars = program.block(idx).vars
            var_name = var.name()
            if dict_rename_var_old_new:
                var_name = dict_rename_var_old_new[var_name]
            if var_name in vars:
                out = vars[var_name]
                if out not in outs:
                    outs.append(out)
    return outs


def append_op_from_block_desc_static(block, src_block_desc):
    """
    Append Operators of 'src_block_desc' to current block.

    Args:
        block(Block): append OP of  'src_block_desc' to it.
        src_block_desc(BlockDesc): append var of  'src_block_desc'

    Returns:
        List: list of the OP that are append to current block.
    """
    ops = []
    for i in range(src_block_desc.op_size()):
        ops.append(append_op_from_desc_static(block, src_block_desc.op(i)))
    return ops


def append_op_from_desc_static(block, op_desc):
    """
    Append Operators to 'block' according to 'op_desc'.

    Args:
        block(Block): append OP of  'src_block_desc' to it.
        op_desc(OpDesc): create OP according to it.

    Returns:
        Operator: OP appended to 'block'.
    """
    op_type = op_desc.type()
    op_append = block.desc.append_op()
    op_append.copy_from(op_desc)
    op = framework.Operator(
        block=block,
        desc=op_append,
        type=op_type,
        inputs=None,
        outputs=None,
        attrs=None,
    )
    block.ops.append(op)
    return op


def append_var_from_block_desc_static(
    block, src_block_desc, include=None, exclude=None
):
    """
    Append Variables of 'src_block_desc' to current block.
    If 'include' is not `None`,variables that are not in include are not append.
    If 'exclude' is not `None`,variables that are in exclude will are not append.

    Args:
        block(Block): append Variables of  'src_block_desc' to it.
        src_block_desc(BlockDesc): append var of  'src_block_desc'
        include(List):list of names of variables
        exclude(List):list of names of variables

    Returns:
        List: list of the variables that are append to current block.
    """
    vars_append = []
    for var_desc in src_block_desc.all_vars():
        var_desc_name = var_desc.name()
        should_append = (include is None or var_desc_name in include) and (
            exclude is None or var_desc_name not in exclude
        )
        if not block.has_var(var_desc_name) and should_append:
            var_type = var_desc.type()
            if var_type in [
                core.VarDesc.VarType.SELECTED_ROWS,
                core.VarDesc.VarType.LOD_TENSOR,
                core.VarDesc.VarType.LOD_TENSOR_ARRAY,
            ]:
                data_type = var_desc.dtype()
                var_shape = var_desc.shape()
            else:
                data_type = None
                var_shape = None
            if var_type in [
                core.VarDesc.VarType.LOD_TENSOR,
                core.VarDesc.VarType.LOD_TENSOR_ARRAY,
            ]:
                lod_level = var_desc.lod_level()
            else:
                lod_level = None

            if var_desc.persistable():
                current_block = block.program.global_block()
            else:
                current_block = block

            vars_append.append(
                current_block.create_var(
                    name=var_desc.name(),
                    dtype=data_type,
                    type=var_type,
                    shape=var_shape,
                    lod_level=lod_level,
                    persistable=var_desc.persistable(),
                    set_need_check_feed=var_desc.need_check_feed(),
                )
            )
    return vars_append


class TranslatedLayer(layers.Layer):
    """
    TranslatedLayer is a ``paddle.nn.Layer`` for holding the model
    loaded by :ref:`api_paddle_jit_load` . It can be used like a
    general Layer object in eval or train mode.

    .. note:
        The TranslatedLayer objects should not be created by constructor, it only can be loaded and constructed by :ref:`api_paddle_jit_load` .

    Examples:
        .. code-block:: python

            import numpy as np
            import paddle
            import paddle.nn as nn
            import paddle.optimizer as opt

            BATCH_SIZE = 16
            BATCH_NUM = 4
            EPOCH_NUM = 4

            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

                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

                def __len__(self):
                    return self.num_samples

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

                @paddle.jit.to_static
                def forward(self, x):
                    return self._linear(x)

            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.

            # create network
            layer = LinearNet()
            loss_fn = nn.CrossEntropyLoss()
            adam = opt.Adam(learning_rate=0.001, parameters=layer.parameters())

            # 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)

            # train
            train(layer, loader, loss_fn, adam)

            # save
            model_path = "linear.example.model"
            paddle.jit.save(layer, model_path)

            # 2. load model as TranslatedLayer

            # load
            translated_layer = paddle.jit.load(model_path)

            # inference
            translated_layer.eval()
            x = paddle.randn([1, IMAGE_SIZE], 'float32')
            pred = translated_layer(x)

            # fine-tune
            translated_layer.train()
            adam = opt.Adam(learning_rate=0.001, parameters=translated_layer.parameters())
            train(translated_layer, loader, loss_fn, adam)

    """

    def __init__(self, programs, persistable_vars):
        super().__init__()

        if not isinstance(programs, dict):
            raise TypeError(
                "TranslatedLayer need to use _ProgramHolder's dict for initialization."
            )
        if not isinstance(persistable_vars, dict):
            raise TypeError(
                "TranslatedLayer need to use persistable variable dict for initialization."
            )

        self._program_holder_dict = programs

        # NOTE(chenweihang): [ why not use var name directly? ]
        # When add parameter or buffer to Layer by follow apis,
        # the variable name can't contain `.`, beccause which may cause
        # AttributeError when access the newly added parameter or buffer
        # in the form of `self.**.**``, but the EagerParamBase or BarBase
        # name contains `.` originally, such as `linear_0.w_0`, so here
        # need to generate new var name for each var
        self._persistable_var_name_dict = {}
        # the TranslatedLayer object holded var names count started from 0
        with unique_name.guard():
            for name, var in persistable_vars.items():
                if isinstance(var, framework.EagerParamBase):
                    dy_name = _generate_unique_var_name(PARAMETER_NAME_PREFIX)
                    self._persistable_var_name_dict[name] = dy_name
                    self.add_parameter(dy_name, var)
                elif isinstance(var, core.eager.Tensor):
                    dy_name = _generate_unique_var_name(BUFFER_NAME_PREFIX)
                    self._persistable_var_name_dict[name] = dy_name
                    self.register_buffer(dy_name, var)
                else:
                    raise TypeError(
                        "Adding persistent variable which  to layer is not supported now"
                    )

        self._is_test = True
        self._input_args_names = None

    @staticmethod
    @framework.dygraph_only
    def _construct(model_path, configs=None):
        # 0. dir and filename check
        model_path = os.path.normpath(model_path)
        if not os.path.isdir(model_path):
            raise ValueError("There is no directory named '%s'" % model_path)
        model_filename = None
        params_filename = None
        if configs is not None:
            model_filename = configs.model_filename
            params_filename = configs.params_filename

        # 1. load program desc & construct _ProgramHolder
        programs = _construct_program_holders(model_path, model_filename)

        # 2. load layer parameters & buffers
        persistable_vars = _construct_params_and_buffers(
            model_path, programs, params_filename
        )

        # 3. construct TranslatedLayer object
        translated_layer = TranslatedLayer(programs, persistable_vars)

        # 4. create TranslatedLayer's execution method
        for method_name, program_holder in programs.items():
            if translated_layer._input_args_names is None:
                translated_layer._input_args_names = [
                    ins.name() for ins in program_holder.input_descs
                ]
            setattr(
                TranslatedLayer,
                method_name,
                TranslatedLayer._execution_method_creator(
                    method_name, program_holder
                ),
            )

        # 5. set TranslatedLayer's default mode to eval
        translated_layer.eval()

        return translated_layer

    @staticmethod
    def _execution_method_creator(method_name, program_holder):
        def __i_m_p_l__(self, *input):
            program_holder = self._program_holder_dict[__i_m_p_l__.__name__]
            # When using jit.save, it runs in static graph mode.
            # Run in dynamic graph mode when the model is inferring.
            if _non_static_mode():
                return _run_dygraph(self, input, program_holder)
            else:
                # NOTE(weixin): [ why not use 'program_holder.infer_program' directly? ]
                # When use '_run_static_graph(input, program_holder, program_holder.infer_program)',
                # because '_run_static_graph' modifies 'ProgramDesc', 'OpDesc.op_size()' will return a very large wrong number.
                # A Segmentation fault error may occur if used 'p=ProgramDesc(program_holder.infer_program)'.
                p = framework.Program._construct_from_desc(
                    core.ProgramDesc(program_holder.infer_program)
                )
                return _run_static_graph(input, program_holder, p.desc)

        __i_m_p_l__.__name__ = method_name
        return __i_m_p_l__

    def train(self):
        self._is_test = False
        self.training = True

    def eval(self):
        self._is_test = True
        self.training = False

    def program(self, method_name='forward'):
        """
        Gets translated program of specified method.

        Args:
            - method_name (string): mehtod name corresponding to the program
                to be obtained. Default: 'forward'.

        Returns:
            Program

        Examples:
            .. code-block:: python

                import numpy as np
                import paddle
                import paddle.nn as nn
                import paddle.optimizer as opt

                BATCH_SIZE = 16
                BATCH_NUM = 4
                EPOCH_NUM = 4

                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

                    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

                    def __len__(self):
                        return self.num_samples

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

                    @paddle.jit.to_static
                    def forward(self, x):
                        return self._linear(x)

                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())))

                # create network
                layer = LinearNet()
                loss_fn = nn.CrossEntropyLoss()
                adam = opt.Adam(learning_rate=0.001, parameters=layer.parameters())

                # 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)

                # train
                train(layer, loader, loss_fn, adam)

                # save
                model_path = "linear.example.model"
                paddle.jit.save(layer, model_path)

                # load
                translated_layer = paddle.jit.load(model_path)

                # get program
                program = translated_layer.program()
        """
        # 1. get program holder
        program_holder = self._get_program_holder(method_name)

        # 2. get inference program desc
        program_desc = program_holder.infer_program

        # 3. construct program
        program = _build_program_by_desc(program_desc)
        return program

    def _get_program_holder(self, method_name='forward'):
        program_holder = self._program_holder_dict.get(method_name, None)
        if program_holder is None:
            raise ValueError(
                "The method `%s` does not exist in loaded TranslatedLayer."
                % method_name
            )
        return program_holder

    def _input_spec(self, method_name='forward'):
        # 1. get program holder
        program_holder = self._get_program_holder(method_name)

        # 2. build input spec by input desc
        input_spec = []
        for var_desc in program_holder.input_descs:
            spec = paddle.static.InputSpec(
                shape=var_desc.shape(),
                dtype=var_desc.dtype(),
                name=var_desc.name(),
            )
            input_spec.append(spec)

        return input_spec

    def _output_spec(self, method_name='forward'):
        # 1. get program holder
        program_holder = self._get_program_holder(method_name)

        # 2. build output spec by output desc
        output_spec = []
        for var_desc in program_holder.output_descs:
            # NOTE(chenweihang): InputSpec describes a tensor, not just input.
            # Maybe the name is not good enough. Here we use InputSpec to
            # construct the description of Output tensor
            spec = paddle.static.InputSpec(
                shape=var_desc.shape(),
                dtype=var_desc.dtype(),
                name=var_desc.name(),
            )
            output_spec.append(spec)

        return output_spec