Morph MORPH
SIGN IN SIGN UP
PaddlePaddle / Paddle UNCLAIMED

PArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice （『飞桨』核心框架，深度学习&机器学习高性能单机、分布式训练和跨平台部署）

23798 0 0 C++
Normal View History Raw
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import abc
import copy
from paddle.nn import Layer
from paddle.nn.quant.format import (
ConvertibleQuantedLayer,
LinearQuanterDequanter,
)
from .base_quanter import BaseQuanter
from .config import QuantConfig
[CodeStyle][UP004] remove useless object inheritance (#51771) * add_up004_for_ruff * 修改配置文件并清除object * fix md
2023-03-20 16:10:34 +08:00
class Quantization(metaclass=abc.ABCMeta):
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
r"""
Abstract class used to prepares a copy of the model for quantization calibration or quantization-aware training.
Args:
[Docs] Fix delimiter of argument in docstring (#60942) * fix docstring * update
2024-01-22 14:26:50 +08:00
config(QuantConfig): Quantization configuration
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
"""
def __init__(self, config: QuantConfig):
self._config = copy.deepcopy(config)
@abc.abstractmethod
def quantize(self, model: Layer, inplace=False):
r"""Create a model for quantization-aware training or post-training quantization."""
pass
Solve the problem of scale saving in PTQ (#59441)
2023-12-01 11:41:12 +08:00
def convert(self, model: Layer, inplace=False, remain_weight=False):
[xdoctest][task 124] Reformat example code with google style in python/paddle/quantization/quantize.py (#56235) * [xdoctest][task 124] test=docs_preview * test=document_fix * fix indent --------- Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2023-08-21 12:02:29 +08:00
r"""Convert the quantization model to ONNX style. And the converted
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
model can be saved as inference model by calling paddle.jit.save.
Args:
[Docs] Fix delimiter of argument in docstring (#60942) * fix docstring * update
2024-01-22 14:26:50 +08:00
model(Layer): The quantized model to be converted.
inplace(bool, optional): Whether to modify the model in-place, default is False.
remain_weight(bool, optional): Whether to remain weights in floats, default is False.
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
Return: The converted model
Examples:
[CodeStyle][Xdoctest][204,326-328,333] Fix example code for API `paddle.quantization.PTQ.convert`, `paddle.Model`, `paddle.callbacks.{EarlyStopping,LRScheduler}` and `paddle.distribution.Gumbel` (#77326) --------- Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2026-01-13 17:49:05 +08:00
.. code-block:: pycon
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
[xdoctest][task 124] Reformat example code with google style in python/paddle/quantization/quantize.py (#56235) * [xdoctest][task 124] test=docs_preview * test=document_fix * fix indent --------- Co-authored-by: SigureMo <sigure.qaq@gmail.com>
2023-08-21 12:02:29 +08:00
>>> import paddle
>>> from paddle.quantization import QAT, QuantConfig
>>> from paddle.quantization.quanters import FakeQuanterWithAbsMaxObserver
>>> from paddle.vision.models import LeNet
>>> quanter = FakeQuanterWithAbsMaxObserver(moving_rate=0.9)
>>> q_config = QuantConfig(activation=quanter, weight=quanter)
>>> qat = QAT(q_config)
>>> model = LeNet()
>>> quantized_model = qat.quantize(model)
>>> converted_model = qat.convert(quantized_model)
>>> dummy_data = paddle.rand([1, 1, 32, 32], dtype="float32")
>>> paddle.jit.save(converted_model, "./quant_deploy", [dummy_data])
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
"""
_model = model if inplace else copy.deepcopy(model)
replaced = {}
for name, child in _model.named_children():
quant_dequant = None
if isinstance(child, ConvertibleQuantedLayer):
fix cachekv quant bug (#61357) * fix cachekv quant problem * add unittest
2024-02-21 19:08:52 +08:00
if child.converted:
continue
Fix Quant Bug [avoid inserting quantized op] (#63082)
2024-04-07 11:10:20 +08:00
if hasattr(child, 'weight_quanter') and (
fix cachekv quant bug (#61357) * fix cachekv quant problem * add unittest
2024-02-21 19:08:52 +08:00
child.weight_quanter is None
or child.weight_quanter.scales() is None
):
Solve the problem of scale saving in PTQ (#59441)
2023-12-01 11:41:12 +08:00
continue
child._convert(remain_weight=remain_weight)
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
elif isinstance(child, BaseQuanter):
quant_dequant = LinearQuanterDequanter.from_quanter(child)
else:
Solve the problem of scale saving in PTQ (#59441)
2023-12-01 11:41:12 +08:00
self.convert(child, inplace=True, remain_weight=remain_weight)
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
if quant_dequant is not None:
replaced[name] = quant_dequant
for key, value in replaced.items():
_model._sub_layers[key] = value
return _model
def _convert_to_quant_layers(self, model: Layer, config: QuantConfig):
replaced = {}
for name, child in model.named_children():
Enhance the quantization API with some new features (#50816)
2023-03-08 17:21:51 +08:00
if (
config._is_quantifiable(child)
and type(child) in config.qat_layer_mappings
):
replaced[name] = config._get_qat_layer(child)
else:
self._convert_to_quant_layers(child, config)
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
for key, value in replaced.items():
model._sub_layers[key] = value
def _insert_activation_observers(self, model: Layer, config: QuantConfig):
replaced = {}
for name, child in model.named_children():
if config._need_observe(child):
replaced[name] = config._get_observe_wrapper(child)
else:
Fix Quant Bug [avoid inserting quantized op] (#63082)
2024-04-07 11:10:20 +08:00
if (
type(child) not in config._qat_layer_mapping.values()
and type(child)
not in config._customized_qat_layer_mapping.values()
):
self._insert_activation_observers(child, config)
Add Post-Training Quantization and export function in dygraph mode (#50107) Add PTQ and exporting function 1. Add Post-Training Quantization 1.1 Abstract some functions from QAT to Quantization class 1.2 Add Post-Training Quantization by extending Quantization class 1.3 Add observers for PTQ 1.4 Add unittest for PTQ 2. Add exporting function for QAT and PTQ
2023-02-16 14:57:20 +08:00
for key, value in replaced.items():
model._sub_layers[key] = value
def _details(self):
return self._config.details()
def __str__(self):
return self._details()
def __repr__(self):
return self.__str__()
Reference in New Issue Copy Permalink