diff --git a/dygraph/configs/slim/README.md b/dygraph/configs/slim/README.md old mode 100644 new mode 100755 index 5e523e849ac10eda22c45f89fbdec3f211d2a132..8a5f43a3a909ea95229a5efb415adb253ffc9fa4 --- a/dygraph/configs/slim/README.md +++ b/dygraph/configs/slim/README.md @@ -3,6 +3,7 @@ 在PaddleDetection中, 提供了基于[PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim)进行模型压缩的完整教程和benchmark。目前支持的方法： - [剪裁](prune) +- [量化](quant) 推荐您使用剪裁和蒸馏联合训练，或者使用剪裁和量化，进行检测模型压缩。 下面以YOLOv3为例，进行剪裁、蒸馏和量化实验。 @@ -17,6 +18,17 @@ | YOLOv3-MobileNetV1 | baseline | 24.13 | 93 | 608 | 289.9ms | 75.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/yolov3_mobilenet_v1_270e_voc.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | - | | YOLOv3-MobileNetV1 | 剪裁-l1_norm(sensity) | 15.78(-34.49%) | 66(-29%) | 608 | - | 77.6(+2.5) | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/slim/yolov3_mobilenet_v1_voc_prune_l1_norm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/slim/prune/yolov3_prune_l1_norm.yml) | +### 量化 + +#### COCO上benchmark + +| 模型 | 压缩策略 | 输入尺寸 | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 | +| ------------------ | ------------ | -------- | :---------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: | +| YOLOv3-MobileNetV1 | baseline | 608 | 28.8 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/yolov3_mobilenet_v1_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - | +| YOLOv3-MobileNetV1 | 普通在线量化 | 608 | 27.5 (-1.3) | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/slim/yolov3_mobilenet_v1_coco_qat.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/quant/yolov3_mobilenet_v1_qat.yml) | +| YOLOv3-MobileNetV3 | baseline | 608 | 31.4 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | - | +| YOLOv3-MobileNetV3 | PACT在线量化 | 608 | 29.0 (-2.4) | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/dygraph/slim/yolov3_mobilenet_v3_coco_qat.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/master/dygraph/configs/quant/yolov3_mobilenet_v3_qat.yml) | + - SD855预测时延为使用PaddleLite部署，使用arm8架构并使用4线程(4 Threads)推理时延 ## 实验环境 diff --git a/dygraph/configs/slim/quant/yolov3_mobilenet_v1_qat.yml b/dygraph/configs/slim/quant/yolov3_mobilenet_v1_qat.yml new file mode 100644 index 0000000000000000000000000000000000000000..25cfc37068e88ae9a991249b9d921be36bcffe8e --- /dev/null +++ b/dygraph/configs/slim/quant/yolov3_mobilenet_v1_qat.yml @@ -0,0 +1,12 @@ +# Weights of yolov3_mobilenet_v1_coco +pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/dygraph/yolov3_mobilenet_v1_270e_coco.pdparams +load_static_weights: False +weight_type: resume +slim: QAT + +QAT: + quant_config: { + 'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max', + 'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9, + 'quantizable_layer_type': ['Conv2D', 'Linear']} + print_model: True diff --git a/dygraph/configs/slim/quant/yolov3_mobilenet_v3_qat.yml b/dygraph/configs/slim/quant/yolov3_mobilenet_v3_qat.yml new file mode 100644 index 0000000000000000000000000000000000000000..5ea8fc7aae05f052f9bceab84b74bc4adb76b44d --- /dev/null +++ b/dygraph/configs/slim/quant/yolov3_mobilenet_v3_qat.yml @@ -0,0 +1,13 @@ +# Weights of yolov3_mobilenet_v3_coco +pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/dygraph/yolov3_mobilenet_v3_large_270e_coco.pdparams +load_static_weights: False +weight_type: resume +slim: QAT + +QAT: + quant_config: { + 'weight_preprocess_type': 'PACT', + 'weight_quantize_type': 'channel_wise_abs_max', 'activation_quantize_type': 'moving_average_abs_max', + 'weight_bits': 8, 'activation_bits': 8, 'dtype': 'int8', 'window_size': 10000, 'moving_rate': 0.9, + 'quantizable_layer_type': ['Conv2D', 'Linear']} + print_model: True diff --git a/dygraph/ppdet/slim/__init__.py b/dygraph/ppdet/slim/__init__.py index 6c1a61023c86261ca3748915c06dad1d62695431..7a58bf591c895d699301c0f88dd55268552581e5 100644 --- a/dygraph/ppdet/slim/__init__.py +++ b/dygraph/ppdet/slim/__init__.py @@ -13,5 +13,7 @@ # limitations under the License. from . import prune +from . import quant from .prune import * +from .quant import * diff --git a/dygraph/ppdet/slim/quant.py b/dygraph/ppdet/slim/quant.py new file mode 100644 index 0000000000000000000000000000000000000000..abf123bd199eaf218569d9ea2e5293abe7e49de9 --- /dev/null +++ b/dygraph/ppdet/slim/quant.py @@ -0,0 +1,48 @@ +# Copyright (c) 2021 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. + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import paddle +from paddle.utils import try_import + +from ppdet.core.workspace import register, serializable +from ppdet.utils.logger import setup_logger +logger = setup_logger(__name__) + + +@register +@serializable +class QAT(object): + def __init__(self, quant_config, print_model): + super(QAT, self).__init__() + self.quant_config = quant_config + self.print_model = print_model + + def __call__(self, model): + paddleslim = try_import('paddleslim') + self.quanter = paddleslim.dygraph.quant.QAT(config=self.quant_config) + if self.print_model: + logger.info("Model before quant:") + logger.info(model) + + self.quanter.quantize(model) + + if self.print_model: + logger.info("Quantized model:") + logger.info(model) + + return model