From 11377e3c6d7a3ccaa5b119230ba6e83a9f8e7996 Mon Sep 17 00:00:00 2001 From: wanghaoshuang Date: Tue, 25 Feb 2020 20:37:28 +0800 Subject: [PATCH] Update pages. --- _sources/api_cn/nas_api.rst.txt | 4 +- _sources/api_cn/pantheon_api.md.txt | 5 +- _sources/api_cn/prune_api.rst.txt | 2 +- _sources/api_en/index_en.rst.txt | 1 + _sources/api_en/table_latency_en.md.txt | 144 ++ _sources/index_en.rst.txt | 6 +- _sources/install_en.md.txt | 19 +- _sources/intro.md.txt | 66 +- _sources/intro_en.md.txt | 94 +- _sources/model_zoo.md.txt | 119 +- _sources/model_zoo_en.md.txt | 309 +++-- _sources/tutorials/darts_nas_turorial.md.txt | 275 ++++ ...on_sensitivity_analysis_tutorial_en.md.txt | 263 ++++ _sources/tutorials/index.rst.txt | 12 +- _sources/tutorials/index_en.rst.txt | 2 +- ...etection_slim_distillation_tutorial.md.txt | 32 + .../paddledetection_slim_nas_tutorial.md.txt | 8 + ...addledetection_slim_pruing_tutorial.md.txt | 3 + ...edetection_slim_prune_dist_tutorial.md.txt | 7 + ...etection_slim_quantization_tutorial.md.txt | 28 + ...ledetection_slim_sensitivy_tutorial.md.txt | 3 + algo/algo.html | 219 +-- api_cn/analysis_api.html | 4 +- api_cn/index.html | 10 +- api_cn/nas_api.html | 13 +- api_cn/one_shot_api.html | 8 +- api_cn/pantheon_api.html | 126 +- api_cn/prune_api.html | 10 +- api_cn/quantization_api.html | 4 +- api_cn/search_space.html | 4 +- api_cn/single_distiller_api.html | 4 +- api_cn/table_latency.html | 4 +- api_en/index_en.html | 6 +- api_en/paddleslim.analysis.html | 1 + api_en/paddleslim.dist.html | 1 + api_en/paddleslim.nas.html | 1 + api_en/paddleslim.nas.one_shot.html | 1 + api_en/paddleslim.pantheon.html | 7 +- api_en/paddleslim.prune.html | 1 + api_en/paddleslim.quant.html | 1 + api_en/search_space_en.html | 5 +- api_en/table_latency_en.html | 372 ++++++ index.html | 174 ++- index_en.html | 189 ++- install.html | 6 +- install_en.html | 18 +- intro.html | 162 ++- intro_en.html | 196 ++- model_zoo.html | 224 +++- model_zoo_en.html | 1169 +++++++++++++---- objects.inv | Bin 2800 -> 2820 bytes quick_start/distillation_tutorial.html | 14 +- quick_start/distillation_tutorial_en.html | 20 +- quick_start/index.html | 2 +- quick_start/nas_tutorial.html | 86 +- quick_start/nas_tutorial_en.html | 12 +- quick_start/pruning_tutorial.html | 48 +- quick_start/pruning_tutorial_en.html | 44 +- quick_start/quant_aware_tutorial.html | 56 +- quick_start/quant_aware_tutorial_en.html | 10 +- quick_start/quant_post_tutorial.html | 44 +- quick_start/quant_post_tutorial_en.html | 6 +- searchindex.js | 2 +- tutorials/darts_nas_turorial.html | 531 ++++++++ ...ication_sensitivity_analysis_tutorial.html | 15 +- ...tion_sensitivity_analysis_tutorial_en.html | 462 +++++++ tutorials/index.html | 29 +- tutorials/index_en.html | 8 +- ...edetection_slim_distillation_tutorial.html | 366 ++++++ .../paddledetection_slim_nas_tutorial.html | 232 ++++ .../paddledetection_slim_pruing_tutorial.html | 224 ++++ ...dledetection_slim_prune_dist_tutorial.html | 231 ++++ ...edetection_slim_quantization_tutorial.html | 360 +++++ ...ddledetection_slim_sensitivy_tutorial.html | 224 ++++ tutorials/sensitivity_tutorial.html | 10 +- 75 files changed, 6499 insertions(+), 879 deletions(-) create mode 100644 _sources/api_en/table_latency_en.md.txt create mode 100644 _sources/tutorials/darts_nas_turorial.md.txt create mode 100644 _sources/tutorials/image_classification_sensitivity_analysis_tutorial_en.md.txt create mode 100644 _sources/tutorials/paddledetection_slim_distillation_tutorial.md.txt create mode 100644 _sources/tutorials/paddledetection_slim_nas_tutorial.md.txt create mode 100644 _sources/tutorials/paddledetection_slim_pruing_tutorial.md.txt create mode 100644 _sources/tutorials/paddledetection_slim_prune_dist_tutorial.md.txt create mode 100644 _sources/tutorials/paddledetection_slim_quantization_tutorial.md.txt create mode 100644 _sources/tutorials/paddledetection_slim_sensitivy_tutorial.md.txt create mode 100644 api_en/table_latency_en.html create mode 100644 tutorials/darts_nas_turorial.html create mode 100644 tutorials/image_classification_sensitivity_analysis_tutorial_en.html create mode 100644 tutorials/paddledetection_slim_distillation_tutorial.html create mode 100644 tutorials/paddledetection_slim_nas_tutorial.html create mode 100644 tutorials/paddledetection_slim_pruing_tutorial.html create mode 100644 tutorials/paddledetection_slim_prune_dist_tutorial.html create mode 100644 tutorials/paddledetection_slim_quantization_tutorial.html create mode 100644 tutorials/paddledetection_slim_sensitivy_tutorial.html diff --git a/_sources/api_cn/nas_api.rst.txt b/_sources/api_cn/nas_api.rst.txt index f1f0214d..d5ddaa58 100644 --- a/_sources/api_cn/nas_api.rst.txt +++ b/_sources/api_cn/nas_api.rst.txt @@ -128,7 +128,7 @@ SANAS（Simulated Annealing Neural Architecture Search）是基于模拟退火 - **tokens(list):** - 一组tokens。tokens的长度和范围取决于搜索空间。 **返回：** - 根据传入的token得到一个模型结构实例。 + 根据传入的token得到一个模型结构实例列表。 **示例代码：** @@ -153,8 +153,10 @@ SANAS（Simulated Annealing Neural Architecture Search）是基于模拟退火 **示例代码：** .. code-block:: python + import paddle.fluid as fluid from paddleslim.nas import SANAS config = [('MobileNetV2Space')] sanas = SANAS(configs=config) print(sanas.current_info()) + diff --git a/_sources/api_cn/pantheon_api.md.txt b/_sources/api_cn/pantheon_api.md.txt index 1405faea..1de04a64 100644 --- a/_sources/api_cn/pantheon_api.md.txt +++ b/_sources/api_cn/pantheon_api.md.txt @@ -1,4 +1,4 @@ -# 多进程蒸馏 +# 大规模可扩展知识蒸馏框架 Pantheon ## Teacher @@ -100,7 +100,8 @@ pantheon.Teacher.start\_knowledge\_service(feed\_list, schema, program, reader\_ - **times (int):** The maximum repeated serving times, default 1. Whenever the public method **get\_knowledge\_generator()** in **Student** object called once, the serving times will be added one, - until reaching the maximum and ending the service. + until reaching the maximum and ending the service. Only + valid in online mode, and will be ignored in offline mode. **Return:** None diff --git a/_sources/api_cn/prune_api.rst.txt b/_sources/api_cn/prune_api.rst.txt index d38480d6..38a76acd 100644 --- a/_sources/api_cn/prune_api.rst.txt +++ b/_sources/api_cn/prune_api.rst.txt @@ -378,7 +378,7 @@ load_sensitivities } } sensitivities_file = "sensitive_api_demo.data" - with open(sensitivities_file, 'w') as f: + with open(sensitivities_file, 'wb') as f: pickle.dump(sen, f) sensitivities = load_sensitivities(sensitivities_file) print(sensitivities) diff --git a/_sources/api_en/index_en.rst.txt b/_sources/api_en/index_en.rst.txt index 9fd0325a..2f21132c 100644 --- a/_sources/api_en/index_en.rst.txt +++ b/_sources/api_en/index_en.rst.txt @@ -17,3 +17,4 @@ API Documents paddleslim.nas.one_shot.rst paddleslim.pantheon.rst search_space_en.rst + table_latency_en.md diff --git a/_sources/api_en/table_latency_en.md.txt b/_sources/api_en/table_latency_en.md.txt new file mode 100644 index 00000000..6a6c6ac6 --- /dev/null +++ b/_sources/api_en/table_latency_en.md.txt @@ -0,0 +1,144 @@ +# Table about hardware lantency + +The table about hardware latency is used to evaluate the inference time in special environment and inference engine. The following text used to introduce the format that PaddleSlim support. + +## Introduce + +The table about hardware latency saved all possible operations, one operation in the table including type and parameters, such as: type can be `conv2d`, and corresponding parameters can be the size of feature map, number of kernel, and the size of kernel. +The latency of every operation depends on hardware and inference engine. + +## Overview format +The table about hardware latency saved in the way of file or multi-line string. +The first line of the table about hardware latency saved the information about version, every line in the following represents a operation and its latency. + +## Version + +The information about version split by comma in the english format, and the detail is hardware, inference engine and timestamp. + +- ** hardware: ** Used to mark the environment of hardware, including type of architecture, version and so on. + +- ** inference engine: ** Used to mark inference engine, including the name of inference engine, version, optimize options and so on. + +- ** timestamp: ** Used to mark the time of this table created. + +## Operation + +The information about operation split by comma in the english format, the information about operation and latency split by tabs. + +### conv2d + +**format** + +```text +op_type,flag_bias,flag_relu,n_in,c_in,h_in,w_in,c_out,groups,kernel,padding,stride,dilation\tlatency +``` + +**introduce** + +- **op_type(str)** - The type of this op. +- **flag_bias (int)** - Whether has bias or not(0: donot has bias, 1: has bias). +- **flag_relu (int)** - Whether has relu or not(0: donot has relu, 1: has relu). +- **n_in (int)** - The batch size of input. +- **c_in (int)** - The number of channel about input. +- **h_in (int)** - The height of input feature map. +- **w_in (int)** - The width of input feature map. +- **c_out (int)** - The number of channel about output. +- **groups (int)** - The group of conv2d. +- **kernel (int)** - The size of kernel. +- **padding (int)** - The size of padding. +- **stride (int)** - The size of stride. +- **dilation (int)** - The size of dilation. +- **latency (float)** - The latency of this op. + +### activaiton + +**format** + +```text +op_type,n_in,c_in,h_in,w_in\tlatency +``` + +**introduce** + +- **op_type(str)** - The type of this op. +- **n_in (int)** - The batch size of input. +- **c_in (int)** - The number of channel about input. +- **h_in (int)** - The height of input feature map. +- **w_in (int)** - The width of input feature map. +- **latency (float)** - The latency of this op. + +### batch_norm + +**format** + +```text +op_type,active_type,n_in,c_in,h_in,w_in\tlatency +``` + +**introduce** + +- **op_type(str)** - The type of this op. +- **active_type (string|None)** - The type of activation function, including relu, prelu, sigmoid, relu6, tanh. +- **n_in (int)** - The batch size of input. +- **c_in (int)** - The number of channel about input. +- **h_in (int)** - The height of input feature map. +- **w_in (int)** - The width of input feature map. +- **latency (float)** - The latency of this op. + +### eltwise + +**format** + +```text +op_type,n_in,c_in,h_in,w_in\tlatency +``` + +**introduce** + +- **op_type(str)** - The type of this op. +- **n_in (int)** - The batch size of input. +- **c_in (int)** - The number of channel about input. +- **h_in (int)** - The height of input feature map. +- **w_in (int)** - The width of input feature map. +- **latency (float)** - The latency of this op. + +### pooling + +**format** + +```text +op_type,flag_global_pooling,n_in,c_in,h_in,w_in,kernel,padding,stride,ceil_mode,pool_type\tlatency +``` + +**introduce** + +- **op_type(str)** - The type of this op. +- **flag_global_pooling (int)** - Whether is global pooling or not(0: is not global, 1: is global pooling). +- **n_in (int)** - The batch size of input. +- **c_in (int)** - The number of channel about input. +- **h_in (int)** - The height of input feature map. +- **w_in (int)** - The width of input feature map. +- **kernel (int)** - The size of kernel. +- **padding (int)** - The size of padding. +- **stride (int)** - The size of stride. +- **ceil_mode (int)** - Whether to compute height and width by using ceil function(0: use floor function, 1: use ceil function). +- **pool_type (int)** - The type of pooling(1: max pooling 2: average pooling including padding 3: average pooling excluding padding). +- **latency (float)** - The latency of this op. + +### softmax + +**format** + +```text +op_type,axis,n_in,c_in,h_in,w_in\tlatency +``` + +**introduce** + +- **op_type(str)** - The type of this op. +- **axis (int)** - The index to compute softmax, index in the range of [-1, rank-1], `rank` is the rank of input. +- **n_in (int)** - The batch size of input. +- **c_in (int)** - The number of channel about input. +- **h_in (int)** - The height of input feature map. +- **w_in (int)** - The width of input feature map. +- **latency (float)** - The latency of this op. diff --git a/_sources/index_en.rst.txt b/_sources/index_en.rst.txt index ec250402..32682529 100644 --- a/_sources/index_en.rst.txt +++ b/_sources/index_en.rst.txt @@ -3,8 +3,8 @@ You can adapt this file completely to your liking, but it should at least contain the root `toctree` directive. -Welcome to use PaddleSlim. -======== +Index +============== .. toctree:: :maxdepth: 1 @@ -16,3 +16,5 @@ Welcome to use PaddleSlim. tutorials/index_en api_en/index_en model_zoo_en.md + +.. mdinclude:: intro_en.md diff --git a/_sources/install_en.md.txt b/_sources/install_en.md.txt index 69d684a4..9d0eedba 100644 --- a/_sources/install_en.md.txt +++ b/_sources/install_en.md.txt @@ -1,23 +1,22 @@ # Install -安装PaddleSlim前，请确认已正确安装Paddle1.6版本或更新版本。Paddle安装请参考：[Paddle安装教程](https://www.paddlepaddle.org.cn/install/quick)。 +Please ensure you have installed PaddlePaddle1.7+. [How to install PaddlePaddle](https://www.paddlepaddle.org.cn/install/quick)。 -- 安装develop版本 - +- Install by pip ```bash -git clone https://github.com/PaddlePaddle/PaddleSlim.git -cd PaddleSlim -python setup.py install +pip install paddleslim -i https://pypi.org/simple ``` -- 安装官方发布的最新版本 +- Install from source ```bash -pip install paddleslim -i https://pypi.org/simple +git clone https://github.com/PaddlePaddle/PaddleSlim.git +cd PaddleSlim +python setup.py install ``` -- 安装历史版本 +- History packages -请点击[pypi.org](https://pypi.org/project/paddleslim/#history)查看可安装历史版本。 +History packages is available in [pypi.org](https://pypi.org/project/paddleslim/#history). diff --git a/_sources/intro.md.txt b/_sources/intro.md.txt index 572c1690..2b2ec1b3 100644 --- a/_sources/intro.md.txt +++ b/_sources/intro.md.txt @@ -1,27 +1,75 @@ +# 介绍 +PaddleSlim是一个模型压缩工具库，包含模型剪裁、定点量化、知识蒸馏、超参搜索和模型结构搜索等一系列模型压缩策略。 -# PaddleSlim简介 +对于业务用户，PaddleSlim提供完整的模型压缩解决方案，可用于图像分类、检测、分割等各种类型的视觉场景。 +同时也在持续探索NLP领域模型的压缩方案。另外，PaddleSlim提供且在不断完善各种压缩策略在经典开源任务的benchmark, +以便业务用户参考。 + +对于模型压缩算法研究者或开发者，PaddleSlim提供各种压缩策略的底层辅助接口，方便用户复现、调研和使用最新论文方法。 +PaddleSlim会从底层能力、技术咨询合作和业务场景等角度支持开发者进行模型压缩策略相关的创新工作。 -PaddleSlim是PaddlePaddle框架的一个子模块，主要用于压缩图像领域模型。在PaddleSlim中，不仅实现了目前主流的网络剪枝、量化、蒸馏三种压缩策略，还实现了超参数搜索和小模型网络结构搜索功能。在后续版本中，会添加更多的压缩策略，以及完善对NLP领域模型的支持。 ## 功能 - 模型剪裁 - - 支持通道均匀模型剪裁（uniform pruning) - - 基于敏感度的模型剪裁 - - 基于进化算法的自动模型剪裁三种方式 + - 卷积通道均匀剪裁 + - 基于敏感度的卷积通道剪裁 + - 基于进化算法的自动剪裁 -- 量化训练 +- 定点量化 - 在线量化训练（training aware） - 离线量化（post training） - - 支持对权重全局量化和Channel-Wise量化 - 知识蒸馏 - 支持单进程知识蒸馏 - 支持多进程分布式知识蒸馏 - 神经网络结构自动搜索（NAS） - - 支持One-Shot网络结构自动搜索（Ont-Shot-NAS） - - 支持基于进化算法的轻量神经网络结构自动搜索（Light-NAS） + - 支持基于进化算法的轻量神经网络结构自动搜索 + - 支持One-Shot网络结构自动搜索 - 支持 FLOPS / 硬件延时约束 - 支持多平台模型延时评估 + - 支持用户自定义搜索算法和搜索空间 + + +## 部分压缩策略效果 + +### 分类模型 + +数据: ImageNet2012; 模型: MobileNetV1; + +|压缩策略 |精度收益(baseline: 70.91%) |模型大小(baseline: 17.0M)| +|:---:|:---:|:---:| +| 知识蒸馏(ResNet50)| **+1.06%** | | +| 知识蒸馏(ResNet50) + int8量化训练 |**+1.10%**| **-71.76%**| +| 剪裁(FLOPs-50%) + int8量化训练|**-1.71%**|**-86.47%**| + + +### 图像检测模型 + +#### 数据：Pascal VOC；模型：MobileNet-V1-YOLOv3 + +| 压缩方法 | mAP(baseline: 76.2%) | 模型大小(baseline: 94MB) | +| :---------------------: | :------------: | :------------:| +| 知识蒸馏(ResNet34-YOLOv3) | **+2.8%** | | +| 剪裁 FLOPs -52.88% | **+1.4%** | **-67.76%** | +|知识蒸馏(ResNet34-YOLOv3)+剪裁(FLOPs-69.57%)| **+2.6%**|**-67.00%**| + + +#### 数据：COCO；模型：MobileNet-V1-YOLOv3 + +| 压缩方法 | mAP(baseline: 29.3%) | 模型大小| +| :---------------------: | :------------: | :------:| +| 知识蒸馏(ResNet34-YOLOv3) | **+2.1%** | | +| 知识蒸馏(ResNet34-YOLOv3)+剪裁(FLOPs-67.56%) | **-0.3%** | **-66.90%**| + +### 搜索 + +数据：ImageNet2012; 模型：MobileNetV2 + +|硬件环境 | 推理耗时 | Top1准确率(baseline:71.90%) | +|:---------------:|:---------:|:--------------------:| +| RK3288 | **-23%** | +0.07% | +| Android cellphone | **-20%** | +0.16% | +| iPhone 6s | **-17%** | +0.32% | diff --git a/_sources/intro_en.md.txt b/_sources/intro_en.md.txt index b0b64cba..23737106 100644 --- a/_sources/intro_en.md.txt +++ b/_sources/intro_en.md.txt @@ -1,28 +1,84 @@ +# Introduction +PaddleSlim is a toolkit for model compression. It contains a collection of compression strategies, such as pruning, fixed point quantization, knowledge distillation, hyperparameter searching and neural architecture search. -# Introduction +PaddleSlim provides solutions of compression on computer vision models, such as image classification, object detection and semantic segmentation. Meanwhile, PaddleSlim Keeps exploring advanced compression strategies for language model. Furthermore, benckmark of compression strategies on some open tasks is available for your reference. + +PaddleSlim also provides auxiliary and primitive API for developer and researcher to survey, implement and apply the method in latest papers. PaddleSlim will support developer in ability of framework and technology consulting. + +## Features + +### Pruning + + - Uniform pruning of convolution + - Sensitivity-based prunning + - Automated pruning based evolution search strategy + - Support pruning of various deep architectures such as VGG, ResNet, and MobileNet. + - Support self-defined range of pruning, i.e., layers to be pruned. + +### Fixed Point Quantization + + - **Training aware** + - Dynamic strategy: During inference, we quantize models with hyperparameters dynamically estimated from small batches of samples. + - Static strategy: During inference, we quantize models with the same hyperparameters estimated from training data. + - Support layer-wise and channel-wise quantization. + - **Post training** + +### Knowledge Distillation + + - **Naive knowledge distillation:** transfers dark knowledge by merging the teacher and student model into the same Program + - **Paddle large-scale scalable knowledge distillation framework Pantheon:** a universal solution for knowledge distillation, more flexible than the naive knowledge distillation, and easier to scale to the large-scale applications. + + - Decouple the teacher and student models --- they run in different processes in the same or different nodes, and transfer knowledge via TCP/IP ports or local files; + - Friendly to assemble multiple teacher models and each of them can work in either online or offline mode independently; + - Merge knowledge from different teachers and make batch data for the student model automatically; + - Support the large-scale knowledge prediction of teacher models on multiple devices. + +### Neural Architecture Search + + - Neural architecture search based on evolution strategy. + - Support distributed search. + - One-Shot neural architecture search. + - Support FLOPs and latency constrained search. + - Support the latency estimation on different hardware and platforms. + +## Performance + +### Image Classification + +Dataset: ImageNet2012; Model: MobileNetV1; + +|Method |Accuracy(baseline: 70.91%) |Model Size(baseline: 17.0M)| +|:---:|:---:|:---:| +| Knowledge Distillation(ResNet50)| **+1.06%** | | +| Knowledge Distillation(ResNet50) + int8 quantization |**+1.10%**| **-71.76%**| +| Pruning(FLOPs-50%) + int8 quantization|**-1.71%**|**-86.47%**| + + +### Object Detection + +#### Dataset: Pascal VOC; Model: MobileNet-V1-YOLOv3 -As a submodule of PaddlePaddle framework, PaddleSlim is an open-source library for deep model compression and architecture search. PaddleSlim supports current popular deep compression techniques such as pruning, quantization, and knowledge distillation. Further, it also automates the search of hyperparameters and the design of lightweight deep architectures. In the future, we will develop more practically useful compression techniques for industrial-level applications and transfer these techniques to models in NLP. +| Method | mAP(baseline: 76.2%) | Model Size(baseline: 94MB) | +| :---------------------: | :------------: | :------------:| +| Knowledge Distillation(ResNet34-YOLOv3) | **+2.8%** | | +| Pruning(FLOPs -52.88%) | **+1.4%** | **-67.76%** | +|Knowledge DistillationResNet34-YOLOv3)+Pruning(FLOPs-69.57%)| **+2.6%**|**-67.00%**| -## Methods +#### Dataset: COCO; Model: MobileNet-V1-YOLOv3 -- Pruning - - Uniform pruning - - Sensitivity-based pruning - - Automated model pruning +| Method | mAP(baseline: 29.3%) | Model Size| +| :---------------------: | :------------: | :------:| +| Knowledge Distillation(ResNet34-YOLOv3) | **+2.1%** |-| +| Knowledge Distillation(ResNet34-YOLOv3)+Pruning(FLOPs-67.56%) | **-0.3%** | **-66.90%**| -- Quantization - - Training-aware quantization: Quantize models with hyperparameters dynamically estimated from small batches of samples. - - Training-aware quantization: Quantize models with the same hyperparameters estimated from training data. - - Support global quantization of weights and Channel-Wise quantization +### NAS -- Knowledge Distillation - - Single-process knowledge distillation - - Multi-process distributed knowledge distillation +Dataset: ImageNet2012; Model: MobileNetV2 -- Network Architecture Search（NAS） - - Simulated Annealing (SA)-based lightweight network architecture search method.（Light-NAS） - - One-Shot network structure automatic search. (One-Shot-NAS) - - PaddleSlim supports FLOPs and latency constrained search. - - PaddleSlim supports the latency estimation on different hardware and platforms. +|Device | Infer time cost | Top1 accuracy(baseline:71.90%) | +|:---------------:|:---------:|:--------------------:| +| RK3288 | **-23%** | +0.07% | +| Android cellphone | **-20%** | +0.16% | +| iPhone 6s | **-17%** | +0.32% | diff --git a/_sources/model_zoo.md.txt b/_sources/model_zoo.md.txt index c8126cb7..27e859ef 100644 --- a/_sources/model_zoo.md.txt +++ b/_sources/model_zoo.md.txt @@ -1,6 +1,6 @@ # 模型库 -## 1. 图象分类 +## 1. 图像分类 数据集：ImageNet1000类 @@ -16,7 +16,7 @@ | MobileNetV2 | quant_aware |72.05%/90.63% (-0.1%/-0.02%)| 4.0 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_quant_aware.tar) | |ResNet50|-|76.50%/93.00%| 99 | 2.71 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar) | |ResNet50|quant_post|76.33%/93.02% (-0.17%/+0.02%)| 25.1| 1.19 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_post.tar) | -|ResNet50|quant_aware| 76.48%/93.11% (-0.02%/+0.11%)| 25.1 | 1.17 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_awre.tar) | +|ResNet50|quant_aware| 76.48%/93.11% (-0.02%/+0.11%)| 25.1 | 1.17 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_awre.tar) | 分类模型Lite时延(ms) @@ -57,20 +57,27 @@ ### 1.2 剪裁 -| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | GFLOPs | 下载 | -|:--:|:---:|:--:|:--:|:--:|:--:| -| MobileNetV1 | Baseline | 70.99%/89.68% | 17 | 1.11 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) | -| MobileNetV1 | uniform -50% | 69.4%/88.66% (-1.59%/-1.02%) | 9 | 0.56 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_uniform-50.tar) | -| MobileNetV1 | sensitive -30% | 70.4%/89.3% (-0.59%/-0.38%) | 12 | 0.74 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-30.tar) | -| MobileNetV1 | sensitive -50% | 69.8% / 88.9% (-1.19%/-0.78%) | 9 | 0.56 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-50.tar) | -| MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) | -| MobileNetV2 | uniform -50% | 65.79%/86.11% (-6.35%/-4.47%) | 11 | 0.296 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_uniform-50.tar) | -| ResNet34 | - | 72.15%/90.65% | 84 | 7.36 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet34_pretrained.tar) | -| ResNet34 | uniform -50% | 70.99%/89.95% (-1.36%/-0.87%) | 41 | 3.67 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_uniform-50.tar) | -| ResNet34 | auto -55.05% | 70.24%/89.63% (-2.04%/-1.06%) | 33 | 3.31 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_auto-55.tar) | +PaddleLite推理耗时说明： + +环境：Qualcomm SnapDragon 845 + armv8 + +速度指标：Thread1/Thread2/Thread4耗时 +PaddleLite版本： v2.3 +| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | GFLOPs |PaddleLite推理耗时|TensorRT推理速度(FPS)| 下载 | +|:--:|:---:|:--:|:--:|:--:|:--:|:--:|:--:| +| MobileNetV1 | Baseline | 70.99%/89.68% | 17 | 1.11 |66.052\35.8014\19.5762|-| [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) | +| MobileNetV1 | uniform -50% | 69.4%/88.66% (-1.59%/-1.02%) | 9 | 0.56 | 33.5636\18.6834\10.5076|-|[下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_uniform-50.tar) | +| MobileNetV1 | sensitive -30% | 70.4%/89.3% (-0.59%/-0.38%) | 12 | 0.74 | 46.5958\25.3098\13.6982|-|[下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-30.tar) | +| MobileNetV1 | sensitive -50% | 69.8% / 88.9% (-1.19%/-0.78%) | 9 | 0.56 |37.9892\20.7882\11.3144|-| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-50.tar) | +| MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 |41.7874\23.375\13.3998|-| [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) | +| MobileNetV2 | uniform -50% | 65.79%/86.11% (-6.35%/-4.47%) | 11 | 0.296 |23.8842\13.8698\8.5572|-| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_uniform-50.tar) | +| ResNet34 | - | 72.15%/90.65% | 84 | 7.36 |217.808\139.943\96.7504|342.32| [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet34_pretrained.tar) | +| ResNet34 | uniform -50% | 70.99%/89.95% (-1.36%/-0.87%) | 41 | 3.67 |114.787\75.0332\51.8438|452.41| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_uniform-50.tar) | +| ResNet34 | auto -55.05% | 70.24%/89.63% (-2.04%/-1.06%) | 33 | 3.31 |105.924\69.3222\48.0246|457.25| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_auto-55.tar) | + ### 1.3 蒸馏 @@ -85,10 +92,25 @@ |ResNet101|teacher|77.56%/93.64%| 173 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar) | | ResNet50 | ResNet101 distill | 77.29%/93.65% (+0.79%/+0.65%) | 99 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_distilled.tar) | -!!! note "Note" +注意：带"_vd"后缀代表该预训练模型使用了Mixup，Mixup相关介绍参考[mixup: Beyond Empirical Risk Minimization](https://arxiv.org/abs/1710.09412) - [1]：带_vd后缀代表该预训练模型使用了Mixup，Mixup相关介绍参考[mixup: Beyond Empirical Risk Minimization](https://arxiv.org/abs/1710.09412) +### 1.4 搜索 +数据集: ImageNet1000 + +| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | GFLOPs | 下载 | +|:--:|:---:|:--:|:--:|:--:|:--:| +| MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) | +| MobileNetV2 | SANAS | 71.518%/90.208% (-0.632%/-0.442%) | 14 | 0.295 | [下载链接](https://paddlemodels.cdn.bcebos.com/PaddleSlim/MobileNetV2_sanas.tar) | + +数据集: Cifar10 + +| 模型 |压缩方法 | Acc | 模型参数（MB） | 下载 | +|:---:|:--:|:--:|:--:|:--:| +| Darts | - | 97.135% | 3.767 | - | +| Darts_SA(基于Darts搜索空间) | SANAS | 97.276%(+0.141%) | 3.344(-11.2%) | - | + +Note: MobileNetV2_NAS 的token是：[4, 4, 5, 1, 1, 2, 1, 1, 0, 2, 6, 2, 0, 3, 4, 5, 0, 4, 5, 5, 1, 4, 8, 0, 0]. Darts_SA的token是：[5, 5, 0, 5, 5, 10, 7, 7, 5, 7, 7, 11, 10, 12, 10, 0, 5, 3, 10, 8]. ## 2. 目标检测 @@ -99,8 +121,8 @@ | 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积（MB） | TensorRT时延(V100, ms) | 下载 | | :----------------------------: | :---------: | :----: | :-------: | :------------: | :------------: | :------------: | :------------: | :----------: |:----------: | | MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.1 | 95 | - | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | -| MobileNet-V1-YOLOv3 | quant_post | COCO | 8 | 27.9 (-1.4)| 28.0 (-1.3) | 26.0 (-1.0) | 25 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | -| MobileNet-V1-YOLOv3 | quant_aware | COCO | 8 | 28.1 (-1.2)| 28.2 (-1.1) | 25.8 (-1.2) | 26.3 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_coco_quant_aware.tar) | +| MobileNet-V1-YOLOv3 | quant_post | COCO | 8 | 27.9 (-1.4)| 28.0 (-1.3) | 26.0 (-1.0) | 25 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | +| MobileNet-V1-YOLOv3 | quant_aware | COCO | 8 | 28.1 (-1.2)| 28.2 (-1.1) | 25.8 (-1.2) | 26.3 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_coco_quant_aware.tar) | | R34-YOLOv3 | - | COCO | 8 | 36.2 | 34.3 | 31.4 | 162 | - | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) | | R34-YOLOv3 | quant_post | COCO | 8 | 35.7 (-0.5) | - | - | 42.7 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_post.tar) | | R34-YOLOv3 | quant_aware | COCO | 8 | 35.2 (-1.0) | 33.3 (-1.0) | 30.3 (-1.1)| 44 | - | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | @@ -127,20 +149,29 @@ ### 2.2 剪裁 + 数据集：Pasacl VOC & COCO 2017 -| 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积(MB) | GFLOPs (608*608) | 下载 | -| :----------------------------: | :---------------: | :--------: | :-------: | :------------: | :------------: | :------------: | :----------: | :--------------: | :----------------------------------------------------------: | -| MobileNet-V1-YOLOv3 | Baseline | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | 40.49 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | -| MobileNet-V1-YOLOv3 | sensitive -52.88% | Pascal VOC | 8 | 77.6 (+1.4) | 77.7 (1.0) | 75.5 (+0.2) | 31 | 19.08 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_voc_prune.tar) | -| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | 41.35 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | -| MobileNet-V1-YOLOv3 | sensitive -51.77% | COCO | 8 | 26.0 (-3.3) | 25.1 (-4.2) | 22.6 (-4.4) | 32 | 19.94 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_prune.tar) | -| R50-dcn-YOLOv3 | - | COCO | 8 | 39.1 | - | - | 177 | 89.60 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn.tar) | -| R50-dcn-YOLOv3 | sensitive -9.37% | COCO | 8 | 39.3 (+0.2) | - | - | 150 | 81.20 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune.tar) | -| R50-dcn-YOLOv3 | sensitive -24.68% | COCO | 8 | 37.3 (-1.8) | - | - | 113 | 67.48 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune578.tar) | -| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | - | - | 177 | 89.60 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn_obj365_pretrained_coco.tar) | -| R50-dcn-YOLOv3 obj365_pretrain | sensitive -9.37% | COCO | 8 | 40.5 (-0.9) | - | - | 150 | 81.20 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune.tar) | -| R50-dcn-YOLOv3 obj365_pretrain | sensitive -24.68% | COCO | 8 | 37.8 (-3.3) | - | - | 113 | 67.48 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune578.tar) | +PaddleLite推理耗时说明： + +环境：Qualcomm SnapDragon 845 + armv8 + +速度指标：Thread1/Thread2/Thread4耗时 + +PaddleLite版本： v2.3 + +| 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积(MB) | GFLOPs (608*608) | PaddleLite推理耗时(ms)(608*608) | TensorRT推理速度(FPS)(608*608) | 下载 | +| :----------------------------: | :---------------: | :--------: | :-------: | :------------: | :------------: | :------------: | :----------: | :--------------: | :--------------: | :--------------: | :-----------------------------------: | +| MobileNet-V1-YOLOv3 | Baseline | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | 40.49 | 1238\796.943\520.101|60.04| [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNet-V1-YOLOv3 | sensitive -52.88% | Pascal VOC | 8 | 77.6 (+1.4) | 77.7 (1.0) | 75.5 (+0.2) | 31 | 19.08 | 602.497\353.759\222.427 |99.36| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_voc_prune.tar) | +| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | 41.35 |-|-| [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNet-V1-YOLOv3 | sensitive -51.77% | COCO | 8 | 26.0 (-3.3) | 25.1 (-4.2) | 22.6 (-4.4) | 32 | 19.94 |-|73.93| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_prune.tar) | +| R50-dcn-YOLOv3 | - | COCO | 8 | 39.1 | - | - | 177 | 89.60 |-|27.68| [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn.tar) | +| R50-dcn-YOLOv3 | sensitive -9.37% | COCO | 8 | 39.3 (+0.2) | - | - | 150 | 81.20 |-|30.08| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune.tar) | +| R50-dcn-YOLOv3 | sensitive -24.68% | COCO | 8 | 37.3 (-1.8) | - | - | 113 | 67.48 |-|34.32| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune578.tar) | +| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | - | - | 177 | 89.60 |-|-| [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn_obj365_pretrained_coco.tar) | +| R50-dcn-YOLOv3 obj365_pretrain | sensitive -9.37% | COCO | 8 | 40.5 (-0.9) | - | - | 150 | 81.20 |-|-| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune.tar) | +| R50-dcn-YOLOv3 obj365_pretrain | sensitive -24.68% | COCO | 8 | 37.8 (-3.3) | - | - | 113 | 67.48 |-|-| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune578.tar) | ### 2.3 蒸馏 @@ -157,6 +188,18 @@ | MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | COCO | 8 | 31.4 (+2.1) | 30.0 (+0.7) | 27.1 (+0.1) | 95 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | +### 2.4 搜索 + +数据集：WIDER-FACE + +| 模型 | 压缩方法 | Image/GPU | 输入尺寸 | Easy/Medium/Hard | 模型体积（KB） | 硬件延时（ms）| 下载 | +| :------------: | :---------: | :-------: | :------: | :-----------------------------: | :------------: | :------------: | :----------------------------------------------------------: | +| BlazeFace | - | 8 | 640 | 91.5/89.2/79.7 | 815 | 71.862 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_original.tar) | +| BlazeFace-NAS | - | 8 | 640 | 83.7/80.7/65.8 | 244 | 21.117 |[下载链接](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas.tar) | +| BlazeFace-NASV2 | SANAS | 8 | 640 | 87.0/83.7/68.5 | 389 | 22.558 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas2.tar) | + +Note: 硬件延时时间是利用提供的硬件延时表得到的，硬件延时表是在855芯片上基于PaddleLite测试的结果。BlazeFace-NASV2的详细配置在[这里](https://github.com/PaddlePaddle/PaddleDetection/blob/master/configs/face_detection/blazeface_nas_v2.yml). + ## 3. 图像分割 数据集：Cityscapes @@ -201,8 +244,16 @@ ### 3.2 剪裁 -| 模型 | 压缩方法 | mIoU | 模型体积（MB） | GFLOPs | 下载 | -| :-------: | :---------------: | :-----------: | :------------: | :----: | :----------------------------------------------------------: | -| fast-scnn | baseline | 69.64 | 11 | 14.41 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape.tar) | -| fast-scnn | uniform -17.07% | 69.58 (-0.06) | 8.5 | 11.95 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_uniform-17.tar) | -| fast-scnn | sensitive -47.60% | 66.68 (-2.96) | 5.7 | 7.55 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_sensitive-47.tar) | +PaddleLite推理耗时说明： + +环境：Qualcomm SnapDragon 845 + armv8 + +速度指标：Thread1/Thread2/Thread4耗时 + +PaddleLite版本： v2.3 + +| 模型 | 压缩方法 | mIoU | 模型体积（MB） | GFLOPs | PaddleLite推理耗时 | TensorRT推理速度(FPS) | 下载 | +| :-------: | :---------------: | :-----------: | :------------: | :----: | :------------: | :----: | :--------------------------------------: | +| fast-scnn | baseline | 69.64 | 11 | 14.41 | 1226.36\682.96\415.664 |39.53| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape.tar) | +| fast-scnn | uniform -17.07% | 69.58 (-0.06) | 8.5 | 11.95 | 1140.37\656.612\415.888 |42.01| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_uniform-17.tar) | +| fast-scnn | sensitive -47.60% | 66.68 (-2.96) | 5.7 | 7.55 | 866.693\494.467\291.748 |51.48| [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_sensitive-47.tar) | diff --git a/_sources/model_zoo_en.md.txt b/_sources/model_zoo_en.md.txt index 52d7806b..ba0bc909 100644 --- a/_sources/model_zoo_en.md.txt +++ b/_sources/model_zoo_en.md.txt @@ -1,145 +1,250 @@ # Model Zoo -## 1. 图象分类 +## 1. Image Classification -数据集：ImageNet1000类 +Dataset：ImageNet1000 -### 1.1 量化 +### 1.1 Quantization -| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | 下载 | +| Model | Method | Top-1/Top-5 Acc | Model Size（MB） | TensorRT latency(V100, ms) | Download | +|:--:|:---:|:--:|:--:|:--:|:--:| +|MobileNetV1|-|70.99%/89.68%| 17 | -| [model](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) | +|MobileNetV1|quant_post|70.18%/89.25% (-0.81%/-0.43%)| 4.4 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_quant_post.tar) | +|MobileNetV1|quant_aware|70.60%/89.57% (-0.39%/-0.11%)| 4.4 | -| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_quant_aware.tar) | +| MobileNetV2 | - |72.15%/90.65%| 15 | - | [model](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) | +| MobileNetV2 | quant_post | 71.15%/90.11% (-1%/-0.54%)| 4.0 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_quant_post.tar) | +| MobileNetV2 | quant_aware |72.05%/90.63% (-0.1%/-0.02%)| 4.0 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_quant_aware.tar) | +|ResNet50|-|76.50%/93.00%| 99 | 2.71 | [model](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar) | +|ResNet50|quant_post|76.33%/93.02% (-0.17%/+0.02%)| 25.1| 1.19 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_post.tar) | +|ResNet50|quant_aware| 76.48%/93.11% (-0.02%/+0.11%)| 25.1 | 1.17 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_quant_awre.tar) | + +PaddleLite latency(ms) + +| Device | Model | Method | armv7 Thread 1 | armv7 Thread 2 | armv7 Thread 4 | armv8 Thread 1 | armv8 Thread 2 | armv8 Thread 4 | +| ------- | ----------- | ------------- | -------------- | -------------- | -------------- | -------------- | -------------- | -------------- | +| Qualcomm 835 | MobileNetV1 | FP32 baseline | 96.1942 | 53.2058 | 32.4468 | 88.4955 | 47.95 | 27.5189 | +| Qualcomm 835 | MobileNetV1 | quant_aware | 60.8186 | 32.1931 | 16.4275 | 56.4311 | 29.5446 | 15.1053 | +| Qualcomm 835 | MobileNetV1 | quant_post | 60.5615 | 32.4016 | 16.6596 | 56.5266 | 29.7178 | 15.1459 | +| Qualcomm 835 | MobileNetV2 | FP32 baseline | 65.715 | 38.1346 | 25.155 | 61.3593 | 36.2038 | 22.849 | +| Qualcomm 835 | MobileNetV2 | quant_aware | 48.3655 | 30.2021 | 21.9303 | 46.1487 | 27.3146 | 18.3053 | +| Qualcomm 835 | MobileNetV2 | quant_post | 48.3495 | 30.3069 | 22.1506 | 45.8715 | 27.4105 | 18.2223 | +| Qualcomm 835 | ResNet50 | FP32 baseline | 526.811 | 319.6486 | 205.8345 | 506.1138 | 335.1584 | 214.8936 | +| Qualcomm 835 | ResNet50 | quant_aware | 475.4538 | 256.8672 | 139.699 | 461.7344 | 247.9506 | 145.9847 | +| Qualcomm 835 | ResNet50 | quant_post | 476.0507 | 256.5963 | 139.7266 | 461.9176 | 248.3795 | 149.353 | +| Qualcomm 855 | MobileNetV1 | FP32 baseline | 33.5086 | 19.5773 | 11.7534 | 31.3474 | 18.5382 | 10.0811 | +| Qualcomm 855 | MobileNetV1 | quant_aware | 36.7067 | 21.628 | 11.0372 | 14.0238 | 8.199 | 4.2588 | +| Qualcomm 855 | MobileNetV1 | quant_post | 37.0498 | 21.7081 | 11.0779 | 14.0947 | 8.1926 | 4.2934 | +| Qualcomm 855 | MobileNetV2 | FP32 baseline | 25.0396 | 15.2862 | 9.6609 | 22.909 | 14.1797 | 8.8325 | +| Qualcomm 855 | MobileNetV2 | quant_aware | 28.1583 | 18.3317 | 11.8103 | 16.9158 | 11.1606 | 7.4148 | +| Qualcomm 855 | MobileNetV2 | quant_post | 28.1631 | 18.3917 | 11.8333 | 16.9399 | 11.1772 | 7.4176 | +| Qualcomm 855 | ResNet50 | FP32 baseline | 185.3705 | 113.0825 | 87.0741 | 177.7367 | 110.0433 | 74.4114 | +| Qualcomm 855 | ResNet50 | quant_aware | 327.6883 | 202.4536 | 106.243 | 243.5621 | 150.0542 | 78.4205 | +| Qualcomm 855 | ResNet50 | quant_post | 328.2683 | 201.9937 | 106.744 | 242.6397 | 150.0338 | 79.8659 | +| Kirin 970 | MobileNetV1 | FP32 baseline | 101.2455 | 56.4053 | 35.6484 | 94.8985 | 51.7251 | 31.9511 | +| Kirin 970 | MobileNetV1 | quant_aware | 62.5012 | 32.1863 | 16.6018 | 57.7477 | 29.2116 | 15.0703 | +| Kirin 970 | MobileNetV1 | quant_post | 62.4412 | 32.2585 | 16.6215 | 57.825 | 29.2573 | 15.1206 | +| Kirin 970 | MobileNetV2 | FP32 baseline | 70.4176 | 42.0795 | 25.1939 | 68.9597 | 39.2145 | 22.6617 | +| Kirin 970 | MobileNetV2 | quant_aware | 52.9961 | 31.5323 | 22.1447 | 49.4858 | 28.0856 | 18.7287 | +| Kirin 970 | MobileNetV2 | quant_post | 53.0961 | 31.7987 | 21.8334 | 49.383 | 28.2358 | 18.3642 | +| Kirin 970 | ResNet50 | FP32 baseline | 586.8943 | 344.0858 | 228.2293 | 573.3344 | 351.4332 | 225.8006 | +| Kirin 970 | ResNet50 | quant_aware | 488.361 | 260.1697 | 142.416 | 479.5668 | 249.8485 | 138.1742 | +| Kirin 970 | ResNet50 | quant_post | 489.6188 | 258.3279 | 142.6063 | 480.0064 | 249.5339 | 138.5284 | + +### 1.2 Pruning + +PaddleLite: + +env: Qualcomm SnapDragon 845 + armv8 + +criterion: time cost in Thread1/Thread2/Thread4 + +PaddleLite version: v2.3 + + +|Model | Method | Top-1/Top-5 Acc | ModelSize(MB) | GFLOPs |PaddleLite cost(ms)|TensorRT speed(FPS)| download | +|:--:|:---:|:--:|:--:|:--:|:--:|:--:|:--:| +| MobileNetV1 | Baseline | 70.99%/89.68% | 17 | 1.11 |66.052\35.8014\19.5762|-| [download](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) | +| MobileNetV1 | uniform -50% | 69.4%/88.66% (-1.59%/-1.02%) | 9 | 0.56 | 33.5636\18.6834\10.5076|-|[download](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_uniform-50.tar) | +| MobileNetV1 | sensitive -30% | 70.4%/89.3% (-0.59%/-0.38%) | 12 | 0.74 | 46.5958\25.3098\13.6982|-|[download](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-30.tar) | +| MobileNetV1 | sensitive -50% | 69.8% / 88.9% (-1.19%/-0.78%) | 9 | 0.56 |37.9892\20.7882\11.3144|-| [download](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-50.tar) | +| MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 |41.7874\23.375\13.3998|-| [download](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) | +| MobileNetV2 | uniform -50% | 65.79%/86.11% (-6.35%/-4.47%) | 11 | 0.296 |23.8842\13.8698\8.5572|-| [download](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_uniform-50.tar) | +| ResNet34 | - | 72.15%/90.65% | 84 | 7.36 |217.808\139.943\96.7504|342.32| [download](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet34_pretrained.tar) | +| ResNet34 | uniform -50% | 70.99%/89.95% (-1.36%/-0.87%) | 41 | 3.67 |114.787\75.0332\51.8438|452.41| [download](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_uniform-50.tar) | +| ResNet34 | auto -55.05% | 70.24%/89.63% (-2.04%/-1.06%) | 33 | 3.31 |105.924\69.3222\48.0246|457.25| [download](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_auto-55.tar) | + +### 1.3 Distillation + +| Model | Method | Top-1/Top-5 Acc | Model Size（MB） | Download | |:--:|:---:|:--:|:--:|:--:| -|MobileNetV1|-|70.99%/89.68%| xx | [下载链接]() | -|MobileNetV1|quant_post|xx%/xx%| xx | [下载链接]() | -|MobileNetV1|quant_aware|xx%/xx%| xx | [下载链接]() | -| MobileNetV2 | - |72.15%/90.65%| xx | [下载链接]() | -| MobileNetV2 | quant_post |xx%/xx%| xx | [下载链接]() | -| MobileNetV2 | quant_aware |xx%/xx%| xx | [下载链接]() | -|ResNet50|-|76.50%/93.00%| xx | [下载链接]() | -|ResNet50|quant_post|xx%/xx%| xx | [下载链接]() | -|ResNet50|quant_aware|xx%/xx%| xx | [下载链接]() | - +| MobileNetV1 | student | 70.99%/89.68% | 17 | [model](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) | +|ResNet50_vd|teacher|79.12%/94.44%| 99 | [model](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar) | +|MobileNetV1|ResNet50_vd[1](#trans1) distill|72.77%/90.68% (+1.78%/+1.00%)| 17 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_distilled.tar) | +| MobileNetV2 | student | 72.15%/90.65% | 15 | [model](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) | +| MobileNetV2 | ResNet50_vd distill | 74.28%/91.53% (+2.13%/+0.88%) | 15 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_distilled.tar) | +| ResNet50 | student | 76.50%/93.00% | 99 | [model](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar) | +|ResNet101|teacher|77.56%/93.64%| 173 | [model](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar) | +| ResNet50 | ResNet101 distill | 77.29%/93.65% (+0.79%/+0.65%) | 99 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_distilled.tar) | +Note: The `_vd` suffix indicates that the pre-trained model uses Mixup. Please refer to the detailed introduction: [mixup: Beyond Empirical Risk Minimization](https://arxiv.org/abs/1710.09412) -### 1.2 剪裁 +### 1.4 NAS -| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | GFLOPs | 下载 | +| Model | Method | Top-1/Top-5 Acc | Volume（MB） | GFLOPs | Download | |:--:|:---:|:--:|:--:|:--:|:--:| -| MobileNetV1 | Baseline | 70.99%/89.68% | 17 | 1.11 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) | -| MobileNetV1 | uniform -50% | 69.4%/88.66% (-1.59%/-1.02%) | 9 | 0.56 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_uniform-50.tar) | -| MobileNetV1 | sensitive -30% | 70.4%/89.3% (-0.59%/-0.38%) | 12 | 0.74 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-30.tar) | -| MobileNetV1 | sensitive -50% | 69.8% / 88.9% (-1.19%/-0.78%) | 9 | 0.56 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_sensitive-50.tar) | -| MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) | -| MobileNetV2 | uniform -50% | 65.79%/86.11% (-6.35%/-4.47%) | 11 | 0.296 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_uniform-50.tar) | -| ResNet34 | - | 72.15%/90.65% | 84 | 7.36 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet34_pretrained.tar) | -| ResNet34 | uniform -50% | 70.99%/89.95% (-1.36%/-0.87%) | 41 | 3.67 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_uniform-50.tar) | -| ResNet34 | auto -55.05% | 70.24%/89.63% (-2.04%/-1.06%) | 33 | 3.31 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet34_auto-55.tar) | +| MobileNetV2 | - | 72.15%/90.65% | 15 | 0.59 | [model](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) | +| MobileNetV2_NAS | SANAS | 71.518%/90.208% (-0.632%/-0.442%) | 14 | 0.295 | [model](https://paddlemodels.cdn.bcebos.com/PaddleSlim/MobileNetV2_sanas.tar) | +Dataset: Cifar10 +| Model | Method | Acc | Params（MB） | Download | +|:---:|:--:|:--:|:--:|:--:| +| Darts | - | 97.135% | 3.767 | - | +| Darts_SA(Based on Darts) | SANAS | 97.276%(+0.141%) | 3.344(-11.2%) | - | +Note: The token of MobileNetV2_NAS is [4, 4, 5, 1, 1, 2, 1, 1, 0, 2, 6, 2, 0, 3, 4, 5, 0, 4, 5, 5, 1, 4, 8, 0, 0]. The token of Darts_SA is [5, 5, 0, 5, 5, 10, 7, 7, 5, 7, 7, 11, 10, 12, 10, 0, 5, 3, 10, 8]. -### 1.3 蒸馏 +## 2. Object Detection -| 模型 | 压缩方法 | Top-1/Top-5 Acc | 模型体积（MB） | 下载 | -|:--:|:---:|:--:|:--:|:--:| -| MobileNetV1 | student | 70.99%/89.68% | 17 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.tar) | -|ResNet50_vd|teacher|79.12%/94.44%| 99 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_pretrained.tar) | -|MobileNetV1|ResNet50_vd[1](#trans1) distill|72.77%/90.68% (+1.78%/+1.00%)| 17 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV1_distilled.tar) | -| MobileNetV2 | student | 72.15%/90.65% | 15 | [下载链接](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.tar) | -| MobileNetV2 | ResNet50_vd distill | 74.28%/91.53% (+2.13%/+0.88%) | 15 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/MobileNetV2_distilled.tar) | -| ResNet50 | student | 76.50%/93.00% | 99 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.tar) | -|ResNet101|teacher|77.56%/93.64%| 173 | [下载链接](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.tar) | -| ResNet50 | ResNet101 distill | 77.29%/93.65% (+0.79%/+0.65%) | 99 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/ResNet50_distilled.tar) | +### 2.1 Quantization + +Dataset： COCO 2017 + +| Model | Method | Dataset | Image/GPU | Input 608 Box AP | Input 416 Box AP | Input 320 Box AP | Model Size（MB） | TensorRT latency(V100, ms) | Download | +| :----------------------------: | :---------: | :----: | :-------: | :------------: | :------------: | :------------: | :------------: | :----------: |:----------: | +| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.1 | 95 | - | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNet-V1-YOLOv3 | quant_post | COCO | 8 | 27.9 (-1.4)| 28.0 (-1.3) | 26.0 (-1.0) | 25 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | +| MobileNet-V1-YOLOv3 | quant_aware | COCO | 8 | 28.1 (-1.2)| 28.2 (-1.1) | 25.8 (-1.2) | 26.3 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_coco_quant_aware.tar) | +| R34-YOLOv3 | - | COCO | 8 | 36.2 | 34.3 | 31.4 | 162 | - | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) | +| R34-YOLOv3 | quant_post | COCO | 8 | 35.7 (-0.5) | - | - | 42.7 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_post.tar) | +| R34-YOLOv3 | quant_aware | COCO | 8 | 35.2 (-1.0) | 33.3 (-1.0) | 30.3 (-1.1)| 44 | - | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | +| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | - | - | 177 | 18.56 |[model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn_obj365_pretrained_coco.tar) | +| R50-dcn-YOLOv3 obj365_pretrain | quant_aware | COCO | 8 | 40.6 (-0.8) | 37.5 | 34.1 | 66 | 14.64 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | + + + +Dataset：WIDER-FACE + + + +| Model | Method | Image/GPU | Input Size | Easy/Medium/Hard | Model Size（MB） | Download | +| :------------: | :---------: | :-------: | :--------: | :-----------------------------: | :--------------: | :----------------------------------------------------------: | +| BlazeFace | - | 8 | 640 | 91.5/89.2/79.7 | 815 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_original.tar) | +| BlazeFace | quant_post | 8 | 640 | 87.8/85.1/74.9 (-3.7/-4.1/-4.8) | 228 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_post.tar) | +| BlazeFace | quant_aware | 8 | 640 | 90.5/87.9/77.6 (-1.0/-1.3/-2.1) | 228 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_origin_quant_aware.tar) | +| BlazeFace-Lite | - | 8 | 640 | 90.9/88.5/78.1 | 711 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_lite.tar) | +| BlazeFace-Lite | quant_post | 8 | 640 | 89.4/86.7/75.7 (-1.5/-1.8/-2.4) | 211 | [model]((https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_post.tar)) | +| BlazeFace-Lite | quant_aware | 8 | 640 | 89.7/87.3/77.0 (-1.2/-1.2/-1.1) | 211 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_lite_quant_aware.tar) | +| BlazeFace-NAS | - | 8 | 640 | 83.7/80.7/65.8 | 244 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas.tar) | +| BlazeFace-NAS | quant_post | 8 | 640 | 81.6/78.3/63.6 (-2.1/-2.4/-2.2) | 71 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_post.tar) | +| BlazeFace-NAS | quant_aware | 8 | 640 | 83.1/79.7/64.2 (-0.6/-1.0/-1.6) | 71 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/blazeface_nas_quant_aware.tar) | + +### 2.2 Pruning + +Dataset：Pasacl VOC & COCO 2017 + +PaddleLite: + +env: Qualcomm SnapDragon 845 + armv8 -!!! note "Note" +criterion: time cost in Thread1/Thread2/Thread4 - [1]：带_vd后缀代表该预训练模型使用了Mixup，Mixup相关介绍参考[mixup: Beyond Empirical Risk Minimization](https://arxiv.org/abs/1710.09412) +PaddleLite version: v2.3 +| Model | Method | Dataset | Image/GPU | Input 608 Box AP | Input 416 Box AP | Input 320 Box AP | Model Size(MB) | GFLOPs (608*608) | PaddleLite cost(ms)(608*608) | TensorRT speed(FPS)(608*608) | Download | +| :----------------------------: | :---------------: | :--------: | :-------: | :--------------: | :--------------: | :--------------: | :------------: | :--------------: | :--------------: | :--------------: | :----------------------------: | +| MobileNet-V1-YOLOv3 | Baseline | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | 40.49 | 1238\796.943\520.101 |60.40| [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| MobileNet-V1-YOLOv3 | sensitive -52.88% | Pascal VOC | 8 | 77.6 (+1.4) | 77.7 (1.0) | 75.5 (+0.2) | 31 | 19.08 | 602.497\353.759\222.427 |99.36| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_voc_prune.tar) | +| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | 41.35 |-|-| [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| MobileNet-V1-YOLOv3 | sensitive -51.77% | COCO | 8 | 26.0 (-3.3) | 25.1 (-4.2) | 22.6 (-4.4) | 32 | 19.94 |-|73.93| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_prune.tar) | +| R50-dcn-YOLOv3 | - | COCO | 8 | 39.1 | - | - | 177 | 89.60 |-|27.68| [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn.tar) | +| R50-dcn-YOLOv3 | sensitive -9.37% | COCO | 8 | 39.3 (+0.2) | - | - | 150 | 81.20 |-|30.08| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune.tar) | +| R50-dcn-YOLOv3 | sensitive -24.68% | COCO | 8 | 37.3 (-1.8) | - | - | 113 | 67.48 |-|34.32| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune578.tar) | +| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | - | - | 177 | 89.60 |-|-| [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn_obj365_pretrained_coco.tar) | +| R50-dcn-YOLOv3 obj365_pretrain | sensitive -9.37% | COCO | 8 | 40.5 (-0.9) | - | - | 150 | 81.20 |-|-| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune.tar) | +| R50-dcn-YOLOv3 obj365_pretrain | sensitive -24.68% | COCO | 8 | 37.8 (-3.3) | - | - | 113 | 67.48 |-|-| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune578.tar) | -## 2. 目标检测 +### 2.3 Distillation -### 2.1 量化 +Dataset：Pasacl VOC & COCO 2017 -数据集： COCO 2017 -| 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积（MB） | 下载 | -| :----------------------------: | :---------: | :----: | :-------: | :------------: | :------------: | :------------: | :------------: | :----------: | -| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.1 | xx | [下载链接]() | -| MobileNet-V1-YOLOv3 | quant_post | COCO | 8 | xx | xx | xx | xx | [下载链接]() | -| MobileNet-V1-YOLOv3 | quant_aware | COCO | 8 | xx | xx | xx | xx | [下载链接]() | -| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | xx | xx | xx | [下载链接]() | -| R50-dcn-YOLOv3 obj365_pretrain | quant_post | COCO | 8 | xx | xx | xx | xx | [下载链接]() | -| R50-dcn-YOLOv3 obj365_pretrain | quant_aware | COCO | 8 | xx | xx | xx | xx | [下载链接]() | +| Model | Method | Dataset | Image/GPU | Input 608 Box AP | Input 416 Box AP | Input 320 Box AP | Model Size（MB） | Download | +| :-----------------: | :---------------------: | :--------: | :-------: | :--------------: | :--------------: | :--------------: | :--------------: | :----------------------------------------------------------: | +| MobileNet-V1-YOLOv3 | - | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | +| ResNet34-YOLOv3 | - | Pascal VOC | 8 | 82.6 | 81.9 | 80.1 | 162 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34_voc.tar) | +| MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | Pascal VOC | 8 | 79.0 (+2.8) | 78.2 (+1.5) | 75.5 (+0.2) | 94 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) | +| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | +| ResNet34-YOLOv3 | - | COCO | 8 | 36.2 | 34.3 | 31.4 | 163 | [model](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) | +| MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | COCO | 8 | 31.4 (+2.1) | 30.0 (+0.7) | 27.1 (+0.1) | 95 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | +### 2.4 NAS -数据集：WIDER-FACE +Dataset: WIDER-FACE +| Model | Method | Image/GPU | Input size | Easy/Medium/Hard | volume（KB） | latency（ms）| Download | +| :------------: | :---------: | :-------: | :------: | :-----------------------------: | :------------: | :------------: | :----------------------------------------------------------: | +| BlazeFace | - | 8 | 640 | 91.5/89.2/79.7 | 815 | 71.862 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_original.tar) | +| BlazeFace-NAS | - | 8 | 640 | 83.7/80.7/65.8 | 244 | 21.117 |[model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas.tar) | +| BlazeFace-NASV2 | SANAS | 8 | 640 | 87.0/83.7/68.5 | 389 | 22.558 | [model](https://paddlemodels.bj.bcebos.com/object_detection/blazeface_nas2.tar) | +Note: latency is based on latency_855.txt, the file is test on 855 by PaddleLite。The config of BlazeFace-NASV2 is in [there](https://github.com/PaddlePaddle/PaddleDetection/blob/master/configs/face_detection/blazeface_nas_v2.yml). -| 模型 | 压缩方法 | Image/GPU | 输入尺寸 | Easy/Medium/Hard | 模型体积（MB） | 下载 | -| :------------: | :---------: | :-------: | :------: | :---------------: | :------------: | :----------: | -| BlazeFace | - | 8 | 640 | 0.915/0.892/0.797 | xx | [下载链接]() | -| BlazeFace | quant_post | 8 | 640 | xx/xx/xx | xx | [下载链接]() | -| BlazeFace | quant_aware | 8 | 640 | xx/xx/xx | xx | [下载链接]() | -| BlazeFace-Lite | - | 8 | 640 | 0.909/0.885/0.781 | xx | [下载链接]() | -| BlazeFace-Lite | quant_post | 8 | 640 | xx/xx/xx | xx | [下载链接]() | -| BlazeFace-Lite | quant_aware | 8 | 640 | xx/xx/xx | xx | [下载链接]() | -| BlazeFace-NAS | - | 8 | 640 | 0.837/0.807/0.658 | xx | [下载链接]() | -| BlazeFace-NAS | quant_post | 8 | 640 | xx/xx/xx | xx | [下载链接]() | -| BlazeFace-NAS | quant_aware | 8 | 640 | xx/xx/xx | xx | [下载链接]() | +## 3. Image Segmentation +Dataset：Cityscapes -### 2.2 剪裁 +### 3.1 Quantization -数据集：Pasacl VOC & COCO 2017 +| Model | Method | mIoU | Model Size（MB） | Download | +| :--------------------: | :---------: | :-----------: | :--------------: | :----------------------------------------------------------: | +| DeepLabv3+/MobileNetv1 | - | 63.26 | 6.6 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv1.tar ) | +| DeepLabv3+/MobileNetv1 | quant_post | 58.63 (-4.63) | 1.8 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv1_2049x1025_quant_post.tar) | +| DeepLabv3+/MobileNetv1 | quant_aware | 62.03 (-1.23) | 1.8 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv1_2049x1025_quant_aware.tar) | +| DeepLabv3+/MobileNetv2 | - | 69.81 | 7.4 | [model](https://paddleseg.bj.bcebos.com/models/mobilenet_cityscapes.tgz) | +| DeepLabv3+/MobileNetv2 | quant_post | 67.59 (-2.22) | 2.1 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv2_2049x1025_quant_post.tar) | +| DeepLabv3+/MobileNetv2 | quant_aware | 68.33 (-1.48) | 2.1 | [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/deeplabv3_mobilenetv2_2049x1025_quant_aware.tar) | -| 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积(MB) | GFLOPs (608*608) | 下载 | -| :----------------------------: | :---------------: | :--------: | :-------: | :------------: | :------------: | :------------: | :----------: | :--------------: | :----------------------------------------------------------: | -| MobileNet-V1-YOLOv3 | Baseline | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | 40.49 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | -| MobileNet-V1-YOLOv3 | sensitive -52.88% | Pascal VOC | 8 | 77.6 (+1.4) | 77.7 (1.0) | 75.5 (+0.2) | 31 | 19.08 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_voc_prune.tar) | -| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | 41.35 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | -| MobileNet-V1-YOLOv3 | sensitive -51.77% | COCO | 8 | 26.0 (-3.3) | 25.1 (-4.2) | 22.6 (-4.4) | 32 | 19.94 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenet_v1_prune.tar) | -| R50-dcn-YOLOv3 | - | COCO | 8 | 39.1 | - | - | 177 | 89.60 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn.tar) | -| R50-dcn-YOLOv3 | sensitive -9.37% | COCO | 8 | 39.3 (+0.2) | - | - | 150 | 81.20 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune.tar) | -| R50-dcn-YOLOv3 | sensitive -24.68% | COCO | 8 | 37.3 (-1.8) | - | - | 113 | 67.48 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_prune578.tar) | -| R50-dcn-YOLOv3 obj365_pretrain | - | COCO | 8 | 41.4 | - | - | 177 | 89.60 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r50vd_dcn_obj365_pretrained_coco.tar) | -| R50-dcn-YOLOv3 obj365_pretrain | sensitive -9.37% | COCO | 8 | 40.5 (-0.9) | - | - | 150 | 81.20 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune.tar) | -| R50-dcn-YOLOv3 obj365_pretrain | sensitive -24.68% | COCO | 8 | 37.8 (-3.3) | - | - | 113 | 67.48 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_prune578.tar) | +Image segmentation model PaddleLite latency (ms), input size 769x769 -### 2.3 蒸馏 +| Device | Model | Method | armv7 Thread 1 | armv7 Thread 2 | armv7 Thread 4 | armv8 Thread 1 | armv8 Thread 2 | armv8 Thread 4 | +| ------------ | ---------------------- | ------------- | -------------- | -------------- | -------------- | -------------- | -------------- | -------------- | +| Qualcomm 835 | Deeplabv3- MobileNetV1 | FP32 baseline | 1227.9894 | 734.1922 | 527.9592 | 1109.96 | 699.3818 | 479.0818 | +| Qualcomm 835 | Deeplabv3- MobileNetV1 | quant_aware | 848.6544 | 512.785 | 382.9915 | 752.3573 | 455.0901 | 307.8808 | +| Qualcomm 835 | Deeplabv3- MobileNetV1 | quant_post | 840.2323 | 510.103 | 371.9315 | 748.9401 | 452.1745 | 309.2084 | +| Qualcomm 835 | Deeplabv3-MobileNetV2 | FP32 baseline | 1282.8126 | 793.2064 | 653.6538 | 1193.9908 | 737.1827 | 593.4522 | +| Qualcomm 835 | Deeplabv3-MobileNetV2 | quant_aware | 976.0495 | 659.0541 | 513.4279 | 892.1468 | 582.9847 | 484.7512 | +| Qualcomm 835 | Deeplabv3-MobileNetV2 | quant_post | 981.44 | 658.4969 | 538.6166 | 885.3273 | 586.1284 | 484.0018 | +| Qualcomm 855 | Deeplabv3- MobileNetV1 | FP32 baseline | 568.8748 | 339.8578 | 278.6316 | 420.6031 | 281.3197 | 217.5222 | +| Qualcomm 855 | Deeplabv3- MobileNetV1 | quant_aware | 608.7578 | 347.2087 | 260.653 | 241.2394 | 177.3456 | 143.9178 | +| Qualcomm 855 | Deeplabv3- MobileNetV1 | quant_post | 609.0142 | 347.3784 | 259.9825 | 239.4103 | 180.1894 | 139.9178 | +| Qualcomm 855 | Deeplabv3-MobileNetV2 | FP32 baseline | 639.4425 | 390.1851 | 322.7014 | 477.7667 | 339.7411 | 262.2847 | +| Qualcomm 855 | Deeplabv3-MobileNetV2 | quant_aware | 703.7275 | 497.689 | 417.1296 | 394.3586 | 300.2503 | 239.9204 | +| Qualcomm 855 | Deeplabv3-MobileNetV2 | quant_post | 705.7589 | 474.4076 | 427.2951 | 394.8352 | 297.4035 | 264.6724 | +| Kirin 970 | Deeplabv3- MobileNetV1 | FP32 baseline | 1682.1792 | 1437.9774 | 1181.0246 | 1261.6739 | 1068.6537 | 690.8225 | +| Kirin 970 | Deeplabv3- MobileNetV1 | quant_aware | 1062.3394 | 1248.1014 | 878.3157 | 774.6356 | 710.6277 | 528.5376 | +| Kirin 970 | Deeplabv3- MobileNetV1 | quant_post | 1109.1917 | 1339.6218 | 866.3587 | 771.5164 | 716.5255 | 500.6497 | +| Kirin 970 | Deeplabv3-MobileNetV2 | FP32 baseline | 1771.1301 | 1746.0569 | 1222.4805 | 1448.9739 | 1192.4491 | 760.606 | +| Kirin 970 | Deeplabv3-MobileNetV2 | quant_aware | 1320.2905 | 921.4522 | 676.0732 | 1145.8801 | 821.5685 | 590.1713 | +| Kirin 970 | Deeplabv3-MobileNetV2 | quant_post | 1320.386 | 918.5328 | 672.2481 | 1020.753 | 820.094 | 591.4114 | -数据集：Pasacl VOC & COCO 2017 -| 模型 | 压缩方法 | 数据集 | Image/GPU | 输入608 Box AP | 输入416 Box AP | 输入320 Box AP | 模型体积（MB） | 下载 | -| :-----------------: | :---------------------: | :--------: | :-------: | :------------: | :------------: | :------------: | :------------: | :----------------------------------------------------------: | -| MobileNet-V1-YOLOv3 | - | Pascal VOC | 8 | 76.2 | 76.7 | 75.3 | 94 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar) | -| ResNet34-YOLOv3 | - | Pascal VOC | 8 | 82.6 | 81.9 | 80.1 | 162 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34_voc.tar) | -| MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | Pascal VOC | 8 | 79.0 (+2.8) | 78.2 (+1.5) | 75.5 (+0.2) | 94 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar) | -| MobileNet-V1-YOLOv3 | - | COCO | 8 | 29.3 | 29.3 | 27.0 | 95 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1.tar) | -| ResNet34-YOLOv3 | - | COCO | 8 | 36.2 | 34.3 | 31.4 | 163 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_r34.tar) | -| MobileNet-V1-YOLOv3 | ResNet34-YOLOv3 distill | COCO | 8 | 31.4 (+2.1) | 30.0 (+0.7) | 27.1 (+0.1) | 95 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar) | -## 3. 图像分割 +### 3.2 Pruning -数据集：Cityscapes +PaddleLite: -### 3.1 量化 +env: Qualcomm SnapDragon 845 + armv8 -| 模型 | 压缩方法 | mIoU | 模型体积（MB） | 下载 | -| :--------------------: | :---------: | :---: | :------------: | :----------: | -| DeepLabv3+/MobileNetv1 | - | 63.26 | xx | [下载链接]() | -| DeepLabv3+/MobileNetv1 | quant_post | xx | xx | [下载链接]() | -| DeepLabv3+/MobileNetv1 | quant_aware | xx | xx | [下载链接]() | -| DeepLabv3+/MobileNetv2 | - | 69.81 | xx | [下载链接]() | -| DeepLabv3+/MobileNetv2 | quant_post | xx | xx | [下载链接]() | -| DeepLabv3+/MobileNetv2 | quant_aware | xx | xx | [下载链接]() | +criterion: time cost in Thread1/Thread2/Thread4 -### 3.2 剪裁 +PaddleLite version: v2.3 -| 模型 | 压缩方法 | mIoU | 模型体积（MB） | GFLOPs | 下载 | -| :-------: | :---------------: | :-----------: | :------------: | :----: | :----------------------------------------------------------: | -| fast-scnn | baseline | 69.64 | 11 | 14.41 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape.tar) | -| fast-scnn | uniform -17.07% | 69.58 (-0.06) | 8.5 | 11.95 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_uniform-17.tar) | -| fast-scnn | sensitive -47.60% | 66.68 (-2.96) | 5.7 | 7.55 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_sensitive-47.tar) | +| Model | Method | mIoU | Model Size（MB） | GFLOPs | PaddleLite cost(ms) | TensorRT speed(FPS) | Download | +| :-------: | :---------------: | :-----------: | :--------------: | :----: | :--------------: | :----: | :-------------------: | +| fast-scnn | baseline | 69.64 | 11 | 14.41 | 1226.36\682.96\415.664 |39.53| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape.tar) | +| fast-scnn | uniform -17.07% | 69.58 (-0.06) | 8.5 | 11.95 | 1140.37\656.612\415.888 |42.01| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_uniform-17.tar) | +| fast-scnn | sensitive -47.60% | 66.68 (-2.96) | 5.7 | 7.55 | 866.693\494.467\291.748 |51.48| [model](https://paddlemodels.bj.bcebos.com/PaddleSlim/fast_scnn_cityscape_sensitive-47.tar) | diff --git a/_sources/tutorials/darts_nas_turorial.md.txt b/_sources/tutorials/darts_nas_turorial.md.txt new file mode 100644 index 00000000..2f99a6c1 --- /dev/null +++ b/_sources/tutorials/darts_nas_turorial.md.txt @@ -0,0 +1,275 @@ +# SANAS进阶版实验教程-压缩DARTS产出模型 + +## 收益情况 +利用DARTS搜索出来的最终模型结构（以下简称为DARTS_model）构造相应的搜索空间，根据PaddleSlim提供的SANAS搜索方法进行搜索实验，最终得到的模型结构（以下简称为DARTS_SA）相比DARTS_model的精度提升0.141%，模型大小下降11.2%。 + +## 搜索教程 +本教程展示了如何在DARTS_model基础上利用SANAS进行搜索实验，并得到DARTS_SA的结果。 + +本教程包含以下步骤： +1. 构造搜索空间 +2. 导入依赖包并定义全局变量 +3. 初始化SANAS实例 +4. 定义计算模型参数量的函数 +5. 定义网络输入数据的函数 +6. 定义造program的函数 +7. 定义训练函数 +8. 定义预测函数 +9. 启动搜索 + 9.1 获取下一个模型结构 + 9.2 构造相应的训练和预测program + 9.3 添加搜索限制 + 9.4 定义环境 + 9.5 定义输入数据 + 9.6 启动训练和评估 + 9.7 回传当前模型的得分reward +10. 利用demo下的脚本启动搜索 +11. 利用demo下的脚本启动最终实验 + +### 1. 构造搜索空间 +进行搜索实验之前，首先需要根据DARTS_model的模型特点构造相应的搜索空间，本次实验仅会对DARTS_model的通道数进行搜索，搜索的目的是得到一个精度更高并且模型参数更少的模型。 +定义如下搜索空间： +- 通道数`filter_num`: 定义了每个卷积操作的通道数变化区间。取值区间为：`[4, 8, 12, 16, 20, 36, 54, 72, 90, 108, 144, 180, 216, 252]` + +按照通道数来区分DARTS_model中block的话，则DARTS_model中共有3个block，第一个block仅包含6个normal cell，之后的两个block每个block都包含和一个reduction cell和6个normal cell，共有20个cell。在构造搜索空间的时候我们定义每个cell中的所有卷积操作都使用相同的通道数，共有20位token。 + +完整的搜索空间可以参考[基于DARTS_model的搜索空间](../../../paddleslim/nas/search_space/darts_space.py) + +### 2. 引入依赖包并定义全局变量 +```python +import numpy as np +import paddle +import paddle.fluid as fluid +from paddleslim.nas import SANAS + +BATCH_SIZE=96 +SERVER_ADDRESS = "" +PORT = 8377 +SEARCH_STEPS = 300 +RETAIN_EPOCH=30 +MAX_PARAMS=3.77 +IMAGE_SHAPE=[3, 32, 32] +AUXILIARY = True +AUXILIARY_WEIGHT= 0.4 +TRAINSET_NUM = 50000 +LR = 0.025 +MOMENTUM = 0.9 +WEIGHT_DECAY = 0.0003 +DROP_PATH_PROBILITY = 0.2 +``` + +### 3. 初始化SANAS实例 +首先需要初始化SANAS示例。 +```python +config = [('DartsSpace')] +sa_nas = SANAS(config, server_addr=(SERVER_ADDRESS, PORT), search_steps=SEARCH_STEPS, is_server=True) +``` + +### 4. 定义计算模型参数量的函数 +根据输入的program计算当前模型中的参数量。本教程使用模型参数量作为搜索的限制条件。 +```python +def count_parameters_in_MB(all_params, prefix='model'): + parameters_number = 0 + for param in all_params: + if param.name.startswith( + prefix) and param.trainable and 'aux' not in param.name: + parameters_number += np.prod(param.shape) + return parameters_number / 1e6 +``` + +### 5. 定义网络输入数据的函数 +根据输入图片的尺寸定义网络中的输入，其中包括图片输入、标签输入和在训练过程中需要随机丢弃单元的比例和掩膜。 +```python +def create_data_loader(IMAGE_SHAPE, is_train): + image = fluid.data( + name="image", shape=[None] + IMAGE_SHAPE, dtype="float32") + label = fluid.data(name="label", shape=[None, 1], dtype="int64") + data_loader = fluid.io.DataLoader.from_generator( + feed_list=[image, label], + capacity=64, + use_double_buffer=True, + iterable=True) + drop_path_prob = '' + drop_path_mask = '' + if is_train: + drop_path_prob = fluid.data( + name="drop_path_prob", shape=[BATCH_SIZE, 1], dtype="float32") + drop_path_mask = fluid.data( + name="drop_path_mask", + shape=[BATCH_SIZE, 20, 4, 2], + dtype="float32") + + return data_loader, image, label, drop_path_prob, drop_path_mask +``` + +### 6. 定义构造program的函数 +根据输入的模型结构、输入图片尺寸和当前program是否是训练模式构造program。 +```python +def build_program(main_program, startup_program, IMAGE_SHAPE, archs, is_train): + with fluid.program_guard(main_program, startup_program): + data_loader, data, label, drop_path_prob, drop_path_mask = create_data_loader( + IMAGE_SHAPE, is_train) + logits, logits_aux = archs(data, drop_path_prob, drop_path_mask, + is_train, 10) + top1 = fluid.layers.accuracy(input=logits, label=label, k=1) + top5 = fluid.layers.accuracy(input=logits, label=label, k=5) + loss = fluid.layers.reduce_mean( + fluid.layers.softmax_with_cross_entropy(logits, label)) + + if is_train: + if AUXILIARY: + loss_aux = fluid.layers.reduce_mean( + fluid.layers.softmax_with_cross_entropy(logits_aux, label)) + loss = loss + AUXILIARY_WEIGHT * loss_aux + step_per_epoch = int(TRAINSET_NUM / BATCH_SIZE) + learning_rate = fluid.layers.cosine_decay(LR, step_per_epoch, RETAIN_EPOCH) + fluid.clip.set_gradient_clip( + clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=5.0)) + optimizer = fluid.optimizer.MomentumOptimizer( + learning_rate, + MOMENTUM, + regularization=fluid.regularizer.L2DecayRegularizer( + WEIGHT_DECAY)) + optimizer.minimize(loss) + outs = [loss, top1, top5, learning_rate] + else: + outs = [loss, top1, top5] + return outs, data_loader + +``` + +### 7. 定义训练函数 +```python +def train(main_prog, exe, epoch_id, train_loader, fetch_list): + loss = [] + top1 = [] + top5 = [] + for step_id, data in enumerate(train_loader()): + devices_num = len(data) + if DROP_PATH_PROBILITY > 0: + feed = [] + for device_id in range(devices_num): + image = data[device_id]['image'] + label = data[device_id]['label'] + drop_path_prob = np.array( + [[DROP_PATH_PROBILITY * epoch_id / RETAIN_EPOCH] + for i in range(BATCH_SIZE)]).astype(np.float32) + drop_path_mask = 1 - np.random.binomial( + 1, drop_path_prob[0], + size=[BATCH_SIZE, 20, 4, 2]).astype(np.float32) + feed.append({ + "image": image, + "label": label, + "drop_path_prob": drop_path_prob, + "drop_path_mask": drop_path_mask + }) + else: + feed = data + loss_v, top1_v, top5_v, lr = exe.run( + main_prog, feed=feed, fetch_list=[v.name for v in fetch_list]) + loss.append(loss_v) + top1.append(top1_v) + top5.append(top5_v) + if step_id % 10 == 0: + print( + "Train Epoch {}, Step {}, Lr {:.8f}, loss {:.6f}, acc_1 {:.6f}, acc_5 {:.6f}". + format(epoch_id, step_id, lr[0], np.mean(loss), np.mean(top1), np.mean(top5))) + return np.mean(top1) +``` + +### 8. 定义预测函数 +```python +def valid(main_prog, exe, epoch_id, valid_loader, fetch_list): + loss = [] + top1 = [] + top5 = [] + for step_id, data in enumerate(valid_loader()): + loss_v, top1_v, top5_v = exe.run( + main_prog, feed=data, fetch_list=[v.name for v in fetch_list]) + loss.append(loss_v) + top1.append(top1_v) + top5.append(top5_v) + if step_id % 10 == 0: + print( + "Valid Epoch {}, Step {}, loss {:.6f}, acc_1 {:.6f}, acc_5 {:.6f}". + format(epoch_id, step_id, np.mean(loss), np.mean(top1), np.mean(top5))) + return np.mean(top1) +``` + +### 9. 启动搜索实验 +以下步骤拆解说明了如何获得当前模型结构以及获得当前模型结构之后应该有的步骤。 + +#### 9.1 获取下一个模型结构 +根据上面的SANAS实例中的函数获取下一个模型结构。 +```python +archs = sa_nas.next_archs()[0] +``` + +#### 9.2 构造训练和预测program +根据上一步中获得的模型结构分别构造训练program和预测program。 +```python +train_program = fluid.Program() +test_program = fluid.Program() +startup_program = fluid.Program() +train_fetch_list, train_loader = build_program(train_program, startup_program, IMAGE_SHAPE, archs, is_train=True) +test_fetch_list, test_loader = build_program(test_program, startup_program, IMAGE_SHAPE, archs, is_train=False) +test_program = test_program.clone(for_test=True) +``` + +#### 9.3 添加搜索限制 +本教程以模型参数量为限制条件。首先计算一下当前program的参数量，如果超出限制条件，则终止本次模型结构的训练，获取下一个模型结构。 +```python +current_params = count_parameters_in_MB( + train_program.global_block().all_parameters(), 'cifar10') +``` + +#### 9.4 定义环境 +定义数据和模型的环境并初始化参数。 +```python +place = fluid.CPUPlace() +exe = fluid.Executor(place) +exe.run(startup_program) +``` + +#### 9.5 定义输入数据 +由于本示例中对cifar10中的图片进行了一些额外的预处理操作，和[快速开始](../quick_start/nas_tutorial.md)示例中的reader不同，所以需要自定义cifar10的reader，不能直接调用paddle中封装好的`paddle.dataset.cifar10`的reader。自定义cifar10的reader文件位于[demo/nas](../../../demo/nas/darts_cifar10_reader.py)中。 + +**注意：**本示例为了简化代码直接调用`paddle.dataset.cifar10`定义训练数据和预测数据，实际训练需要使用自定义cifar10的reader。 +```python +train_reader = paddle.batch(paddle.reader.shuffle(paddle.dataset.cifar.train10(cycle=False), buf_size=1024), batch_size=BATCH_SIZE, drop_last=True) +test_reader = paddle.batch(paddle.dataset.cifar.test10(cycle=False), batch_size=BATCH_SIZE, drop_last=False) +train_loader.set_sample_list_generator(train_reader, places=place) +test_loader.set_sample_list_generator(test_reader, places=place) +``` + +#### 9.6 启动训练和评估 +```python +for epoch_id in range(RETAIN_EPOCH): + train_top1 = train(train_program, exe, epoch_id, train_loader, train_fetch_list) + print("TRAIN: Epoch {}, train_acc {:.6f}".format(epoch_id, train_top1)) + valid_top1 = valid(test_program, exe, epoch_id, test_loader, test_fetch_list) + print("TEST: Epoch {}, valid_acc {:.6f}".format(epoch_id, valid_top1)) + valid_top1_list.append(valid_top1) +``` + +#### 9.7 回传当前模型的得分reward +本教程利用最后两个epoch的准确率均值作为最终的得分回传给SANAS。 +```python +sa_nas.reward(float(valid_top1_list[-1] + valid_top1_list[-2]) / 2) +``` + + +### 10. 利用demo下的脚本启动搜索 + +搜索文件位于: [darts_sanas_demo](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/demo/nas/darts_nas.py)，搜索过程中限制模型参数量为不大于3.77M。 +```python +cd demo/nas/ +python darts_nas.py +``` + +### 11. 利用demo下的脚本启动最终实验 +最终实验文件位于: [darts_sanas_demo](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/demo/nas/darts_nas.py)，最终实验需要训练600epoch。以下示例输入token为`[5, 5, 0, 5, 5, 10, 7, 7, 5, 7, 7, 11, 10, 12, 10, 0, 5, 3, 10, 8]`。 +```python +cd demo/nas/ +python darts_nas.py --token 5 5 0 5 5 10 7 7 5 7 7 11 10 12 10 0 5 3 10 8 --retain_epoch 600 +``` diff --git a/_sources/tutorials/image_classification_sensitivity_analysis_tutorial_en.md.txt b/_sources/tutorials/image_classification_sensitivity_analysis_tutorial_en.md.txt new file mode 100644 index 00000000..2b084074 --- /dev/null +++ b/_sources/tutorials/image_classification_sensitivity_analysis_tutorial_en.md.txt @@ -0,0 +1,263 @@ +# Pruning of image classification model - sensitivity + +In this tutorial, you will learn how to use [sensitivity API of PaddleSlim](https://paddlepaddle.github.io/PaddleSlim/api/prune_api/#sensitivity) by a demo of MobileNetV1 model on MNIST dataset。 +This tutorial following workflow: + +1. Import dependency +2. Build model +3. Define data reader +4. Define function for test +5. Training model +6. Get names of parameter +7. Compute sensitivities +8. Pruning model + + +## 1. Import dependency + +PaddleSlim dependents on Paddle1.7. Please ensure that you have installed paddle correctly. Import Paddle and PaddleSlim as below: + +```python +import paddle +import paddle.fluid as fluid +import paddleslim as slim +``` + +## 2. Build model + +This section will build a classsification model based `MobileNetV1` for MNIST task. The shape of the input is `[1, 28, 28]` and the output number is 10. + +To make the code simple, we define a function in package `paddleslim.models` to build classification model. +Excute following code to build a model, + + +```python +exe, train_program, val_program, inputs, outputs = slim.models.image_classification("MobileNet", [1, 28, 28], 10, use_gpu=True) +place = fluid.CUDAPlace(0) +``` + +>Note：The functions in paddleslim.models is just used in tutorials or demos. + +## 3 Define data reader + +MNIST dataset is used for making the demo can be executed quickly. It defines some functions for downloading and reading MNIST dataset in package `paddle.dataset.mnist`. +Show as below： + +```python +import paddle.dataset.mnist as reader +train_reader = paddle.batch( + reader.train(), batch_size=128, drop_last=True) +test_reader = paddle.batch( + reader.test(), batch_size=128, drop_last=True) +data_feeder = fluid.DataFeeder(inputs, place) +``` + +## 4. Define test function + +To get the performance of model on test dataset after pruning a convolution layer, we define a test function as below: + +```python +import numpy as np +def test(program): + acc_top1_ns = [] + acc_top5_ns = [] + for data in test_reader(): + acc_top1_n, acc_top5_n, _ = exe.run( + program, + feed=data_feeder.feed(data), + fetch_list=outputs) + acc_top1_ns.append(np.mean(acc_top1_n)) + acc_top5_ns.append(np.mean(acc_top5_n)) + print("Final eva - acc_top1: {}; acc_top5: {}".format( + np.mean(np.array(acc_top1_ns)), np.mean(np.array(acc_top5_ns)))) + return np.mean(np.array(acc_top1_ns)) +``` + +## 5. Training model + +Sensitivity analysis is dependent on pretrained model. So we should train the model defined in section 2 for some epochs. One epoch training is enough for this simple demo while more epochs may be necessary for other model. Or you can load pretrained model from filesystem. + +Training model as below: + + +```python +for data in train_reader(): + acc1, acc5, loss = exe.run(train_program, feed=data_feeder.feed(data), fetch_list=outputs) +print(np.mean(acc1), np.mean(acc5), np.mean(loss)) +``` + +Get the performance using the test function defined in section 4: + +```python +test(val_program) +``` + +## 6. Get names of parameters + +```python +params = [] +for param in train_program.global_block().all_parameters(): + if "_sep_weights" in param.name: + params.append(param.name) +print(params) +params = params[:5] +``` + +## 7. Compute sensitivities + +### 7.1 Compute in single process + +Apply sensitivity analysis on pretrained model by calling [sensitivity API](https://paddlepaddle.github.io/PaddleSlim/api/prune_api/#sensitivity). + +The sensitivities will be appended into the file given by option `sensitivities_file` during computing. +The information in this file won`t be computed repeatedly. + +Remove the file `sensitivities_0.data` in current directory: + +```python +!rm -rf sensitivities_0.data +``` + +Apart from the parameters to be analyzed, it also support for setting the ratios that each convolutoin will be pruned. + +If one model losses 90% accuracy on test dataset when its single convolution layer is pruned by 40%, then we can set `pruned_ratios` to `[0.1, 0.2, 0.3, 0.4]`. + +The granularity of `pruned_ratios` should be small to get more reasonable sensitivities. But small granularity of `pruned_ratios` will slow down the computing. + +```python +sens_0 = slim.prune.sensitivity( + val_program, + place, + params, + test, + sensitivities_file="sensitivities_0.data", + pruned_ratios=[0.1, 0.2]) +print(sens_0) +``` + +### 7.2 Expand sensitivities + +We can expand `pruned_ratios` to `[0.1, 0.2, 0.3]` based the sensitivities generated in section 7.1. + +```python +sens_0 = slim.prune.sensitivity( + val_program, + place, + params, + test, + sensitivities_file="sensitivities_0.data", + pruned_ratios=[0.3]) +print(sens_0) +``` + +### 7.3 Computing sensitivity in multi-process + +The time cost of computing sensitivities is dependent on the count of parameters and the speed of model evaluation on test dataset. We can speed up computing by multi-process. + +Split `pruned_ratios` into multi-process, and merge the sensitivities from multi-process. + +#### 7.3.1 Computing in each process + +We have compute the sensitivities when `pruned_ratios=[0.1, 0.2, 0.3]` and saved the sensitivities into file named `sensitivities_0.data`. + +在另一个进程中，The we start a task by setting `pruned_ratios=[0.4]` in another process and save result into file named `sensitivities_1.data`. Show as below： + + +```python +sens_1 = slim.prune.sensitivity( + val_program, + place, + params, + test, + sensitivities_file="sensitivities_1.data", + pruned_ratios=[0.4]) +print(sens_1) +``` + +#### 7.3.2 Load sensitivity file generated in multi-process + +```python +s_0 = slim.prune.load_sensitivities("sensitivities_0.data") +s_1 = slim.prune.load_sensitivities("sensitivities_1.data") +print(s_0) +print(s_1) +``` + +#### 7.3.3 Merge sensitivies + + +```python +s = slim.prune.merge_sensitive([s_0, s_1]) +print(s) +``` + +## 8. Pruning model + +Pruning model according to the sensitivities generated in section 7.3.3. + +### 8.1 Get pruning ratios + +Get a group of ratios by calling [get_ratios_by_loss](https://paddlepaddle.github.io/PaddleSlim/api/prune_api/#get_ratios_by_loss) fuction： + + +```python +loss = 0.01 +ratios = slim.prune.get_ratios_by_loss(s_0, loss) +print(ratios) +``` + +### 8.2 Pruning training network + + +```python +pruner = slim.prune.Pruner() +print("FLOPs before pruning: {}".format(slim.analysis.flops(train_program))) +pruned_program, _, _ = pruner.prune( + train_program, + fluid.global_scope(), + params=ratios.keys(), + ratios=ratios.values(), + place=place) +print("FLOPs after pruning: {}".format(slim.analysis.flops(pruned_program))) +``` + +### 8.3 Pruning test network + +Note：The `only_graph` should be set to True while pruning test network. [Pruner API](https://paddlepaddle.github.io/PaddleSlim/api/prune_api/#pruner) + + +```python +pruner = slim.prune.Pruner() +print("FLOPs before pruning: {}".format(slim.analysis.flops(val_program))) +pruned_val_program, _, _ = pruner.prune( + val_program, + fluid.global_scope(), + params=ratios.keys(), + ratios=ratios.values(), + place=place, + only_graph=True) +print("FLOPs after pruning: {}".format(slim.analysis.flops(pruned_val_program))) +``` + +Get accuracy of pruned model on test dataset: + +```python +test(pruned_val_program) +``` + +### 8.4 Training pruned model + +Training pruned model: + + +```python +for data in train_reader(): + acc1, acc5, loss = exe.run(pruned_program, feed=data_feeder.feed(data), fetch_list=outputs) +print(np.mean(acc1), np.mean(acc5), np.mean(loss)) +``` + +Get accuracy of model after training: + +```python +test(pruned_val_program) +``` diff --git a/_sources/tutorials/index.rst.txt b/_sources/tutorials/index.rst.txt index bc4a29ea..e6109b73 100644 --- a/_sources/tutorials/index.rst.txt +++ b/_sources/tutorials/index.rst.txt @@ -3,9 +3,13 @@ ======== .. toctree:: - :maxdepth: 2 - :caption: Contents: + :maxdepth: 1 image_classification_sensitivity_analysis_tutorial.md - image_classification_nas_quick_start.ipynb - + darts_nas_turorial.md + paddledetection_slim_distillation_tutorial.md + paddledetection_slim_nas_tutorial.md + paddledetection_slim_pruing_tutorial.md + paddledetection_slim_prune_dist_tutorial.md + paddledetection_slim_quantization_tutorial.md + paddledetection_slim_sensitivy_tutorial.md diff --git a/_sources/tutorials/index_en.rst.txt b/_sources/tutorials/index_en.rst.txt index 44241358..10bdbd38 100644 --- a/_sources/tutorials/index_en.rst.txt +++ b/_sources/tutorials/index_en.rst.txt @@ -4,5 +4,5 @@ Aadvanced Tutorials .. toctree:: :maxdepth: 1 - sensitivity_tutorial_en.md + image_classification_sensitivity_analysis_tutorial_en.md diff --git a/_sources/tutorials/paddledetection_slim_distillation_tutorial.md.txt b/_sources/tutorials/paddledetection_slim_distillation_tutorial.md.txt new file mode 100644 index 00000000..dbdd30ff --- /dev/null +++ b/_sources/tutorials/paddledetection_slim_distillation_tutorial.md.txt @@ -0,0 +1,32 @@ +# 目标检测模型蒸馏教程 + +教程内容请参考：https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.2/slim/distillation/README.md + + +## 示例结果 + +### MobileNetV1-YOLO-V3-VOC + +| FLOPS |输入尺寸|每张GPU图片个数|推理时间（fps）|Box AP|下载| +|:-:|:-:|:-:|:-:|:-:|:-:| +|baseline|608 |16|104.291|76.2|[下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar)| +|蒸馏后|608 |16|106.914|79.0|[下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar)| +|baseline|416 |16|-|76.7|[下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar)| +|蒸馏后|416 |16|-|78.2|[下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar)| +|baseline|320 |16|-|75.3|[下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar)| +|蒸馏后|320 |16|-|75.5|[下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_voc_distilled.tar)| + +> 蒸馏后的结果用ResNet34-YOLO-V3做teacher，4GPU总batch_size64训练90000 iter得到 + +### MobileNetV1-YOLO-V3-COCO + +| FLOPS |输入尺寸|每张GPU图片个数|推理时间（fps）|Box AP|下载| +|:-:|:-:|:-:|:-:|:-:|:-:| +|baseline|608 |16|78.302|29.3|[下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar)| +|蒸馏后|608 |16|78.523|31.4|[下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar)| +|baseline|416 |16|-|29.3|[下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar)| +|蒸馏后|416 |16|-|30.0|[下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar)| +|baseline|320 |16|-|27.0|[下载链接](https://paddlemodels.bj.bcebos.com/object_detection/yolov3_mobilenet_v1_voc.tar)| +|蒸馏后|320 |16|-|27.1|[下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_distilled.tar)| + +> 蒸馏后的结果用ResNet34-YOLO-V3做teacher，4GPU总batch_size64训练600000 iter得到 diff --git a/_sources/tutorials/paddledetection_slim_nas_tutorial.md.txt b/_sources/tutorials/paddledetection_slim_nas_tutorial.md.txt new file mode 100644 index 00000000..b4aa6ad1 --- /dev/null +++ b/_sources/tutorials/paddledetection_slim_nas_tutorial.md.txt @@ -0,0 +1,8 @@ +# 人脸检测模型小模型结构搜索教程 + +教程内容请参考：https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.2/slim/nas/README.md + +## 概述 + +我们选取人脸检测的BlazeFace模型作为神经网络搜索示例，该示例使用PaddleSlim 辅助完成神经网络搜索实验。 +基于PaddleSlim进行搜索实验过程中，搜索限制条件可以选择是浮点运算数(FLOPs)限制还是硬件延时(latency)限制，硬件延时限制需要提供延时表。本示例提供一份基于blazeface搜索空间的硬件延时表，名称是latency_855.txt(基于PaddleLite在骁龙855上测试的延时)，可以直接用该表进行blazeface的硬件延时搜索实验。 diff --git a/_sources/tutorials/paddledetection_slim_pruing_tutorial.md.txt b/_sources/tutorials/paddledetection_slim_pruing_tutorial.md.txt new file mode 100644 index 00000000..defd371c --- /dev/null +++ b/_sources/tutorials/paddledetection_slim_pruing_tutorial.md.txt @@ -0,0 +1,3 @@ +# 目标检测模型卷积通道剪裁教程 + +请参考：https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.2/slim/prune/README.md diff --git a/_sources/tutorials/paddledetection_slim_prune_dist_tutorial.md.txt b/_sources/tutorials/paddledetection_slim_prune_dist_tutorial.md.txt new file mode 100644 index 00000000..5697e968 --- /dev/null +++ b/_sources/tutorials/paddledetection_slim_prune_dist_tutorial.md.txt @@ -0,0 +1,7 @@ +# 目标检测模型蒸馏剪裁教程 + +教程内容请参考：https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.2/slim/extensions/distill_pruned_model/README.md + +## 概述 + +该文档介绍如何使用PaddleSlim的蒸馏接口和卷积通道剪裁接口对检测库中的模型进行卷积层的通道剪裁并使用较高精度模型对其蒸馏。 diff --git a/_sources/tutorials/paddledetection_slim_quantization_tutorial.md.txt b/_sources/tutorials/paddledetection_slim_quantization_tutorial.md.txt new file mode 100644 index 00000000..c3ea66ce --- /dev/null +++ b/_sources/tutorials/paddledetection_slim_quantization_tutorial.md.txt @@ -0,0 +1,28 @@ +# 目标检测模型定点量化教程 + +教程内容请参考：https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.2/slim/quantization/README.md + + +## 示例结果 + +### 训练策略 + +- 量化策略`post`为使用离线量化得到的模型，`aware`为在线量化训练得到的模型。 + +### YOLOv3 on COCO + +| 骨架网络 | 预训练权重 | 量化策略 | 输入尺寸 | Box AP | 下载 | +| :----------------| :--------: | :------: | :------: |:------: | :-----------------------------------------------------: | +| MobileNetV1 | ImageNet | post | 608 | 27.9 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | +| MobileNetV1 | ImageNet | post | 416 | 28.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | +| MobileNetV1 | ImageNet | post | 320 | 26.0 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_post.tar) | +| MobileNetV1 | ImageNet | aware | 608 | 28.1 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) | +| MobileNetV1 | ImageNet | aware | 416 | 28.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) | +| MobileNetV1 | ImageNet | aware | 320 | 25.8 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_mobilenetv1_coco_quant_aware.tar) | +| ResNet34 | ImageNet | post | 608 | 35.7 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_post.tar) | +| ResNet34 | ImageNet | aware | 608 | 35.2 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | +| ResNet34 | ImageNet | aware | 416 | 33.3 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | +| ResNet34 | ImageNet | aware | 320 | 30.3 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r34_coco_quant_aware.tar) | +| R50vd-dcn | object365 | aware | 608 | 40.6 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | +| R50vd-dcn | object365 | aware | 416 | 37.5 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | +| R50vd-dcn | object365 | aware | 320 | 34.1 | [下载链接](https://paddlemodels.bj.bcebos.com/PaddleSlim/yolov3_r50vd_dcn_obj365_pretrained_coco_quant_aware.tar) | diff --git a/_sources/tutorials/paddledetection_slim_sensitivy_tutorial.md.txt b/_sources/tutorials/paddledetection_slim_sensitivy_tutorial.md.txt new file mode 100644 index 00000000..5e2c7bd5 --- /dev/null +++ b/_sources/tutorials/paddledetection_slim_sensitivy_tutorial.md.txt @@ -0,0 +1,3 @@ +# 目标检测模型敏感度分析教程 + +教程内容请参考：https://github.com/PaddlePaddle/PaddleDetection/blob/release/0.2/slim/sensitive/README.md diff --git a/algo/algo.html b/algo/algo.html index 6c5543e7..359f04b2 100644 --- a/algo/algo.html +++ b/algo/algo.html @@ -81,7 +81,7 @@
  • Model Zoo
    • diff --git a/api_en/search_space_en.html b/api_en/search_space_en.html index 8a662778..b3866466 100644 --- a/api_en/search_space_en.html +++ b/api_en/search_space_en.html @@ -35,7 +35,7 @@ - + @@ -101,6 +101,7 @@
  • How to write yourself search space
    • +
  • Table about hardware lantency
  • Model Zoo
    • @@ -280,7 +281,7 @@ return bn