2019-10-23 19:25:59 +08:00
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
2022-01-11 14:02:45 +08:00
# Copyright (c) 2021 NVIDIA Corporation. All rights reserved.
2019-10-23 19:25:59 +08:00
#
# 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.
2023-01-04 11:55:21 +08:00
from __future__ import annotations
2023-10-22 20:37:05 +08:00
import inspect
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
import os
import pickle
2023-10-20 20:05:25 +08:00
import sys
2022-01-11 14:02:45 +08:00
import threading
2023-08-07 15:10:33 +08:00
import types
2023-10-22 20:37:05 +08:00
import warnings
from collections import OrderedDict
2024-06-14 14:32:24 +08:00
from collections . abc import Sequence
2023-10-23 22:55:23 +08:00
from contextlib import contextmanager
2024-06-05 17:15:45 +08:00
from types import ModuleType
2024-06-21 00:40:47 +08:00
from typing import (
2024-06-29 14:02:32 +08:00
TYPE_CHECKING ,
2024-06-21 00:40:47 +08:00
Any ,
Callable ,
Protocol ,
TypedDict ,
TypeVar ,
overload ,
)
2024-06-14 14:32:24 +08:00
from typing_extensions import (
Literal ,
NotRequired ,
ParamSpec ,
TypeAlias ,
Unpack ,
2024-06-05 17:15:45 +08:00
)
2020-08-27 18:20:05 +08:00
import paddle
2023-09-07 17:26:19 +08:00
from paddle . base import core , dygraph
from paddle . base . compiler import (
2022-10-23 20:01:27 +08:00
BuildStrategy ,
)
2023-09-07 17:26:19 +08:00
from paddle . base . dygraph . base import (
2022-10-23 20:01:27 +08:00
switch_to_static_graph ,
)
2023-09-07 17:26:19 +08:00
from paddle . base . executor import Executor , scope_guard
from paddle . base . framework import (
2023-10-22 20:37:05 +08:00
EagerParamBase ,
Parameter ,
2022-10-23 20:01:27 +08:00
Variable ,
_current_expected_place ,
2023-10-22 20:37:05 +08:00
dygraph_only ,
2022-10-23 20:01:27 +08:00
)
2023-09-07 17:26:19 +08:00
from paddle . base . wrapped_decorator import wrap_decorator
2024-02-29 21:51:42 +08:00
from paddle . framework import use_pir_api
2023-10-22 20:37:05 +08:00
from paddle . nn import Layer
from paddle . static . io import save_inference_model
2023-10-23 22:55:23 +08:00
from paddle . utils . environments import (
EnvironmentVariableGuard ,
)
2023-10-22 20:37:05 +08:00
from . dy2static import logging_utils
2025-05-20 10:23:15 +08:00
from . dy2static . convert_call_func import add_ignore_module
2023-10-22 20:37:05 +08:00
from . dy2static . program_translator import (
ASTStaticFunction ,
ProgramTranslator ,
StaticFunction ,
SymbolicStaticFunction ,
unwrap_decorators ,
)
2025-03-25 10:36:13 +08:00
from . dy2static . utils import (
ENV_ENABLE_SOT ,
Backend ,
infer_use_cinn_backend ,
)
2024-04-25 13:58:01 +08:00
from . pir_translated_layer import PIR_INFER_MODEL_SUFFIX , PirTranslatedLayer
2023-10-22 20:37:05 +08:00
from . translated_layer import (
INFER_MODEL_SUFFIX ,
INFER_PARAMS_INFO_SUFFIX ,
INFER_PARAMS_SUFFIX ,
INFER_PROPERTY_SUFFIX ,
TranslatedLayer ,
)
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2024-06-29 14:02:32 +08:00
if TYPE_CHECKING :
2024-08-01 18:33:37 +08:00
from paddle import Tensor
2024-06-29 14:02:32 +08:00
from paddle . _typing import NestedStructure
from paddle . static import InputSpec
2024-07-28 21:08:52 +08:00
class _SaveOptions ( TypedDict ) :
2024-08-01 18:33:37 +08:00
output_spec : NotRequired [ Sequence [ Tensor | int ] ]
2024-07-28 21:08:52 +08:00
with_hook : NotRequired [ bool ]
combine_params : NotRequired [ bool ]
clip_extra : NotRequired [ bool ]
skip_forward : NotRequired [ bool ]
input_names_after_prune : NotRequired [ list [ str ] ]
skip_prune_program : NotRequired [ bool ]
2024-11-18 14:51:08 +08:00
separate_parameters : NotRequired [ bool ]
2024-07-28 21:08:52 +08:00
class _LoadOptions ( TypedDict ) :
model_filename : NotRequired [ str ]
params_filename : NotRequired [ str ]
2024-06-05 17:15:45 +08:00
_LayerT = TypeVar ( " _LayerT " , bound = Layer )
_RetT = TypeVar ( " _RetT " )
_InputT = ParamSpec ( " _InputT " )
Backends : TypeAlias = Literal [ " CINN " ]
2023-10-23 22:55:23 +08:00
@contextmanager
def sot_mode_guard ( value : bool ) :
with EnvironmentVariableGuard ( ENV_ENABLE_SOT , value ) :
yield
2019-10-23 19:25:59 +08:00
2020-08-27 18:20:05 +08:00
def copy_decorator_attrs ( original_func , decorated_obj ) :
"""
Copies some necessary attributes from original function into decorated function.
Args:
original_func(callable): the original decorated function.
2020-09-24 12:02:16 +08:00
decorated_obj(StaticFunction): the target decorated StaticFunction object.
2020-08-27 18:20:05 +08:00
"""
2022-12-19 19:12:43 +08:00
decorator_name = " to_static "
2020-08-27 18:20:05 +08:00
decorated_obj . __name__ = original_func . __name__
decorated_obj . _decorator_name = decorator_name
2024-10-23 23:47:18 +08:00
if not inspect . ismethod ( original_func ) :
decorated_obj . __wrapped__ = original_func
2020-08-27 18:20:05 +08:00
decorated_obj . __doc__ = original_func . __doc__
if hasattr ( original_func , " __module__ " ) :
decorated_obj . __module__ = original_func . __module__
return decorated_obj
2024-06-05 17:15:45 +08:00
def ignore_module ( modules : list [ ModuleType ] ) - > None :
2023-01-10 14:45:39 +08:00
"""
Adds modules that ignore transcription.
Builtin modules that have been ignored are collections, pdb, copy, inspect, re, numpy, logging, six
Args:
2024-06-05 17:15:45 +08:00
modules (list[ModuleType]): Ignored modules that you want to add
2023-01-10 14:45:39 +08:00
Examples:
2026-02-20 17:21:20 +08:00
.. code-block:: pycon
2023-01-10 14:45:39 +08:00
2023-07-31 14:44:48 +08:00
>>> import scipy
2025-07-08 11:33:54 +08:00
>>> import networkx
2023-01-10 14:45:39 +08:00
2023-07-31 14:44:48 +08:00
>>> import paddle
>>> from paddle.jit import ignore_module
>>> modules = [
... scipy,
2025-07-08 11:33:54 +08:00
... networkx,
2023-07-31 14:44:48 +08:00
... ]
>>> ignore_module(modules)
2023-01-10 14:45:39 +08:00
"""
add_ignore_module ( modules )
2024-06-21 00:40:47 +08:00
class _ToStaticOptions ( TypedDict ) :
2024-06-05 17:15:45 +08:00
property : NotRequired [ bool ]
full_graph : NotRequired [ bool ]
2024-06-21 00:40:47 +08:00
class _ToStaticDecorator ( Protocol ) :
2024-06-05 17:15:45 +08:00
@overload
2024-08-08 10:10:45 +08:00
def __call__ ( self , function : _LayerT ) - > _LayerT : . . .
2024-06-05 17:15:45 +08:00
@overload
def __call__ (
self , function : Callable [ _InputT , _RetT ]
2024-08-08 10:10:45 +08:00
) - > StaticFunction [ _InputT , _RetT ] : . . .
2024-06-05 17:15:45 +08:00
@overload
def to_static (
function : _LayerT ,
2024-06-21 00:40:47 +08:00
input_spec : NestedStructure [ InputSpec ] | None = . . . ,
2024-06-05 17:15:45 +08:00
build_strategy : BuildStrategy | None = . . . ,
backend : Backends | None = . . . ,
2024-06-21 00:40:47 +08:00
* * kwargs : Unpack [ _ToStaticOptions ] ,
2024-08-08 10:10:45 +08:00
) - > _LayerT : . . .
2024-06-05 17:15:45 +08:00
@overload
def to_static (
function : Callable [ _InputT , _RetT ] ,
2024-06-21 00:40:47 +08:00
input_spec : NestedStructure [ InputSpec ] | None = . . . ,
2024-06-05 17:15:45 +08:00
build_strategy : BuildStrategy | None = . . . ,
backend : Backends | None = . . . ,
2024-06-21 00:40:47 +08:00
* * kwargs : Unpack [ _ToStaticOptions ] ,
2024-08-08 10:10:45 +08:00
) - > StaticFunction [ _InputT , _RetT ] : . . .
2024-06-05 17:15:45 +08:00
@overload
def to_static (
function : None = . . . ,
2024-06-21 00:40:47 +08:00
input_spec : NestedStructure [ InputSpec ] | None = . . . ,
2024-06-05 17:15:45 +08:00
build_strategy : BuildStrategy | None = . . . ,
backend : Backends | None = . . . ,
2024-06-21 00:40:47 +08:00
* * kwargs : Unpack [ _ToStaticOptions ] ,
2024-08-08 10:10:45 +08:00
) - > _ToStaticDecorator : . . .
2024-06-05 17:15:45 +08:00
2022-12-19 19:12:43 +08:00
def to_static (
2023-04-11 11:16:53 +08:00
function = None ,
input_spec = None ,
build_strategy = None ,
2025-04-10 10:54:32 +08:00
backend = " CINN " ,
2023-04-11 11:16:53 +08:00
* * kwargs ,
2022-10-23 20:01:27 +08:00
) :
2020-04-18 10:13:02 +08:00
"""
2023-10-21 13:29:12 +08:00
Converts dynamic graph APIs into static graph function APIs. Decorator
2022-12-30 11:02:06 +08:00
@to_static handles the Program and Executor of static graph mode and returns
2023-10-21 13:29:12 +08:00
the result as dynamic graph Tensor(s). Users could use the returned dynamic
graph Tensor(s) to do dynamic graph training, inference, or other operations.
If the decorated function calls other dynamic graph function, the called one
will be converted into static graph function as well.
2020-04-18 10:13:02 +08:00
Args:
2023-10-21 13:29:12 +08:00
function (callable): Callable dynamic graph function. If it used as a
decorator, the decorated function will be parsed as this parameter.
input_spec (list[InputSpec]|tuple[InputSpec]): list/tuple of InputSpec to
specific the shape/dtype/name information of each input Tensor.
build_strategy (BuildStrategy|None): This argument is used to compile the
2021-07-28 20:35:37 +08:00
converted program with the specified options, such as operators ' fusion
in the computational graph and memory optimization during the execution
2025-04-10 10:54:32 +08:00
of the computational graph. For more information about :attr:`build_strategy`,
please refer to :ref:`paddle.static.BuildStrategy <cn_api_paddle_static_BuildStrategy>`.
The default is ``None``.
backend(str, Optional): Specifies compilation backend, which can be `` " CINN " `` or
``None``. When backend is `` " CINN " ``, CINN compiler will be used to speed up
training and inference. default value is `` " CINN " ``.
kwargs: Support keys including :attr:`property` and :attr:`full_graph`.
- property (bool): If True, the function will be treated as a property
function. The default is False.
- full_graph (bool): If True, the function will be converted into a
full static graph. The default is False.
2020-02-28 16:35:24 +08:00
2020-04-18 10:13:02 +08:00
Returns:
2020-08-25 11:25:20 +08:00
Tensor(s): containing the numerical result.
2020-03-04 15:03:59 +08:00
2020-04-18 10:13:02 +08:00
Examples:
2026-02-20 17:21:20 +08:00
.. code-block:: pycon
2020-03-04 15:03:59 +08:00
2023-09-26 11:19:57 +08:00
>>> # doctest: +SKIP( ' `paddle.jit.to_static` can not run in xdoctest ' )
2023-07-31 14:44:48 +08:00
>>> import paddle
>>> from paddle.jit import to_static
>>> @to_static
>>> def func(x):
... if paddle.mean(x) < 0:
... x_v = x - 1
... else:
... x_v = x + 1
... return x_v
>>> x = paddle.ones([1, 2], dtype= ' float32 ' )
>>> x_v = func(x)
>>> print(x_v)
Tensor(shape=[1, 2], dtype=float32, place=Place(cpu), stop_gradient=True,
[[2., 2.]])
2020-02-10 11:55:14 +08:00
2020-04-18 10:13:02 +08:00
"""
2023-04-11 11:16:53 +08:00
property = kwargs . get ( " property " , False )
2023-10-20 20:05:25 +08:00
full_graph = kwargs . get ( " full_graph " , None )
2025-03-25 10:36:13 +08:00
build_strategy = build_strategy or BuildStrategy ( )
if not isinstance ( build_strategy , BuildStrategy ) :
raise TypeError (
f " Required type(build_strategy) shall be `paddle.static.BuildStrategy`, but received { type ( build_strategy ) . __name__ } "
)
backend = Backend . from_arg ( backend )
2025-05-13 13:41:42 +08:00
if infer_use_cinn_backend ( backend , build_strategy ) :
backend = Backend . CINN
elif backend . is_pcc ( ) :
pass
else :
backend = Backend . PHI
2020-03-25 17:26:09 +08:00
2020-08-27 18:20:05 +08:00
def decorated ( python_func ) :
"""
2025-02-20 09:08:37 +08:00
Decorates a python function into a ASTStaticFunction or SymbolicStaticFunction object.
2020-08-27 18:20:05 +08:00
"""
2023-08-07 15:10:33 +08:00
2023-10-20 20:05:25 +08:00
nonlocal full_graph
if full_graph is None :
2023-10-23 22:55:23 +08:00
flag = ENV_ENABLE_SOT . get ( )
full_graph = not flag
2023-10-20 20:05:25 +08:00
2025-10-30 23:10:11 +08:00
if sys . version_info > = ( 3 , 15 ) and not full_graph :
2023-10-20 20:05:25 +08:00
warnings . warn (
2025-10-30 23:10:11 +08:00
" full_graph=False is not supported in Python 3.15+. Set full_graph=True automatically "
2023-10-20 20:05:25 +08:00
)
full_graph = True
2023-08-07 15:10:33 +08:00
2025-02-20 09:08:37 +08:00
StaticClass = (
ASTStaticFunction if full_graph else SymbolicStaticFunction
)
2023-08-07 15:10:33 +08:00
2020-08-27 18:20:05 +08:00
# Step 1. unwrap the function if it is already decorated.
_ , python_func = unwrap_decorators ( python_func )
2020-02-27 17:25:13 +08:00
2020-08-27 18:20:05 +08:00
# Step 2. copy some attributes from original python function.
2022-10-23 20:01:27 +08:00
static_layer = copy_decorator_attrs (
original_func = python_func ,
2023-08-07 15:10:33 +08:00
decorated_obj = StaticClass (
2022-10-23 20:01:27 +08:00
function = python_func ,
input_spec = input_spec ,
build_strategy = build_strategy ,
property = property ,
2023-04-11 11:16:53 +08:00
backend = backend ,
2022-10-23 20:01:27 +08:00
) ,
)
2020-08-27 18:20:05 +08:00
return static_layer
2020-02-27 17:25:13 +08:00
2022-12-19 19:12:43 +08:00
# for usage: `to_static(foo, ...)`
2020-08-27 18:20:05 +08:00
if function is not None :
2020-09-11 11:11:09 +08:00
if isinstance ( function , Layer ) :
2020-09-24 12:02:16 +08:00
if isinstance ( function . forward , StaticFunction ) :
2020-09-11 11:11:09 +08:00
class_name = function . __class__ . __name__
2020-09-17 23:01:26 +08:00
logging_utils . warn (
2024-04-01 10:20:33 +08:00
f " ` { class_name } .forward` has already been decorated somewhere. It will be redecorated to replace previous one. "
2022-10-23 20:01:27 +08:00
)
2024-12-04 21:55:13 +08:00
function . _original_funcs [ " forward " ] = function . forward
2020-09-11 11:11:09 +08:00
function . forward = decorated ( function . forward )
return function
else :
return decorated ( function )
2020-02-27 17:25:13 +08:00
2022-12-19 19:12:43 +08:00
# for usage: `@to_static`
2020-08-27 18:20:05 +08:00
return decorated
2020-02-10 11:55:14 +08:00
2022-11-08 11:29:41 +08:00
class _SaveLoadConfig :
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
def __init__ ( self ) :
self . _output_spec = None
self . _model_filename = None
self . _params_filename = None
self . _separate_params = False
2020-09-03 16:06:45 +08:00
# used for `paddle.load`
self . _keep_name_table = False
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
# NOTE: Users rarely use following configs, so these configs are not open to users,
# reducing user learning costs, but we retain the configuration capabilities
2021-01-20 09:56:38 +08:00
# If True, programs are modified to only support direct inference deployment.
# Otherwise,more information will be stored for flexible optimization and re-training.
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
# 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
2022-04-29 09:44:01 +08:00
self . with_hook = False
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2022-06-30 19:17:49 +08:00
# if True, multi `StaticFunction` will share params in one file.
self . combine_params = False
2023-08-21 17:24:02 +08:00
# when need to save a prune model, use input_names_after_prune to specify the inputs left after pruning
self . input_names_after_prune = None
2024-02-04 14:51:49 +08:00
# in the scene of llm-inference, pruning program can cause unexpectable result, an option to skip prune is necessary
2023-08-31 02:57:09 -05:00
self . skip_prune_program = False
2024-11-18 14:51:08 +08:00
# if True, the params will be saved separately in multiple files.
self . separate_parameters = False
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
@property
def output_spec ( self ) :
return self . _output_spec
@output_spec.setter
def output_spec ( self , spec ) :
2020-10-11 23:12:19 +08:00
if spec is None :
return
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
if not isinstance ( spec , list ) :
raise TypeError (
2024-06-03 19:07:26 +08:00
f " The config `output_spec` should be ' list ' , but received input type is { type ( input ) } . "
2022-10-23 20:01:27 +08:00
)
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
for var in spec :
2023-03-30 10:11:14 +08:00
if not isinstance ( var , core . eager . Tensor ) :
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
raise TypeError (
2024-06-03 19:07:26 +08:00
f " The element in config `output_spec` list should be ' Variable ' , but received element ' s type is { type ( var ) } . "
2022-10-23 20:01:27 +08:00
)
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
self . _output_spec = spec
@property
def model_filename ( self ) :
return self . _model_filename
@model_filename.setter
def model_filename ( self , filename ) :
2020-10-11 23:12:19 +08:00
if filename is None :
return
2022-10-19 15:54:41 +08:00
if not isinstance ( filename , str ) :
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
raise TypeError (
2024-06-03 19:07:26 +08:00
f " The config `model_filename` should be str, but received input ' s type is { type ( filename ) } . "
2022-10-23 20:01:27 +08:00
)
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
if len ( filename ) == 0 :
2020-10-11 23:12:19 +08:00
raise ValueError ( " The config `model_filename` is empty string. " )
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
self . _model_filename = filename
@property
def params_filename ( self ) :
return self . _params_filename
@params_filename.setter
def params_filename ( self , filename ) :
2020-10-11 23:12:19 +08:00
if filename is None :
return
2022-10-19 15:54:41 +08:00
if not isinstance ( filename , str ) :
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
raise TypeError (
2024-06-03 19:07:26 +08:00
f " The config `params_filename` should be str, but received input ' s type is { type ( filename ) } . "
2022-10-23 20:01:27 +08:00
)
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
if len ( filename ) == 0 :
2020-10-11 23:12:19 +08:00
raise ValueError ( " The config `params_filename` is empty string. " )
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
self . _params_filename = filename
2020-09-03 16:06:45 +08:00
@property
def keep_name_table ( self ) :
return self . _keep_name_table
@keep_name_table.setter
def keep_name_table ( self , value ) :
2020-10-11 23:12:19 +08:00
if value is None :
return
2020-09-03 16:06:45 +08:00
if not isinstance ( value , bool ) :
raise TypeError (
2024-06-03 19:07:26 +08:00
f " The config `keep_name_table` should be bool value, but received input ' s type is { type ( value ) } . "
2022-10-23 20:01:27 +08:00
)
2020-09-03 16:06:45 +08:00
self . _keep_name_table = value
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2024-07-28 21:08:52 +08:00
def _parse_save_configs ( configs : _SaveOptions ) - > _SaveLoadConfig :
2022-07-01 19:35:28 +08:00
supported_configs = [
2023-08-21 17:24:02 +08:00
" output_spec " ,
2022-10-23 20:01:27 +08:00
" with_hook " ,
" combine_params " ,
" clip_extra " ,
" skip_forward " ,
2023-08-21 17:24:02 +08:00
" input_names_after_prune " ,
2023-08-31 02:57:09 -05:00
" skip_prune_program " ,
2024-11-18 14:51:08 +08:00
" separate_parameters " ,
2022-07-01 19:35:28 +08:00
]
2020-10-11 23:12:19 +08:00
# input check
for key in configs :
if key not in supported_configs :
raise ValueError (
2024-06-03 19:07:26 +08:00
f " The additional config ( { key } ) of `paddle.jit.save` is not supported. "
2022-10-23 20:01:27 +08:00
)
2020-10-11 23:12:19 +08:00
# construct inner config
inner_config = _SaveLoadConfig ( )
2023-08-21 17:24:02 +08:00
inner_config . output_spec = configs . get ( " output_spec " , None )
inner_config . with_hook = configs . get ( " with_hook " , False )
2022-07-01 19:35:28 +08:00
inner_config . combine_params = configs . get ( " combine_params " , False )
2022-09-23 19:32:16 +08:00
inner_config . clip_extra = configs . get ( " clip_extra " , True )
2022-09-15 20:43:44 +08:00
inner_config . skip_forward = configs . get ( " skip_forward " , False )
2023-08-21 17:24:02 +08:00
inner_config . input_names_after_prune = configs . get (
" input_names_after_prune " , None
)
2023-08-31 02:57:09 -05:00
inner_config . skip_prune_program = configs . get ( " skip_prune_program " , False )
2024-11-18 14:51:08 +08:00
inner_config . separate_parameters = configs . get ( " separate_parameters " , False )
2020-10-11 23:12:19 +08:00
return inner_config
2024-07-28 21:08:52 +08:00
def _parse_load_config ( configs : _LoadOptions ) - > _SaveLoadConfig :
2020-10-11 23:12:19 +08:00
supported_configs = [ ' model_filename ' , ' params_filename ' ]
# input check
for key in configs :
if key not in supported_configs :
raise ValueError (
2024-06-03 19:07:26 +08:00
f " The additional config ( { key } ) of `paddle.jit.load` is not supported. "
2022-10-23 20:01:27 +08:00
)
2020-10-11 23:12:19 +08:00
# construct inner config
inner_config = _SaveLoadConfig ( )
inner_config . model_filename = configs . get ( ' model_filename ' , None )
inner_config . params_filename = configs . get ( ' params_filename ' , None )
return inner_config
2024-04-19 16:32:32 +08:00
def _get_input_var_and_names ( inputs , input_spec , input_names_after_prune ) :
2022-10-23 20:01:27 +08:00
name_none_error = (
" The %s ' s name is None. "
2024-02-04 14:51:49 +08:00
" When using jit.save, please set InputSpec ' s name in "
2022-10-23 20:01:27 +08:00
" to_static(input_spec=[]) and jit.save(input_spec=[]) "
2020-09-10 15:12:56 +08:00
" and make sure they are consistent. "
2022-10-23 20:01:27 +08:00
)
name_no_exists_error = (
" The tensor ` %s ` does not exists. "
" Please make sure the name of InputSpec or example Tensor "
" in input_spec is the same as the name of InputSpec in "
2020-09-10 15:12:56 +08:00
" `to_static` decorated on the Layer.forward method. "
2022-10-23 20:01:27 +08:00
)
2023-08-21 17:24:02 +08:00
if input_names_after_prune is not None :
input_spec = [
x
for x in input_spec
if isinstance ( x , paddle . static . InputSpec )
and x . name in input_names_after_prune
]
2024-04-19 16:32:32 +08:00
input_vars = [
var
for var in paddle . utils . flatten ( inputs )
if isinstance ( var , ( Variable , paddle . pir . Value ) )
]
2020-11-19 10:07:19 +08:00
input_var_names = [
2023-03-09 20:30:52 +08:00
var . name
for var in paddle . utils . flatten ( inputs )
2024-04-19 16:32:32 +08:00
if isinstance ( var , ( Variable , paddle . pir . Value ) )
2020-11-19 10:07:19 +08:00
]
2024-04-19 16:32:32 +08:00
2020-09-10 15:12:56 +08:00
if input_spec is None :
# no prune
2024-04-19 16:32:32 +08:00
return input_vars , input_var_names
2021-06-17 09:44:52 +08:00
else :
2024-02-04 14:51:49 +08:00
# filter out non-tensor type spec infos.
2021-06-17 09:44:52 +08:00
input_spec = [
2022-10-23 20:01:27 +08:00
spec
for spec in input_spec
2021-06-17 09:44:52 +08:00
if isinstance ( spec , paddle . static . InputSpec )
]
2024-04-19 16:32:32 +08:00
result_var_list = [ ]
result_name_list = [ ]
2021-06-17 09:44:52 +08:00
if len ( input_spec ) == len ( input_var_names ) :
2020-09-10 15:12:56 +08:00
# no prune
2024-04-19 16:32:32 +08:00
result_var_list = input_vars
result_name_list = input_var_names
2021-01-20 09:56:38 +08:00
# if input spec name not in input_var_names, only raise warning
2020-09-10 15:12:56 +08:00
for spec in input_spec :
if spec . name is None :
warnings . warn ( name_none_error % spec )
elif spec . name not in input_var_names :
warnings . warn ( name_no_exists_error % spec . name )
else :
pass
else :
# prune
for spec in input_spec :
if spec . name is None :
# name is None, the input_spec only can be InputSpec
raise ValueError ( name_none_error % spec )
elif spec . name not in input_var_names :
2023-03-30 10:11:14 +08:00
# the input_spec can be `InputSpec` or `Tensor`
2020-09-10 15:12:56 +08:00
raise ValueError ( name_no_exists_error % spec . name )
else :
2024-04-19 16:32:32 +08:00
result_var_list . append ( spec )
result_name_list . append ( spec . name )
2020-09-10 15:12:56 +08:00
2024-04-19 16:32:32 +08:00
return result_var_list , result_name_list
2020-09-10 15:12:56 +08:00
2024-06-20 16:32:13 +08:00
def _contains_dict ( output ) :
if isinstance ( output , dict ) :
return True
if isinstance ( output , Sequence ) and not isinstance ( output , str ) :
return any ( _contains_dict ( i ) for i in output )
return False
2022-04-29 09:44:01 +08:00
def _get_output_vars ( outputs , output_spec , with_hook = False ) :
2022-10-23 20:01:27 +08:00
name_no_exists_error = (
" The tensor ` %s ` does not exists. "
" Please make sure the name of example Tensor "
" in configs.output_spec is the output tensor of "
2020-09-10 15:12:56 +08:00
" Layer.forward method. "
2022-10-23 20:01:27 +08:00
)
2024-05-07 19:51:38 +08:00
output_spec_is_not_value_error = (
" tensor ` %s ` is not support in pir mode, "
2025-01-20 11:12:06 +08:00
" because pir value has no name sometimes, especially as output, "
2024-05-17 10:55:29 +08:00
" so we can ' t check tensor ' s name with output var name, please "
" change as pir.value(to_static layer ' s output) "
" or int(the position of to_static layer ' s output) "
2024-05-07 19:51:38 +08:00
)
2022-04-29 09:44:01 +08:00
if output_spec and with_hook :
raise RuntimeError (
" Currently not support specify output_spec while founding pre/post hooks in your outermost layer. "
)
2020-09-10 15:12:56 +08:00
result_list = [ ]
2024-06-20 16:32:13 +08:00
if _contains_dict ( outputs ) :
warnings . warn (
" Found ' dict ' in given outputs, the values will be returned in a sequence sorted in lexicographical order by their keys. "
)
2024-04-19 16:32:32 +08:00
if use_pir_api ( ) :
2024-05-07 19:51:38 +08:00
from paddle . autograd . backward_utils import ValueSet
2024-04-19 16:32:32 +08:00
for var in paddle . utils . flatten ( outputs ) :
2024-10-16 18:50:57 +08:00
if isinstance ( var , paddle . pir . Value ) and var not in ValueSet (
result_list
) :
2024-04-19 16:32:32 +08:00
result_list . append ( var )
2024-05-07 19:51:38 +08:00
if output_spec is not None :
2024-05-17 10:55:29 +08:00
output_size = len ( result_list )
if len ( output_spec ) == output_size :
2024-05-07 19:51:38 +08:00
for var in output_spec :
2024-05-31 10:17:13 +08:00
if not isinstance ( var , ( paddle . pir . Value , int ) ) :
2024-05-07 19:51:38 +08:00
warnings . warn ( output_spec_is_not_value_error % var . name )
else :
if var not in ValueSet ( result_list ) :
warnings . warn ( name_no_exists_error % var . name )
else :
result_set = ValueSet ( result_list )
2024-05-17 10:55:29 +08:00
part_result_list = [ ]
2024-05-07 19:51:38 +08:00
for var in output_spec :
2024-05-17 10:55:29 +08:00
if isinstance ( var , paddle . pir . Value ) :
2024-05-07 19:51:38 +08:00
if var not in result_set :
raise ValueError ( name_no_exists_error % var . name )
else :
2024-05-17 10:55:29 +08:00
part_result_list . append ( var )
elif isinstance ( var , int ) :
if var > = output_size :
raise ValueError (
" position %d should smaller than output ' s size % d " ,
var ,
output_size ,
)
else :
part_result_list . append ( result_list [ var ] )
else :
raise ValueError (
output_spec_is_not_value_error % var . name
)
2024-05-07 19:51:38 +08:00
2024-05-17 10:55:29 +08:00
return part_result_list
2020-09-10 15:12:56 +08:00
else :
2024-04-19 16:32:32 +08:00
output_vars_dict = OrderedDict ( )
for var in paddle . utils . flatten ( outputs ) :
2024-05-17 10:55:29 +08:00
if isinstance ( var , ( Variable ) ) :
2024-04-19 16:32:32 +08:00
output_vars_dict [ var . name ] = var
if output_spec is None :
result_list = list ( output_vars_dict . values ( ) )
elif output_spec is not None and len ( output_spec ) == len (
output_vars_dict
) :
result_list = list ( output_vars_dict . values ( ) )
for var in output_spec :
if var . name not in output_vars_dict :
warnings . warn ( name_no_exists_error % var . name )
else :
for var in output_spec :
if var . name not in output_vars_dict :
raise ValueError ( name_no_exists_error % var . name )
else :
result_list . append ( output_vars_dict [ var . name ] )
2024-05-07 19:51:38 +08:00
2020-09-10 15:12:56 +08:00
return result_list
2020-10-11 23:12:19 +08:00
# NOTE(chenweihang): [ Handling of use cases of API paddle.jit.load ]
# `paddle.jit.load` may be used to load saved results of:
# 1. Expected cases:
# - paddle.jit.save
# - paddle.static.save_inference_model
# 2. Error cases:
# - paddle.save: no .pdmodel for prefix
# - paddle.static.save: no .pdiparams but .pdparams exists
2023-09-07 17:26:19 +08:00
# - paddle.base.io.save_params/save_persistables: no __model__
2020-10-11 23:12:19 +08:00
# TODO(chenweihang): polish error message in above error cases
def _build_load_path_and_config ( path , config ) :
# NOTE(chenweihang): If both [prefix save format] and [directory save format] exist,
# raise error, avoid confusing behavior
2024-04-25 13:58:01 +08:00
if use_pir_api ( ) :
model_suffix = PIR_INFER_MODEL_SUFFIX
else :
model_suffix = INFER_MODEL_SUFFIX
prefix_format_path = path + model_suffix
2020-10-11 23:12:19 +08:00
prefix_format_exist = os . path . exists ( prefix_format_path )
directory_format_exist = os . path . isdir ( path )
if prefix_format_exist and directory_format_exist :
raise ValueError (
2025-03-07 10:07:38 +08:00
f " The { path } .pdmodel(json) and { path } directory exist at the same time, "
2020-10-11 23:12:19 +08:00
" don ' t know which one to load, please make sure that the specified target "
2023-09-22 10:14:38 +08:00
" of ``path`` is unique. "
2022-10-23 20:01:27 +08:00
)
2020-10-11 23:12:19 +08:00
elif not prefix_format_exist and not directory_format_exist :
2022-10-23 20:01:27 +08:00
raise ValueError (
2024-06-03 19:07:26 +08:00
f " The ``path`` ( { path } ) to load model not exists. "
2025-03-07 10:07:38 +08:00
" Please make sure that *.pdmodel(json) exists or "
2024-06-03 19:07:26 +08:00
" don ' t using ``skip_forward=True`` to jit.save. "
2022-10-23 20:01:27 +08:00
)
2020-10-11 23:12:19 +08:00
else :
if prefix_format_exist :
file_prefix = os . path . basename ( path )
model_path = os . path . dirname ( path )
if config . model_filename is not None :
warnings . warn (
" When loading the result saved with the "
" specified file prefix, the ``model_filename`` config does "
2022-10-23 20:01:27 +08:00
" not take effect. "
)
2024-04-25 13:58:01 +08:00
config . model_filename = file_prefix + model_suffix
2020-10-11 23:12:19 +08:00
if config . params_filename is not None :
warnings . warn (
" When loading the result saved with the "
" specified file prefix, the ``params_filename`` config does "
2022-10-23 20:01:27 +08:00
" not take effect. "
)
2020-10-11 23:12:19 +08:00
config . params_filename = file_prefix + INFER_PARAMS_SUFFIX
else :
# Compatible with the old save_inference_model format
model_path = path
2020-09-07 19:51:53 +08:00
2020-10-11 23:12:19 +08:00
return model_path , config
2020-09-07 19:51:53 +08:00
2022-01-11 14:02:45 +08:00
_save_pre_hooks_lock = threading . Lock ( )
_save_pre_hooks = [ ]
2022-11-08 11:29:41 +08:00
class HookRemoveHelper :
2022-10-23 20:01:27 +08:00
""" A HookRemoveHelper that can be used to remove hook. """
2022-01-11 14:02:45 +08:00
def __init__ ( self , hook ) :
self . _hook = hook
def remove ( self ) :
_remove_save_pre_hook ( self . _hook )
def _register_save_pre_hook ( hook ) :
"""
Register a save pre-hook for `paddle.jit.save`.
This hook will be executed before `save` function has been invoked.
hook(layer, input_spec, configs) -> None
- layer (Layer|function): This argument is corresponding to `layer` in `paddle.jit.save`.
- input_spec (list or tuple[InputSpec|Tensor|Python built-in variable]): This argument is corresponding to `input_spec` in `paddle.jit.save`.
- configs (dict): This argument is corresponding to `configs` in `paddle.jit.save`.
Args:
hook(function): a function registered as a save pre-hook
Returns:
HookRemoveHelper: a HookRemoveHelper object that can be used to remove the added hook by calling `hook_remove_helper.remove()`.
Examples:
2026-02-20 17:21:20 +08:00
.. code-block:: pycon
2022-01-11 14:02:45 +08:00
2023-09-26 11:19:57 +08:00
>>> # doctest: +SKIP( ' `paddle.jit.api.to_static` can not run in xdoctest ' )
2023-07-31 14:44:48 +08:00
>>> import numpy as np
>>> import paddle
>>> IMAGE_SIZE = 256
>>> CLASS_NUM = 10
>>> class LinearNet(paddle.nn.Layer):
... def __init__(self):
... super().__init__()
... self._linear = paddle.nn.Linear(IMAGE_SIZE, CLASS_NUM)
...
... def forward(self, x):
... return self._linear(x)
>>> saving_count = 0
>>> def save_pre_hook(layer, input_spec, configs):
... global saving_count
... saving_count += 1
>>> remove_handler = paddle.jit.api._register_save_pre_hook(save_pre_hook)
>>> layer = LinearNet()
>>> paddle.jit.save(layer, " /tmp " , [paddle.static.InputSpec(shape=[-1, IMAGE_SIZE])])
>>> print(saving_count)
1
>>> remove_handler.remove()
>>> paddle.jit.save(layer, " /tmp " , [paddle.static.InputSpec(shape=[-1, IMAGE_SIZE])])
>>> print(saving_count)
1
2022-01-11 14:02:45 +08:00
"""
global _save_pre_hooks_lock
global _save_pre_hooks
_save_pre_hooks_lock . acquire ( )
if hook not in _save_pre_hooks :
_save_pre_hooks . append ( hook )
_save_pre_hooks_lock . release ( )
return HookRemoveHelper ( hook )
def _clear_save_pre_hooks ( ) :
global _save_pre_hooks_lock
global _save_pre_hooks
_save_pre_hooks_lock . acquire ( )
_save_pre_hooks . clear ( )
_save_pre_hooks_lock . release ( )
def _remove_save_pre_hook ( hook ) :
global _save_pre_hooks_lock
global _save_pre_hooks
_save_pre_hooks_lock . acquire ( )
if hook in _save_pre_hooks :
_save_pre_hooks . remove ( hook )
_save_pre_hooks_lock . release ( )
2024-06-21 00:40:47 +08:00
class _SaveFunction ( Protocol ) :
def __call__ (
self ,
layer : Layer | Callable [ . . . , Any ] ,
path : str ,
2024-08-01 18:33:37 +08:00
input_spec : Sequence [ InputSpec | Tensor | object ] | None = . . . ,
2024-07-28 21:08:52 +08:00
* * configs : Unpack [ _SaveOptions ] ,
2024-08-08 10:10:45 +08:00
) - > None : . . .
2024-06-21 00:40:47 +08:00
2022-09-06 11:27:50 +08:00
@wrap_decorator
2024-06-21 00:40:47 +08:00
def _run_save_pre_hooks ( func : _SaveFunction ) - > _SaveFunction :
2024-06-14 14:32:24 +08:00
def wrapper (
layer : Layer | Callable [ . . . , Any ] ,
path : str ,
2024-08-01 18:33:37 +08:00
input_spec : Sequence [ InputSpec | Tensor | object ] | None = None ,
2024-07-28 21:08:52 +08:00
* * configs : Unpack [ _SaveOptions ] ,
2024-06-14 14:32:24 +08:00
) - > None :
2022-01-11 14:02:45 +08:00
global _save_pre_hooks
for hook in _save_pre_hooks :
hook ( layer , input_spec , configs )
func ( layer , path , input_spec , * * configs )
2024-06-21 00:40:47 +08:00
return wrapper
2022-01-11 14:02:45 +08:00
2023-01-04 11:55:21 +08:00
def _save_property ( filename : str , property_vals : list [ tuple [ Any , str ] ] ) :
2022-07-27 14:18:07 +08:00
""" class property serialization.
Args:
2023-01-04 11:55:21 +08:00
filename (str): *.meta
property_vals (list[tuple[Any, str]]): class property.
2022-07-27 14:18:07 +08:00
"""
def set_property ( meta , key , val ) :
if isinstance ( val , float ) :
meta . set_float ( key , val )
elif isinstance ( val , int ) :
meta . set_int ( key , val )
elif isinstance ( val , str ) :
meta . set_string ( key , val )
elif isinstance ( val , ( tuple , list ) ) :
if isinstance ( val [ 0 ] , float ) :
meta . set_floats ( key , val )
elif isinstance ( val [ 0 ] , int ) :
meta . set_ints ( key , val )
elif isinstance ( val [ 0 ] , str ) :
meta . set_strings ( key , val )
else :
raise ValueError ( f " Note support val type: { type ( val ) } " )
with open ( filename , ' wb ' ) as f :
meta = paddle . framework . core . Property ( )
for item in property_vals :
val , key = item [ 0 ] , item [ 1 ]
set_property ( meta , key , val )
f . write ( meta . serialize_to_string ( ) )
2024-07-18 11:18:04 +08:00
def _get_function_names_from_layer ( layer : Layer ) - > list [ str ] :
cls = layer . __class__
return [
member_name
for member_name , member in inspect . getmembers ( cls )
if (
inspect . isfunction ( member )
or inspect . ismethod ( member )
or inspect . ismethoddescriptor ( member )
)
]
2022-01-11 14:02:45 +08:00
@_run_save_pre_hooks
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
@switch_to_static_graph
2024-06-05 17:15:45 +08:00
def save (
2024-06-14 14:32:24 +08:00
layer : Layer | Callable [ . . . , Any ] ,
2024-06-05 17:15:45 +08:00
path : str ,
2024-08-01 18:33:37 +08:00
input_spec : Sequence [ InputSpec | Tensor | object ] | None = None ,
2024-07-28 21:08:52 +08:00
* * configs : Unpack [ _SaveOptions ] ,
2024-06-05 17:15:45 +08:00
) - > None :
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
"""
2021-04-27 14:08:52 +08:00
Saves input Layer or function as ``paddle.jit.TranslatedLayer``
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
format model, which can be used for inference or fine-tuning after loading.
2021-04-07 16:38:02 +08:00
It will save the translated program and all related persistable
2020-10-13 12:32:00 +08:00
variables of input Layer to given ``path`` .
2021-04-07 16:38:02 +08:00
``path`` is the prefix of saved objects, and the saved translated program file
2020-10-13 12:32:00 +08:00
suffix is ``.pdmodel`` , the saved persistable variables file suffix is ``.pdiparams`` ,
2021-04-07 16:38:02 +08:00
and here also saved some additional variable description information to a file,
2020-10-11 23:12:19 +08:00
its suffix is ``.pdiparams.info``, these additional information is used in fine-tuning.
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
The saved model can be loaded by follow APIs:
2021-04-07 16:38:02 +08:00
- ``paddle.jit.load``
- ``paddle.static.load_inference_model``
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
- Other C++ inference APIs
2021-04-27 14:08:52 +08:00
.. note::
2022-09-14 21:56:19 +08:00
When using ``paddle.jit.save`` to save a function, parameters will not be saved. If you have to
2021-04-27 14:08:52 +08:00
save the parameter, please pass the Layer containing function and parameter to ``paddle.jit.save``.
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
Args:
2021-04-27 14:08:52 +08:00
layer (Layer|function): The Layer or function to be saved.
2020-10-11 23:12:19 +08:00
path (str): The path prefix to save model. The format is ``dirname/file_prefix`` or ``file_prefix``.
2021-06-17 09:44:52 +08:00
input_spec (list or tuple[InputSpec|Tensor|Python built-in variable], optional): Describes the input of the saved model ' s forward
method, which can be described by InputSpec or example Tensor. Moreover, we support to specify non-tensor type argument,
such as int, float, string, or list/dict of them.If None, all input variables of
2020-10-13 12:32:00 +08:00
the original Layer ' s forward method would be the inputs of the saved model. Default None.
2021-04-07 16:38:02 +08:00
**configs (dict, optional): Other save configuration options for compatibility. We do not
recommend using these configurations, they may be removed in the future. If not necessary,
2020-10-11 23:12:19 +08:00
DO NOT use them. Default None.
The following options are currently supported:
2024-05-17 10:55:29 +08:00
(1) output_spec (list[Tensor|Value|int]): Selects the output targets of the saved model,
2021-04-07 16:38:02 +08:00
By default, all return variables of original Layer ' s forward method are kept as the
output of the saved model. If the provided ``output_spec`` list is not all output variables,
the saved model will be pruned according to the given ``output_spec`` list.
2024-05-17 10:55:29 +08:00
in pir mode, Tensor is not supported, because value has no name in most cases,
which can ' t be used to judge which tensor corresponds to which value; the value can ' t be found
if the saved program is not the same as the program that includes output_spec, so we need to
use the position of the output.
2020-10-11 23:12:19 +08:00
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
Returns:
None
Examples:
2026-02-20 17:21:20 +08:00
.. code-block:: pycon
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2023-09-26 11:19:57 +08:00
>>> # doctest: +SKIP( ' `paddle.jit.to_static` can not run in xdoctest ' )
2023-07-31 14:44:48 +08:00
>>> # example 1: save layer
>>> 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
2026-02-20 17:21:20 +08:00
>>> class RandomDataset(paddle.io.Dataset): # type: ignore[type-arg]
2023-07-31 14:44:48 +08:00
... def __init__(self, num_samples):
... self.num_samples = num_samples
...
... def __getitem__(self, idx):
... image = np.random.random([IMAGE_SIZE]).astype( ' float32 ' )
2026-02-20 17:21:20 +08:00
... label = np.random.randint(0, CLASS_NUM - 1, (1,)).astype( ' int64 ' )
2023-07-31 14:44:48 +08:00
... 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()
2026-02-20 17:21:20 +08:00
... print( " Epoch {} batch {} : loss = {} " .format(epoch_id, batch_id, np.mean(loss.numpy())))
2023-07-31 14:44:48 +08:00
>>> # 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)
2026-02-20 17:21:20 +08:00
>>> loader = paddle.io.DataLoader(
... dataset,
2023-07-31 14:44:48 +08:00
... batch_size=BATCH_SIZE,
... shuffle=True,
... drop_last=True,
2026-02-20 17:21:20 +08:00
... num_workers=2,
2023-07-31 14:44:48 +08:00
... )
>>> # train
>>> train(layer, loader, loss_fn, adam)
>>> # save
>>> path = " example_model/linear "
>>> paddle.jit.save(layer, path)
>>> # example 2: save function
>>> import paddle
>>> from paddle.static import InputSpec
>>> def save_function():
... @paddle.jit.to_static
... def fun(inputs):
... return paddle.tanh(inputs)
...
... path = ' test_jit_save_load_function_1/func '
... inps = paddle.rand([3, 6])
... origin = fun(inps)
...
... paddle.jit.save(fun, path)
... load_func = paddle.jit.load(path)
...
... load_result = load_func(inps)
... print((load_result - origin).abs().max() < 1e-10)
>>> save_function()
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
"""
2020-10-11 23:12:19 +08:00
# 1. input build & check
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
prog_translator = ProgramTranslator ( )
2023-04-04 15:31:44 +08:00
is_prim_infer = core . _is_fwd_prim_enabled ( ) and core . _is_bwd_prim_enabled ( )
2020-09-17 10:23:42 +08:00
if not prog_translator . enable_to_static :
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
raise RuntimeError (
2023-01-11 11:03:21 +08:00
" The paddle.jit.save doesn ' t work when setting ' paddle.jit.enable_to_static ' to False. "
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
)
2021-04-27 14:08:52 +08:00
2022-10-23 20:01:27 +08:00
if not (
2023-03-28 14:40:48 +08:00
isinstance ( layer , ( Layer , StaticFunction ) ) or inspect . isfunction ( layer )
2022-10-23 20:01:27 +08:00
) :
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
raise TypeError (
2024-06-03 19:07:26 +08:00
f " The input of paddle.jit.save should be ' Layer ' or ' Function ' , but received input type is { type ( layer ) } . "
2022-10-23 20:01:27 +08:00
)
2021-07-14 15:32:31 +08:00
elif inspect . isfunction ( layer ) or isinstance ( layer , StaticFunction ) :
warnings . warn (
' What you save is a function, and `jit.save` will generate the name of the model file according to `path` you specify. When loading these files with `jit.load`, you get a `TranslatedLayer` whose inference result is the same as the inference result of the function you saved. '
)
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2020-11-27 09:59:31 +08:00
# NOTE(chenweihang): If the input layer be wrapped by DataParallel,
# the args and kwargs of forward method will can't be parsed by
2021-01-20 09:56:38 +08:00
# function_spec, so here we save DataParallel._layers instead
2020-11-27 09:59:31 +08:00
# DataParallel it self
# NOTE(chenweihang): using inner_layer, do not change input layer
if isinstance ( layer , paddle . DataParallel ) :
inner_layer = layer . _layers
else :
inner_layer = layer
2020-10-11 23:12:19 +08:00
# path check
file_prefix = os . path . basename ( path )
if file_prefix == " " :
raise ValueError (
" The input path MUST be format of dirname/file_prefix "
" [dirname \\ file_prefix in Windows system], but received "
2022-10-23 20:01:27 +08:00
" file_prefix is empty string. "
)
2020-10-11 23:12:19 +08:00
dirname = os . path . dirname ( path )
if dirname and not os . path . exists ( dirname ) :
os . makedirs ( dirname )
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2020-09-10 15:12:56 +08:00
# avoid change user given input_spec
inner_input_spec = None
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
if input_spec is not None :
2021-04-27 14:08:52 +08:00
if isinstance ( layer , Layer ) :
2024-07-18 11:18:04 +08:00
for member_name in _get_function_names_from_layer ( inner_layer ) :
static_func = getattr ( inner_layer , member_name , None )
2022-10-23 20:01:27 +08:00
if (
isinstance ( static_func , StaticFunction )
2024-07-18 11:18:04 +08:00
and ' forward ' != member_name
2022-10-23 20:01:27 +08:00
) :
2021-04-27 14:08:52 +08:00
raise ValueError (
2024-06-03 19:07:26 +08:00
f " If there are static functions other than ' forward ' that need to be saved, the input ' input_spec ' should be None, but received the type of ' input_spec ' is { type ( input_spec ) } . "
2022-10-23 20:01:27 +08:00
)
2021-04-27 14:08:52 +08:00
2021-04-26 15:32:13 +08:00
if not isinstance ( input_spec , ( list , tuple ) ) :
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
raise TypeError (
2024-06-03 19:07:26 +08:00
f " The input input_spec should be ' list ' , but received input_spec ' s type is { type ( input_spec ) } . "
2022-10-23 20:01:27 +08:00
)
2020-09-10 15:12:56 +08:00
inner_input_spec = [ ]
2023-03-09 20:30:52 +08:00
for var in paddle . utils . flatten ( input_spec ) :
2020-09-10 15:12:56 +08:00
if isinstance ( var , paddle . static . InputSpec ) :
inner_input_spec . append ( var )
2024-05-07 19:51:38 +08:00
elif isinstance (
var , ( core . eager . Tensor , Variable , paddle . pir . Value )
) :
2020-09-10 15:12:56 +08:00
inner_input_spec . append (
2022-10-23 20:01:27 +08:00
paddle . static . InputSpec . from_tensor ( var )
)
2020-09-10 15:12:56 +08:00
else :
2021-06-17 09:44:52 +08:00
# NOTE(Aurelius84): Support non-Tensor type in `input_spec`.
inner_input_spec . append ( var )
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2020-10-11 23:12:19 +08:00
# parse configs
configs = _parse_save_configs ( configs )
2022-04-29 09:44:01 +08:00
# whether outermost layer has pre/post hook, if does, we need also save
2022-06-05 10:58:58 +08:00
# these operators in program.
2022-04-29 09:44:01 +08:00
with_hook = configs . with_hook
2022-06-30 19:17:49 +08:00
combine_params = configs . combine_params
if combine_params :
configs . _program_only = True
2022-04-29 09:44:01 +08:00
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
scope = core . Scope ( )
2023-03-23 10:16:17 +08:00
extra_var_info = { }
2021-04-27 14:08:52 +08:00
if isinstance ( layer , Layer ) :
2024-07-18 11:18:04 +08:00
functions = list ( set ( _get_function_names_from_layer ( inner_layer ) ) )
2023-12-06 19:32:07 +08:00
functions = sorted ( functions )
2022-04-29 09:44:01 +08:00
if inner_layer . _forward_pre_hooks or inner_layer . _forward_post_hooks :
with_hook = True
2021-04-27 14:08:52 +08:00
else :
# layer is function
2024-07-18 11:18:04 +08:00
functions = [ layer ]
2022-06-30 19:17:49 +08:00
2022-07-25 10:31:25 +08:00
combine_vars = { }
2024-05-07 19:51:38 +08:00
combine_program = [ ]
2022-06-30 19:17:49 +08:00
property_vals = [ ] # (value, key)
2022-09-15 20:43:44 +08:00
concrete_program = None
2021-04-27 14:08:52 +08:00
for attr_func in functions :
if isinstance ( layer , Layer ) :
2023-08-07 15:10:33 +08:00
static_func = get_ast_static_function (
getattr ( inner_layer , attr_func , None )
)
2021-04-27 14:08:52 +08:00
if isinstance ( static_func , StaticFunction ) :
2022-06-30 19:17:49 +08:00
if static_func . is_property :
# property method to be exported
immediate_val = static_func ( )
property_vals . append (
2022-10-23 20:01:27 +08:00
(
immediate_val ,
layer . __class__ . __name__ + ' . ' + attr_func ,
)
)
2022-06-30 19:17:49 +08:00
continue
2022-10-23 20:01:27 +08:00
concrete_program = (
static_func . concrete_program_specify_input_spec (
2023-04-04 15:31:44 +08:00
inner_input_spec ,
with_hook = with_hook ,
is_prim_infer = is_prim_infer ,
2022-10-23 20:01:27 +08:00
)
)
2024-05-07 19:51:38 +08:00
2021-04-27 14:08:52 +08:00
elif ' forward ' == attr_func :
2022-09-15 20:43:44 +08:00
if configs . skip_forward :
# do not jit.save forward function
continue
2021-04-27 14:08:52 +08:00
# transform in jit.save, if input_spec is incomplete, declarative will throw error
2021-06-17 09:44:52 +08:00
# inner_input_spec is list[InputSpec], it should be packed with same structure
2021-04-27 14:08:52 +08:00
# as original input_spec here.
if inner_input_spec :
2023-03-09 20:30:52 +08:00
inner_input_spec = paddle . utils . pack_sequence_as (
2022-10-23 20:01:27 +08:00
input_spec , inner_input_spec
)
2022-12-19 19:12:43 +08:00
static_forward = to_static (
2023-08-07 15:10:33 +08:00
inner_layer . forward ,
input_spec = inner_input_spec ,
2023-10-20 20:05:25 +08:00
full_graph = True ,
2025-04-16 12:14:30 +08:00
backend = None ,
2022-10-23 20:01:27 +08:00
)
2024-05-07 19:51:38 +08:00
2022-10-23 20:01:27 +08:00
concrete_program = (
static_forward . concrete_program_specify_input_spec (
2023-04-04 15:31:44 +08:00
with_hook = with_hook , is_prim_infer = is_prim_infer
2022-10-23 20:01:27 +08:00
)
)
2021-04-27 14:08:52 +08:00
# the input_spec has been used in declarative, which is equal to
2022-12-19 19:12:43 +08:00
# @to_static with input_spec and jit.save without input_spec,
2021-04-27 14:08:52 +08:00
# avoid needless warning
inner_input_spec = None
else :
continue
2021-07-14 15:32:31 +08:00
else :
# When layer is a function
if isinstance ( attr_func , StaticFunction ) :
2023-08-07 15:10:33 +08:00
static_func = get_ast_static_function ( attr_func )
if static_func . is_property :
2022-06-30 19:17:49 +08:00
# property method to be exported
2023-08-07 15:10:33 +08:00
immediate_val = static_func ( )
property_vals . append ( ( immediate_val , static_func ) )
2022-06-30 19:17:49 +08:00
continue
2022-10-23 20:01:27 +08:00
concrete_program = (
2023-08-07 15:10:33 +08:00
static_func . concrete_program_specify_input_spec (
2023-04-04 15:31:44 +08:00
inner_input_spec , is_prim_infer = is_prim_infer
2022-10-23 20:01:27 +08:00
)
)
2021-07-14 15:32:31 +08:00
else :
2023-08-07 15:10:33 +08:00
static_func = get_ast_static_function ( attr_func )
2021-07-14 15:32:31 +08:00
if inner_input_spec :
2023-03-09 20:30:52 +08:00
inner_input_spec = paddle . utils . pack_sequence_as (
2022-10-23 20:01:27 +08:00
input_spec , inner_input_spec
)
2022-12-19 19:12:43 +08:00
static_function = to_static (
2023-08-07 15:10:33 +08:00
static_func ,
input_spec = inner_input_spec ,
2023-10-20 20:05:25 +08:00
full_graph = True ,
2025-04-16 12:14:30 +08:00
backend = None ,
2022-10-23 20:01:27 +08:00
)
2021-07-14 15:32:31 +08:00
concrete_program = static_function . concrete_program
2024-10-23 23:47:18 +08:00
if static_function . class_instance is None :
2021-07-14 15:32:31 +08:00
warnings . warn (
2023-11-02 18:28:29 +08:00
f ' `jit.save` will only save the `Program`, not the parameters. If you have to save the parameters, please make sure that { layer } is a member function of `paddle.nn.Layer` and the saved parameters are in `state_dict` '
2022-10-23 20:01:27 +08:00
)
2021-07-14 15:32:31 +08:00
2022-06-30 19:17:49 +08:00
# when save multi `StaticFunction`, all `StaticFunction` share params.
2021-07-14 15:32:31 +08:00
dygraph_state_dict = None
if isinstance ( inner_layer , Layer ) :
2021-09-17 23:49:30 +08:00
dygraph_state_dict = inner_layer . to_static_state_dict ( )
2021-07-14 15:32:31 +08:00
elif isinstance ( attr_func , StaticFunction ) :
2024-10-23 23:47:18 +08:00
if static_func . class_instance :
2022-10-23 20:01:27 +08:00
dygraph_state_dict = (
2024-10-23 23:47:18 +08:00
static_func . class_instance . to_static_state_dict ( )
2021-09-17 23:49:30 +08:00
)
2021-07-14 15:32:31 +08:00
if dygraph_state_dict :
2021-04-27 14:08:52 +08:00
# NOTE(chenweihang): we maintain the mapping of variable name to
# structured name, the buffer variable (non-persistable)
# saved to inference program may not need by dygraph Layer,
# we only record the state_dict variable's structured name
2023-03-23 10:16:17 +08:00
state_names_dict = { }
state_var_dict = { }
2022-10-19 15:54:41 +08:00
for structured_name , var in dygraph_state_dict . items ( ) :
2021-04-27 14:08:52 +08:00
state_names_dict [ var . name ] = structured_name
2021-09-17 23:49:30 +08:00
state_var_dict [ var . name ] = var
2022-06-30 19:17:49 +08:00
# 3. share parameters from Layer to scope & record var info
with dygraph . guard ( ) :
2024-04-19 16:32:32 +08:00
if use_pir_api ( ) :
for tensor , value in zip ( * concrete_program . parameters ) :
2024-05-23 10:16:18 +08:00
if not value . persistable :
continue
2024-04-19 16:32:32 +08:00
param_or_buffer_tensor = scope . var ( value . name ) . get_tensor ( )
2022-10-23 20:01:27 +08:00
src_tensor = (
2024-04-19 16:32:32 +08:00
state_var_dict [ tensor . name ] . value ( ) . get_tensor ( )
2022-10-23 20:01:27 +08:00
)
2022-06-30 19:17:49 +08:00
param_or_buffer_tensor . _share_data_with ( src_tensor )
2024-05-23 10:16:18 +08:00
2024-04-19 16:32:32 +08:00
else :
for param_or_buffer in concrete_program . parameters :
# share to scope
if param_or_buffer . type == core . VarDesc . VarType . VOCAB :
scr_tensor = param_or_buffer . value ( ) . get_map_tensor ( )
tgt_var = scope . var ( param_or_buffer . name )
tgt_var . set_vocab ( scr_tensor )
else :
param_or_buffer_tensor = scope . var (
2022-10-23 20:01:27 +08:00
param_or_buffer . name
2024-04-19 16:32:32 +08:00
) . get_tensor ( )
# src_tensor = param_or_buffer.value().get_tensor()
src_tensor = (
state_var_dict [ param_or_buffer . name ]
. value ( )
. get_tensor ( )
)
param_or_buffer_tensor . _share_data_with ( src_tensor )
# record var info
if param_or_buffer . name not in extra_var_info :
extra_info_dict = { }
if param_or_buffer . name in state_names_dict :
2024-08-08 10:10:45 +08:00
extra_info_dict [ ' structured_name ' ] = (
state_names_dict [ param_or_buffer . name ]
)
extra_info_dict [ ' stop_gradient ' ] = (
param_or_buffer . stop_gradient
)
2024-04-19 16:32:32 +08:00
if isinstance ( param_or_buffer , EagerParamBase ) :
2024-08-08 10:10:45 +08:00
extra_info_dict [ ' trainable ' ] = (
param_or_buffer . trainable
)
2024-04-19 16:32:32 +08:00
extra_var_info [ param_or_buffer . name ] = extra_info_dict
2024-02-04 14:51:49 +08:00
# 4. build input & output of save_inference_model
2020-11-06 19:41:28 +08:00
# NOTE(chenweihang): [ Get input variables name ]
# There are two cases, whether to prune the inputs or not
# - not prune inputs (recommend):
# - the len(input_spec) == len((concrete_program.inputs) - 1
# - here can use concrete_program.inputs directly
# - prune inputs:
# - the input_spec length < len((concrete_program.inputs) - 1
# - the input_spec's name should be in concrete_program.inputs
2024-04-19 16:32:32 +08:00
input_vars , input_var_names = _get_input_var_and_names (
2023-08-21 17:24:02 +08:00
concrete_program . inputs ,
inner_input_spec ,
configs . input_names_after_prune ,
2022-10-23 20:01:27 +08:00
)
2020-11-06 19:41:28 +08:00
# NOTE(chenweihang): [ Get output variables ]
2021-01-20 09:56:38 +08:00
# the rule is like [ Get input variables name ]. For output var,
2023-03-30 10:11:14 +08:00
# we only support Tensor spec, and actually, we only need the
2020-11-06 19:41:28 +08:00
# var name of output, and we don't recommended to use output_spec
2024-05-17 10:55:29 +08:00
# NOTE(Ruting): in pir mode, Tensor is not supported, because value has no name in most cases,
# which can't be used to judge which tensor corresponds to which value; the value can't be found
# if the saved program is not the same as the program that includes output_spec, so we need to
# use the position of the output.
2024-04-19 16:32:32 +08:00
2022-10-23 20:01:27 +08:00
output_vars = _get_output_vars (
concrete_program . outputs , configs . output_spec , with_hook
)
2020-11-06 19:41:28 +08:00
# 5. save inference model
# construct new save_inference_model arguments
model_path = dirname
# NOTE(chenweihang): because prefix contains model and params filename,
2021-01-20 09:56:38 +08:00
# so we don't support set model_filename & params_filename
2021-04-27 14:08:52 +08:00
if ' forward ' == attr_func or not isinstance ( layer , Layer ) :
2020-11-06 19:41:28 +08:00
model_filename = file_prefix + INFER_MODEL_SUFFIX
params_filename = file_prefix + INFER_PARAMS_SUFFIX
2023-08-11 11:09:21 +08:00
path_prefix = file_prefix
2020-11-06 19:41:28 +08:00
else :
model_filename = file_prefix + ' . ' + attr_func + INFER_MODEL_SUFFIX
2022-10-23 20:01:27 +08:00
params_filename = (
file_prefix + ' . ' + attr_func + INFER_PARAMS_SUFFIX
)
2024-05-07 19:51:38 +08:00
path_prefix = file_prefix + ' . ' + attr_func
file_path = os . path . join ( model_path , path_prefix )
2020-11-06 19:41:28 +08:00
with scope_guard ( scope ) :
2024-05-07 19:51:38 +08:00
if use_pir_api ( ) :
value_map = paddle . pir . IrMapping ( )
clone_program = concrete_program . main_program . clone ( value_map )
clone_input_vars = [ ]
for v in input_vars :
if type ( v ) is paddle . static . InputSpec :
name = v . name
for op in clone_program . global_block ( ) . ops :
if (
op . name ( ) == ' pd_op.data '
and op . attrs ( ) [ " name " ] == name
) :
clone_input_vars . append ( op . result ( 0 ) )
else :
clone_input_vars . append ( value_map . look_up ( v ) )
clone_output_vars = [ value_map . look_up ( v ) for v in output_vars ]
else :
2024-04-19 16:32:32 +08:00
input_vars = [
concrete_program . main_program . global_block ( ) . var ( name )
for name in input_var_names
]
clone_program = concrete_program . main_program . clone ( )
clone_input_vars = input_vars
clone_output_vars = output_vars
2020-11-06 19:41:28 +08:00
save_inference_model (
2024-05-07 19:51:38 +08:00
path_prefix = file_path ,
2024-04-19 16:32:32 +08:00
feed_vars = clone_input_vars ,
fetch_vars = clone_output_vars ,
2020-11-06 19:41:28 +08:00
executor = Executor ( _current_expected_place ( ) ) ,
2024-04-19 16:32:32 +08:00
program = clone_program ,
2022-10-23 20:01:27 +08:00
clip_extra = configs . clip_extra ,
2023-08-31 02:57:09 -05:00
skip_prune_program = configs . skip_prune_program ,
2024-11-18 14:51:08 +08:00
separate_parameters = configs . separate_parameters ,
2022-10-23 20:01:27 +08:00
)
2020-11-06 19:41:28 +08:00
2022-07-25 10:31:25 +08:00
if combine_params :
2024-05-07 19:51:38 +08:00
if use_pir_api ( ) :
# NOTE(Ruting): concrete_program has been pruned when init partialProgramLayer,
[CodeStyle][Typos][[N-[1-5],N-[9,10],N-[12-14]] Fix typo(`namess`, `nams`, `neccessary`, `Neet`, `neet`, `neeed`, `nedd`,`neighor`,`netwrok`,`normlize`,`noraml`,`numer`) (#70283)
* fix typos
* Update _typos.toml
* fix typos
2024-12-20 09:55:47 +08:00
# so we do not need to prune again.
2024-05-07 19:51:38 +08:00
for var in concrete_program . main_program . list_vars ( ) :
if var . persistable :
combine_vars [ var . name ] = var
# NOTE(Ruting): concrete_program will delete after this loop item,
# value delete at the same time, so we use list to Extend its lifecycle
combine_program . append ( concrete_program . main_program )
else :
clone_main_program = concrete_program . main_program . clone ( )
clone_main_program = clone_main_program . _prune_with_input (
input_var_names , output_vars
)
for block in clone_main_program . blocks :
combine_vars . update ( block . vars )
2022-06-30 19:17:49 +08:00
# save shared params
if combine_params :
2022-07-25 10:31:25 +08:00
# sort vars by name
combine_vars = sorted ( combine_vars . items ( ) , key = lambda item : item [ 0 ] )
ordered_vars = [ ]
for name , var in combine_vars :
ordered_vars . append ( var )
2022-06-30 19:17:49 +08:00
params_filename = file_prefix + INFER_PARAMS_SUFFIX
with scope_guard ( scope ) :
2024-05-07 19:51:38 +08:00
if use_pir_api ( ) :
paddle . static . save_vars (
Executor ( _current_expected_place ( ) ) ,
dirname = model_path ,
vars = ordered_vars ,
filename = params_filename ,
)
else :
paddle . static . save_vars (
Executor ( _current_expected_place ( ) ) ,
dirname = model_path ,
vars = list (
filter (
paddle . framework . io_utils . is_persistable ,
ordered_vars ,
)
) ,
filename = params_filename ,
)
2022-07-27 14:18:07 +08:00
# save property
2022-10-23 20:01:27 +08:00
property_save_path = os . path . join (
os . path . normpath ( model_path ) , file_prefix + INFER_PROPERTY_SUFFIX
)
2022-09-19 11:11:40 +08:00
_save_property ( property_save_path , property_vals )
2022-06-30 19:17:49 +08:00
2020-11-06 19:41:28 +08:00
# NOTE(chenweihang): [ Save extra variable info ]
# save_inference_model will lose some important variable information, including:
# - Variable name and correspondence (when saved variables as one file)
# - Variable.stop_gradient information
# - Which persistent variable are parameter and which are not
# - Parameter.trainable information
#
2021-01-20 09:56:38 +08:00
# 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
2020-11-06 19:41:28 +08:00
# configure redundant information to proceed.
#
2021-01-20 09:56:38 +08:00
# Due to compatibility issues, we cannot change the original storage structure,
# but we can save these information in `jit.save` without changing the original
2020-11-06 19:41:28 +08:00
# storage to improve user experience. So we save extra information into
# file `***.pdiparams.info`
2021-07-14 15:32:31 +08:00
# "layer" can only be Layer or function or StaticFunction.
contain_parameter = False
2022-09-15 20:43:44 +08:00
if concrete_program is not None :
for var in concrete_program . main_program . list_vars ( ) :
2024-04-19 16:32:32 +08:00
if use_pir_api ( ) :
is_persistable = (
var . get_defining_op ( ) . has_attr ( " persistable " )
and var . get_defining_op ( ) . attrs ( ) [ " persistable " ] is True
)
contain_parameter | = is_persistable
else :
contain_parameter | = isinstance ( var , Parameter )
2021-07-14 15:32:31 +08:00
if ( isinstance ( layer , Layer ) or contain_parameter ) and extra_var_info :
2021-04-27 14:08:52 +08:00
with scope_guard ( scope ) :
extra_var_info_path = path + INFER_PARAMS_INFO_SUFFIX
with open ( extra_var_info_path , ' wb ' ) as f :
pickle . dump ( extra_var_info , f , protocol = 2 )
2024-10-17 11:15:34 +08:00
scope . erase ( scope . local_var_names ( ) )
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
@dygraph_only
2024-06-05 17:15:45 +08:00
def load (
2024-07-28 21:08:52 +08:00
path : str , * * configs : Unpack [ _LoadOptions ]
2024-06-05 17:15:45 +08:00
) - > TranslatedLayer | PirTranslatedLayer :
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
"""
:api_attr: imperative
2021-04-07 16:38:02 +08:00
Load model saved by ``paddle.jit.save`` or ``paddle.static.save_inference_model`` or
2023-09-26 11:19:57 +08:00
paddle 1.x API ``paddle.static.save_inference_model`` as ``paddle.jit.TranslatedLayer``,
2020-10-11 23:12:19 +08:00
then performing inference or fine-tune training.
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
.. note::
2020-10-11 23:12:19 +08:00
If you load model saved by ``paddle.static.save_inference_model`` ,
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
there will be the following limitations when using it in fine-tuning:
2024-11-20 11:46:17 +08:00
1. Imperative mode do not support DenseTensor. All original model ' s feed targets or parameters that depend on LoD are temporarily unavailable.
2020-08-27 18:20:05 +08:00
2. All saved model ' s feed targets need to be passed into TranslatedLayer ' s forward function.
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
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:
2020-10-13 12:32:00 +08:00
path (str): The path prefix to load model. The format is ``dirname/file_prefix`` or ``file_prefix`` .
2021-04-07 16:38:02 +08:00
**configs (dict, optional): Other load configuration options for compatibility. We do not
recommend using these configurations, they may be removed in the future. If not necessary,
2020-10-11 23:12:19 +08:00
DO NOT use them. Default None.
The following options are currently supported:
2021-04-07 16:38:02 +08:00
(1) model_filename (str): The inference model file name of the paddle 1.x
``save_inference_model`` save format. Default file name is :code:`__model__` .
(2) params_filename (str): The persistable variables file name of the paddle 1.x
``save_inference_model`` save format. No default file name, save variables separately
2020-10-11 23:12:19 +08:00
by default.
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
Returns:
TranslatedLayer: A Layer object can run saved translated model.
Examples:
2020-10-11 23:12:19 +08:00
1. Load model saved by ``paddle.jit.save`` then performing inference and fine-tune training.
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2026-02-20 17:21:20 +08:00
.. code-block:: pycon
2023-07-31 14:44:48 +08:00
:name: code-example1
2023-09-26 11:19:57 +08:00
>>> # doctest: +SKIP( ' `paddle.jit.to_static` can not run in xdoctest ' )
2023-07-31 14:44:48 +08:00
>>> 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
2026-02-20 17:21:20 +08:00
>>> class RandomDataset(paddle.io.Dataset): # type: ignore[type-arg]
2023-07-31 14:44:48 +08:00
... def __init__(self, num_samples):
... self.num_samples = num_samples
...
... def __getitem__(self, idx):
... image = np.random.random([IMAGE_SIZE]).astype( ' float32 ' )
2026-02-20 17:21:20 +08:00
... label = np.random.randint(0, CLASS_NUM - 1, (1,)).astype( ' int64 ' )
2023-07-31 14:44:48 +08:00
... 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()
2026-02-20 17:21:20 +08:00
... print( " Epoch {} batch {} : loss = {} " .format(epoch_id, batch_id, np.mean(loss.numpy())))
2023-07-31 14:44:48 +08:00
>>> # 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,
2026-02-20 17:21:20 +08:00
... num_workers=2,
2023-07-31 14:44:48 +08:00
... )
>>> # train
>>> train(layer, loader, loss_fn, adam)
>>> # save
>>> path = " example_model/linear "
>>> paddle.jit.save(layer, path)
>>> # 2. load model
>>> # load
>>> loaded_layer = paddle.jit.load(path)
>>> # inference
>>> loaded_layer.eval()
>>> x = paddle.randn([1, IMAGE_SIZE], ' float32 ' )
>>> pred = loaded_layer(x)
>>> # fine-tune
>>> loaded_layer.train()
>>> adam = opt.Adam(learning_rate=0.001, parameters=loaded_layer.parameters())
>>> train(loaded_layer, loader, loss_fn, adam)
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2023-09-26 11:19:57 +08:00
2. Load model saved by ``paddle.static.save_inference_model`` then performing and fine-tune training.
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2026-02-20 17:21:20 +08:00
.. code-block:: pycon
2023-07-31 14:44:48 +08:00
:name: code-example2
2023-09-26 11:19:57 +08:00
>>> # doctest: +SOLO( ' can not use multiprocessing testing `DataLoader` ' )
2023-07-31 14:44:48 +08:00
>>> import numpy as np
>>> import paddle
>>> import paddle.static as static
>>> import paddle.nn as nn
>>> import paddle.optimizer as opt
>>> import paddle.nn.functional as F
>>> BATCH_SIZE = 16
>>> BATCH_NUM = 4
>>> EPOCH_NUM = 4
>>> IMAGE_SIZE = 784
>>> CLASS_NUM = 10
>>> # define a random dataset
2026-02-20 17:21:20 +08:00
>>> class RandomDataset(paddle.io.Dataset): # type: ignore[type-arg]
2023-07-31 14:44:48 +08:00
... def __init__(self, num_samples):
... self.num_samples = num_samples
...
... def __getitem__(self, idx):
... image = np.random.random([IMAGE_SIZE]).astype( ' float32 ' )
2026-02-20 17:21:20 +08:00
... label = np.random.randint(0, CLASS_NUM - 1, (1,)).astype( ' int64 ' )
2023-07-31 14:44:48 +08:00
... return image, label
...
... def __len__(self):
... return self.num_samples
>>> paddle.enable_static()
>>> image = static.data(name= ' image ' , shape=[None, 784], dtype= ' float32 ' )
>>> label = static.data(name= ' label ' , shape=[None, 1], dtype= ' int64 ' )
>>> pred = static.nn.fc(x=image, size=10, activation= ' softmax ' )
>>> loss = F.cross_entropy(input=pred, label=label)
>>> avg_loss = paddle.mean(loss)
>>> optimizer = paddle.optimizer.SGD(learning_rate=0.001)
>>> optimizer.minimize(avg_loss)
>>> place = paddle.CPUPlace()
>>> exe = static.Executor(place)
>>> exe.run(static.default_startup_program())
>>> # create data loader
>>> dataset = RandomDataset(BATCH_NUM * BATCH_SIZE)
2026-02-20 17:21:20 +08:00
>>> loader = paddle.io.DataLoader(
... dataset,
2023-07-31 14:44:48 +08:00
... feed_list=[image, label],
... places=place,
... batch_size=BATCH_SIZE,
... shuffle=True,
... drop_last=True,
... return_list=False,
2026-02-20 17:21:20 +08:00
... num_workers=2,
2023-07-31 14:44:48 +08:00
... )
>>> # 1. train and save inference model
>>> for data in loader():
2023-09-26 11:19:57 +08:00
... exe.run(
2023-07-31 14:44:48 +08:00
... static.default_main_program(),
... feed=data,
2026-02-20 17:21:20 +08:00
... fetch_list=[avg_loss],
2023-07-31 14:44:48 +08:00
... )
>>> model_path = " fc.example.model "
2023-09-26 11:19:57 +08:00
>>> paddle.static.save_inference_model(
... model_path,
... [image],
... [pred],
2026-02-20 17:21:20 +08:00
... exe,
2023-09-26 11:19:57 +08:00
... )
2023-07-31 14:44:48 +08:00
>>> # 2. load model
>>> # enable dygraph mode
>>> paddle.disable_static(place)
>>> # load
>>> fc = paddle.jit.load(model_path)
>>> # inference
>>> fc.eval()
>>> x = paddle.randn([1, IMAGE_SIZE], ' float32 ' )
>>> pred = fc(x)
>>> # fine-tune
>>> fc.train()
>>> loss_fn = nn.CrossEntropyLoss()
>>> adam = opt.Adam(learning_rate=0.001, parameters=fc.parameters())
2026-02-20 17:21:20 +08:00
>>> loader = paddle.io.DataLoader(
... dataset,
2023-07-31 14:44:48 +08:00
... places=place,
... batch_size=BATCH_SIZE,
... shuffle=True,
... drop_last=True,
2026-02-20 17:21:20 +08:00
... num_workers=2,
2023-07-31 14:44:48 +08:00
... )
>>> for epoch_id in range(EPOCH_NUM):
... for batch_id, (image, label) in enumerate(loader()):
... out = fc(image)
... loss = loss_fn(out, label)
... loss.backward()
... adam.step()
... adam.clear_grad()
2026-02-20 17:21:20 +08:00
... print( " Epoch {} batch {} : loss = {} " .format(epoch_id, batch_id, np.mean(loss.numpy())))
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
"""
2020-10-11 23:12:19 +08:00
# 1. construct correct config
config = _parse_load_config ( configs )
model_path , config = _build_load_path_and_config ( path , config )
2024-04-25 13:58:01 +08:00
if use_pir_api ( ) :
return PirTranslatedLayer . _construct ( model_path , config )
else :
return TranslatedLayer . _construct ( model_path , config )
[Dy2static] Refactor ProgramTranslator save_inference_model API (#24989)
* experimental refactoring, test=develop
* add TranslatedLayer & remove StaticModelRunner, test=develop
* revert tracedlayer change, test=develop
* fix test_mnist unittest error, test=develop
* add doc & examples, test=develop
* polish doc details, test=develop
* add imperative.jit module, test=develop
* change TranslatedLayer pos, test=develop
* adjust jit module import path, test=develop
* polish doc based review result
* add SaveLoadConfig.separate_params to save paraams separately
* add Layer.buffer support, test=develop
* polish doc details based review result, test=develop
* polish details baesd review comments, test=develop
* add empty str check for param, test=develop
* add unittests, test=develop
* polish details based review comment, test=develop
* remove blanks in comment, test=develop
* polish doc details, test=develop
* update imperative doc link, test=develop
* add api attr for load, test=develop
2020-07-14 15:42:44 +08:00
2024-01-28 00:38:45 +08:00
def set_dynamic_shape ( variable , shape_list ) :
if paddle . base . dygraph . base . in_to_static_mode ( ) :
if isinstance ( variable , paddle . base . framework . Variable ) :
variable . desc . set_shape ( shape_list )
elif isinstance ( variable , paddle . pir . Value ) :
variable . set_shape ( shape_list )
else :
raise TypeError (
" In to_static mode, variable must be a Variable or Value "
)
else :
# in dygraph mode, dynamic shape is not needed, just do nothing.
return
2023-08-07 15:10:33 +08:00
def get_ast_static_function ( function ) :
if isinstance ( function , SymbolicStaticFunction ) :
2024-10-23 23:47:18 +08:00
if function . class_instance :
2023-08-07 15:10:33 +08:00
dygraph_function = types . MethodType (
2024-10-23 23:47:18 +08:00
function . _dygraph_function , function . class_instance
2023-08-07 15:10:33 +08:00
)
else :
dygraph_function = function . _dygraph_function
if function . _function_spec . _input_spec is None :
ast_static_function = ASTStaticFunction (
dygraph_function ,
function . last_call_input_spec ,
* * function . _kwargs ,
)
return ast_static_function
else :
ast_static_function = ASTStaticFunction (
dygraph_function ,
function . _function_spec . _input_spec ,
* * function . _kwargs ,
)
return ast_static_function
return function
2024-10-25 16:33:41 +08:00
def json_to_pdmodel ( net , input_spec , load_path , save_path ) :
net1 = paddle . jit . load ( load_path )
state_dict = { }
for val in net1 . state_dict ( ) . values ( ) :
name = val . name [ : val . name . rfind ( ' _ ' ) ]
state_dict [ name ] = val
name_state_dict = { }
for name , var in net . to_static_state_dict ( ) . items ( ) :
name_state_dict [ name ] = state_dict [ var . name ]
net . set_state_dict ( name_state_dict )
with paddle . pir_utils . OldIrGuard ( ) :
paddle . jit . save ( net , save_path , input_spec )
