提交 6fcca493 编写于 作者: D dongshuilong

Merge branch 'develop' of https://github.com/PaddlePaddle/PaddleClas into deploy_doc

...@@ -53,52 +53,65 @@ Res2Net200_vd预训练模型Top-1精度高达85.1%。 ...@@ -53,52 +53,65 @@ Res2Net200_vd预训练模型Top-1精度高达85.1%。
PP-ShiTu图像识别快速体验:[点击这里](./docs/zh_CN/quick_start/quick_start_recognition.md) PP-ShiTu图像识别快速体验:[点击这里](./docs/zh_CN/quick_start/quick_start_recognition.md)
## 文档教程 ## 文档教程
- 安装说明 - [环境准备]()
- [安装Paddle](./docs/zh_CN/installation/install_paddle.md) - [PULC超轻量图像分类实用方案]()
- [安装PaddleClas](./docs/zh_CN/installation/install_paddleclas.md) - [超轻量图像分类快速体验 (@崔程)]()
- 快速体验 - [超轻量图像分类模型库](包含benchmark @崔程)
- [PP-ShiTu图像识别快速体验](./docs/zh_CN/quick_start/quick_start_recognition.md) - xx
- 图像分类快速体验 - [方案介绍和模型训练]()
- [尝鲜版](./docs/zh_CN/quick_start/quick_start_classification_new_user.md) - [推理部署](@水龙)
- [进阶版](./docs/zh_CN/quick_start/quick_start_classification_professional.md) - 基于python预测引擎推理
- [多标签分类](./docs/zh_CN/quick_start/quick_start_multilabel_classification.md) - 基于C++预测引擎推理
- 服务化部署
- 端侧部署
- Paddle2ONNX模型转化与预测
- [模型压缩](@崔程)
- [PP-ShiTu图像识别系统介绍](#图像识别系统介绍) - [PP-ShiTu图像识别系统介绍](#图像识别系统介绍)
- 图像识别快速体验
- 模块介绍
- [主体检测](./docs/zh_CN/image_recognition_pipeline/mainbody_detection.md) - [主体检测](./docs/zh_CN/image_recognition_pipeline/mainbody_detection.md)
- [特征提取](./docs/zh_CN/image_recognition_pipeline/feature_extraction.md) - [特征提取模型](./docs/zh_CN/image_recognition_pipeline/feature_extraction.md)
- [向量检索](./docs/zh_CN/image_recognition_pipeline/vector_search.md) - [向量检索](./docs/zh_CN/image_recognition_pipeline/vector_search.md)
- [骨干网络和预训练模型库](./docs/zh_CN/algorithm_introduction/ImageNet_models.md) - 哈希编码
- 数据准备 - 模型训练(包含数据集格式说明等)
- [图像分类数据集介绍](./docs/zh_CN/data_preparation/classification_dataset.md) - 推理部署
- [图像识别数据集介绍](./docs/zh_CN/data_preparation/recognition_dataset.md) - 基于python预测引擎推理
- 模型训练 - 基于C++预测引擎推理
- [图像分类任务](./docs/zh_CN/models_training/classification.md)
- [图像识别任务](./docs/zh_CN/models_training/recognition.md)
- [训练参数调整策略](./docs/zh_CN/models_training/train_strategy.md)
- [配置文件说明](./docs/zh_CN/models_training/config_description.md)
- 模型预测部署
- [模型导出](./docs/zh_CN/inference_deployment/export_model.md)
- Python/C++ 预测引擎
- [基于Python预测引擎预测推理](./docs/zh_CN/inference_deployment/python_deploy.md)
- [基于C++分类预测引擎预测推理](./docs/zh_CN/inference_deployment/cpp_deploy.md)[基于C++的PP-ShiTu预测引擎预测推理](deploy/cpp_shitu/readme.md)
- 服务化部署 - 服务化部署
- [Paddle Serving服务化部署(推荐)](./docs/zh_CN/inference_deployment/paddle_serving_deploy.md) - 端侧部署
- [Hub serving服务化部署](./docs/zh_CN/inference_deployment/paddle_hub_serving_deploy.md) - Paddle2ONNX模型转化与预测
- [端侧部署](./deploy/lite/readme.md) - 模型压缩
- [whl包预测](./docs/zh_CN/inference_deployment/whl_deploy.md) - 模型量化
- 算法介绍 - 模型裁剪
- [图像分类任务介绍](./docs/zh_CN/algorithm_introduction/image_classification.md) - [骨干网络和预训练模型库](./docs/zh_CN/algorithm_introduction/ImageNet_models.md)
- [度量学习介绍](./docs/zh_CN/algorithm_introduction/metric_learning.md) - PP系列骨干网络模型(包括算法介绍,使用,训推一体链接等)(@崔程)
- 高阶使用 - PP-HGNet
- [数据增广](./docs/zh_CN/advanced_tutorials/DataAugmentation.md) - PP-LCNet v2
- [模型量化](./docs/zh_CN/advanced_tutorials/model_prune_quantization.md) - PP-LCNet
- [知识蒸馏](./docs/zh_CN/advanced_tutorials/knowledge_distillation.md) - SSLD半监督知识蒸馏方案 (@若愚)
- [PaddleClas结构解析](./docs/zh_CN/advanced_tutorials/code_overview.md) - SSLD算法简介
- [社区贡献指南](./docs/zh_CN/advanced_tutorials/how_to_contribute.md) - 预训练模型库
- 使用方法(?)
- 前沿算法
- 骨干网络和预训练模型库 (@崔程)
- 服务端CNN模型库
- 移动端CNN模型库
- Vision Transformer模型库
- 度量学习(arcmargin等算法)(@水龙)
- ReID (@水龙)
- 向量检索 (@水龙)
- 哈希特征 (@水龙)
- 模型蒸馏 (@若愚)
- 数据增强 (@崔程)
- 产业实用范例库 (@胜禹)
- 30分钟快速体验图像分类(原尝鲜版)(@崔程)
- FAQ - FAQ
- [图像识别精选问题](docs/zh_CN/faq_series/faq_2021_s2.md) - [图像识别精选问题](docs/zh_CN/faq_series/faq_2021_s2.md)
- [图像分类精选问题](docs/zh_CN/faq_series/faq_selected_30.md) - [图像分类精选问题](docs/zh_CN/faq_series/faq_selected_30.md)
- [图像分类FAQ第一季](docs/zh_CN/faq_series/faq_2020_s1.md) - [图像分类FAQ第一季](docs/zh_CN/faq_series/faq_2020_s1.md)
- [图像分类FAQ第二季](docs/zh_CN/faq_series/faq_2021_s1.md) - [图像分类FAQ第二季](docs/zh_CN/faq_series/faq_2021_s1.md)
- [PaddleClas结构解析](./docs/zh_CN/advanced_tutorials/code_overview.md)
- [社区贡献指南](./docs/zh_CN/advanced_tutorials/how_to_contribute.md)
- [许可证书](#许可证书) - [许可证书](#许可证书)
- [贡献代码](#贡献代码) - [贡献代码](#贡献代码)
......
Global: Global:
infer_imgs: "./images/PULC/person/objects365_02035329.jpg" infer_imgs: "./images/PULC/person_exists/objects365_02035329.jpg"
inference_model_dir: "./models/person_cls_infer" inference_model_dir: "./models/person_exists_infer"
batch_size: 1 batch_size: 1
use_gpu: True use_gpu: True
enable_mkldnn: False enable_mkldnn: False
......
Global:
infer_imgs: "./images/PULC/traffic_sign/99603_17806.jpg"
inference_model_dir: "./models/traffic_sign_infer"
batch_size: 1
use_gpu: True
enable_mkldnn: True
cpu_num_threads: 10
benchmark: False
use_fp16: False
ir_optim: True
use_tensorrt: False
gpu_mem: 8000
enable_profile: False
PreProcess:
transform_ops:
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 0.00392157
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
channel_num: 3
- ToCHWImage:
PostProcess:
main_indicator: Topk
Topk:
topk: 5
class_id_map_file: "../dataset/traffic_sign/label_name_id.txt"
SavePreLabel:
save_dir: ./pre_label/
Global:
infer_imgs: "./images/PULC/vehicle_attr/0002_c002_00030670_0.jpg"
inference_model_dir: "./models/vehicle_attr_infer"
batch_size: 1
use_gpu: True
enable_mkldnn: True
cpu_num_threads: 10
benchmark: False
use_fp16: False
ir_optim: True
use_tensorrt: False
gpu_mem: 8000
enable_profile: False
PreProcess:
transform_ops:
- ResizeImage:
size: [256, 192]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
channel_num: 3
- ToCHWImage:
PostProcess:
main_indicator: VehicleAttribute
VehicleAttribute:
color_threshold: 0.5
type_threshold: 0.5
...@@ -30,4 +30,4 @@ op: ...@@ -30,4 +30,4 @@ op:
client_type: local_predictor client_type: local_predictor
#Fetch结果列表,以client_config中fetch_var的alias_name为准 #Fetch结果列表,以client_config中fetch_var的alias_name为准
fetch_list: ["prediction"] fetch_list: ["prediction"]
...@@ -31,7 +31,7 @@ op: ...@@ -31,7 +31,7 @@ op:
#Fetch结果列表,以client_config中fetch_var的alias_name为准 #Fetch结果列表,以client_config中fetch_var的alias_name为准
fetch_list: ["features"] fetch_list: ["features"]
det: det:
concurrency: 1 concurrency: 1
local_service_conf: local_service_conf:
......
...@@ -280,3 +280,45 @@ class Attribute(object): ...@@ -280,3 +280,45 @@ class Attribute(object):
batch_res.append([label_res, pred_res]) batch_res.append([label_res, pred_res])
return batch_res return batch_res
class VehicleAttribute(object):
def __init__(self, color_threshold=0.5, type_threshold=0.5):
self.color_threshold = color_threshold
self.type_threshold = type_threshold
self.color_list = [
"yellow", "orange", "green", "gray", "red", "blue", "white",
"golden", "brown", "black"
]
self.type_list = [
"sedan", "suv", "van", "hatchback", "mpv", "pickup", "bus",
"truck", "estate"
]
def __call__(self, batch_preds, file_names=None):
# postprocess output of predictor
batch_res = []
for res in batch_preds:
res = res.tolist()
label_res = []
color_idx = np.argmax(res[:10])
type_idx = np.argmax(res[10:])
if res[color_idx] >= self.color_threshold:
color_info = f"Color: ({self.color_list[color_idx]}, prob: {res[color_idx]})"
else:
color_info = "Color unknown"
if res[type_idx + 10] >= self.type_threshold:
type_info = f"Type: ({self.type_list[type_idx]}, prob: {res[type_idx + 10]})"
else:
type_info = "Type unknown"
label_res = f"{color_info}, {type_info}"
threshold_list = [self.color_threshold
] * 10 + [self.type_threshold] * 9
pred_res = (np.array(res) > np.array(threshold_list)
).astype(np.int8).tolist()
batch_res.append([label_res, pred_res])
return batch_res
...@@ -138,7 +138,9 @@ def main(config): ...@@ -138,7 +138,9 @@ def main(config):
continue continue
batch_results = cls_predictor.predict(batch_imgs) batch_results = cls_predictor.predict(batch_imgs)
for number, result_dict in enumerate(batch_results): for number, result_dict in enumerate(batch_results):
if "Attribute" in config["PostProcess"]: if "Attribute" in config[
"PostProcess"] or "VehicleAttribute" in config[
"PostProcess"]:
filename = batch_names[number] filename = batch_names[number]
attr_message = result_dict[0] attr_message = result_dict[0]
pred_res = result_dict[1] pred_res = result_dict[1]
......
# PaddleClas构建有人/无人分类案例
此处提供了用户使用 PaddleClas 快速构建轻量级、高精度、可落地的有人/无人的分类模型教程,主要基于有人/无人场景的数据,融合了轻量级骨干网络PPLCNet、SSLD预训练权重、EDA数据增强策略、SKL-UGI知识蒸馏策略、SHAS超参数搜索策略,得到精度高、速度快、易于部署的二分类模型。
------
## 目录
- [1. 环境配置](#1)
- [2. 有人/无人场景推理预测](#2)
- [2.1 下载模型](#2.1)
- [2.2 模型推理预测](#2.2)
- [2.2.1 预测单张图像](#2.2.1)
- [2.2.2 基于文件夹的批量预测](#2.2.2)
- [3.有人/无人场景训练](#3)
- [3.1 数据准备](#3.1)
- [3.2 模型训练](#3.2)
- [3.2.1 基于默认超参数训练](#3.2.1)
- [3.2.1.1 基于默认超参数训练轻量级模型](#3.2.1.1)
- [3.2.1.2 基于默认超参数训练教师模型](#3.2.1.2)
- [3.2.1.3 基于默认超参数进行蒸馏训练](#3.2.1.3)
- [3.2.2 超参数搜索训练](#3.2)
- [4. 模型评估与推理](#4)
- [4.1 模型评估](#3.1)
- [4.2 模型预测](#3.2)
- [4.3 使用 inference 模型进行推理](#4.3)
- [4.3.1 导出 inference 模型](#4.3.1)
- [4.3.2 模型推理预测](#4.3.2)
<a name="1"></a>
## 1. 环境配置
* 安装:请先参考 [Paddle 安装教程](../installation/install_paddle.md) 以及 [PaddleClas 安装教程](../installation/install_paddleclas.md) 配置 PaddleClas 运行环境。
<a name="2"></a>
## 2. 有人/无人场景推理预测
<a name="2.1"></a>
### 2.1 下载模型
* 进入 `deploy` 运行目录。
```
cd deploy
```
下载有人/无人分类的模型。
```
mkdir models
cd models
# 下载inference 模型并解压
wget https://paddleclas.bj.bcebos.com/models/PULC/person_cls_infer.tar && tar -xf person_cls_infer.tar
```
解压完毕后,`models` 文件夹下应有如下文件结构:
```
├── person_cls_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
<a name="2.2"></a>
### 2.2 模型推理预测
<a name="2.2.1"></a>
#### 2.2.1 预测单张图像
返回 `deploy` 目录:
```
cd ../
```
运行下面的命令,对图像 `./images/PULC/person/objects365_02035329.jpg` 进行有人/无人分类。
```shell
# 使用下面的命令使用 GPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/person/inference_person_cls.yaml -o PostProcess.ThreshOutput.threshold=0.9794
# 使用下面的命令使用 CPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/person/inference_person_cls.yaml -o PostProcess.ThreshOutput.threshold=0.9794 -o Global.use_gpu=False
```
输出结果如下。
```
objects365_02035329.jpg: class id(s): [1], score(s): [1.00], label_name(s): ['someone']
```
**备注:** 真实场景中往往需要在假正类率(Fpr)小于某一个指标下求真正类率(Tpr),该场景中的`val`数据集在千分之一Fpr下得到的最佳Tpr所得到的阈值为`0.9794`,故此处的`threshold``0.9794`。该阈值的确定方法可以参考[3.2节](#3.2)
<a name="2.2.2"></a>
#### 2.2.2 基于文件夹的批量预测
如果希望预测文件夹内的图像,可以直接修改配置文件中的 `Global.infer_imgs` 字段,也可以通过下面的 `-o` 参数修改对应的配置。
```shell
# 使用下面的命令使用 GPU 进行预测,如果希望使用 CPU 预测,可以在命令后面添加 -o Global.use_gpu=False
python3.7 python/predict_cls.py -c configs/PULC/person/inference_person_cls.yaml -o Global.infer_imgs="./images/PULC/person/"
```
终端中会输出该文件夹内所有图像的分类结果,如下所示。
```
objects365_01780782.jpg: class id(s): [0], score(s): [1.00], label_name(s): ['nobody']
objects365_02035329.jpg: class id(s): [1], score(s): [1.00], label_name(s): ['someone']
```
其中,`someone` 表示该图里存在人,`nobody` 表示该图里不存在人。
<a name="3"></a>
## 3.有人/无人场景训练
<a name="3.1"></a>
### 3.1 数据准备
进入 PaddleClas 目录。
```
cd path_to_PaddleClas
```
进入 `dataset/` 目录,下载并解压有人/无人场景的数据。
```shell
cd dataset
wget https://paddleclas.bj.bcebos.com/data/cls_demo/person.tar
tar -xf person.tar
cd ../
```
执行上述命令后,`dataset/`下存在`person`目录,该目录中具有以下数据:
```
├── train
│   ├── 000000000009.jpg
│   ├── 000000000025.jpg
...
├── val
│   ├── objects365_01780637.jpg
│   ├── objects365_01780640.jpg
...
├── ImageNet_val
│   ├── ILSVRC2012_val_00000001.JPEG
│   ├── ILSVRC2012_val_00000002.JPEG
...
├── train_list.txt
├── train_list.txt.debug
├── train_list_for_distill.txt
├── val_list.txt
└── val_list.txt.debug
```
其中`train/``val/`分别为训练集和验证集。`train_list.txt``val_list.txt`分别为训练集和验证集的标签文件,`train_list.txt.debug``val_list.txt.debug`分别为训练集和验证集的`debug`标签文件,其分别是`train_list.txt``val_list.txt`的子集,用该文件可以快速体验本案例的流程。`ImageNet_val/`是ImageNet的验证集,该集合和`train`集合的混合数据用于本案例的`SKL-UGI知识蒸馏策略`,对应的训练标签文件为`train_list_for_distill.txt`
* **注意**:
* 本案例中所使用的所有数据集均为开源数据,`train`集合为[MS-COCO数据](https://cocodataset.org/#overview)的训练集的子集,`val`集合为[Object365数据](https://www.objects365.org/overview.html)的训练集的子集,`ImageNet_val`[ImageNet数据](https://www.image-net.org/)的验证集。数据集的筛选流程可以参考[有人/无人场景数据集筛选方法]()。
<a name="3.2"></a>
### 3.2 模型训练
<a name="3.2.1"></a>
#### 3.2.1 基于默认超参数训练
<a name="3.2.1.1"></a>
##### 3.2.1.1 基于默认超参数训练轻量级模型
`ppcls/configs/PULC/person/PPLCNet/PPLCNet_x1_0.yaml`中提供了基于该场景的训练配置,可以通过如下脚本启动训练:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/person/PPLCNet/PPLCNet_x1_0.yaml
```
验证集的最佳指标在0.94-0.95之间(数据集较小,容易造成波动)。
**备注:**
* 此时使用的指标为Tpr,该指标描述了在假正类率(Fpr)小于某一个指标时的真正类率(Tpr),是产业中二分类问题常用的指标之一。在本案例中,Fpr为千分之一。关于Fpr和Tpr的更多介绍,可以参考[这里](https://baike.baidu.com/item/AUC/19282953)
* 在eval时,会打印出来当前最佳的TprAtFpr指标,具体地,其会打印当前的`Fpr``Tpr`值,以及当前的`threshold`值,`Tpr`值反映了在当前`Fpr`值下的召回率,该值越高,代表模型越好。`threshold` 表示当前最佳`Fpr`所对应的分类阈值,可用于后续模型部署落地等。
<a name="3.2.1.2"></a>
##### 3.2.1.2 基于默认超参数训练教师模型
复用`ppcls/configs/PULC/person/PPLCNet/PPLCNet_x1_0.yaml`中的超参数,训练教师模型,训练脚本如下:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/person/PPLCNet/PPLCNet_x1_0.yaml \
-o Arch.name=ResNet101_vd
```
验证集的最佳指标为0.96-0.98之间,当前教师模型最好的权重保存在`output/ResNet101_vd/best_model.pdparams`
<a name="3.2.1.3"></a>
##### 3.2.1.3 基于默认超参数进行蒸馏训练
配置文件`ppcls/configs/PULC/PULC/Distillation/PPLCNet_x1_0_distillation.yaml`提供了`SKL-UGI知识蒸馏策略`的配置。该配置将`ResNet101_vd`当作教师模型,`PPLCNet_x1_0`当作学生模型,使用ImageNet数据集的验证集作为新增的无标签数据。训练脚本如下:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/person/Distillation/PPLCNet_x1_0_distillation.yaml \
-o Arch.models.0.Teacher.pretrained=output/ResNet101_vd/best_model
```
验证集的最佳指标为0.95-0.97之间,当前模型最好的权重保存在`output/DistillationModel/best_model_student.pdparams`
<a name="3.2.2"></a>
#### 3.2.2 超参数搜索训练
[3.2 小节](#3.2) 提供了在已经搜索并得到的超参数上进行了训练,此部分内容提供了搜索的过程,此过程是为了得到更好的训练超参数。
* 搜索运行脚本如下:
```shell
python tools/search_strategy.py -c ppcls/configs/StrategySearch/person.yaml
```
`ppcls/configs/StrategySearch/person.yaml`中指定了具体的 GPU id 号和搜索配置, 默认搜索的训练日志和模型存放于`output/search_person`中,最终的蒸馏模型存放于`output/search_person/search_res/DistillationModel/best_model_student.pdparams`
* **注意**:
* 3.1小节提供的默认配置已经经过了搜索,所以此过程不是必要的过程,如果自己的训练数据集有变化,可以尝试此过程。
* 此过程基于当前数据集在 V100 4 卡上大概需要耗时 10 小时,如果缺少机器资源,希望体验搜索过程,可以将`ppcls/configs/cls_demo/person/PPLCNet/PPLCNet_x1_0_search.yaml`中的`train_list.txt``val_list.txt`分别替换为`train_list.txt.debug``val_list.txt.debug`。替换list只是为了加速跑通整个搜索过程,由于数据量较小,其搜素的结果没有参考性。另外,搜索空间可以根据当前的机器资源来调整,如果机器资源有限,可以尝试缩小搜索空间,如果机器资源较充足,可以尝试扩大搜索空间。
* 如果此过程搜索的得到的超参数与[3.2.1小节](#3.2.1)提供的超参数不一致,主要是由于训练数据较小造成的波动导致,可以忽略。
<a name="4"></a>
## 4. 模型评估与推理
<a name="4.1"></a>
### 4.1 模型评估
训练好模型之后,可以通过以下命令实现对模型指标的评估。
```bash
python3 tools/eval.py \
-c ./ppcls/configs/PULC/person/PPLCNet/PPLCNet_x1_0.yaml \
-o Global.pretrained_model="output/DistillationModel/best_model_student"
```
<a name="4.2"></a>
### 4.2 模型预测
模型训练完成之后,可以加载训练得到的预训练模型,进行模型预测。在模型库的 `tools/infer.py` 中提供了完整的示例,只需执行下述命令即可完成模型预测:
```python
python3 tools/infer.py \
-c ./ppcls/configs/PULC/person/PPLCNet/PPLCNet_x1_0.yaml \
-o Infer.infer_imgs=./dataset/person/val/objects365_01780637.jpg \
-o Global.pretrained_model=output/DistillationModel/best_model_student \
-o Global.pretrained_model=Infer.PostProcess.threshold=0.9794
```
输出结果如下:
```
[{'class_ids': [0], 'scores': [0.9878496769815683], 'label_names': ['nobody'], 'file_name': './dataset/person/val/objects365_01780637.jpg'}]
```
**备注:** 这里的`Infer.PostProcess.threshold`的值需要根据实际场景来确定,此处的`0.9794`是在该场景中的`val`数据集在千分之一Fpr下得到的最佳Tpr所得到的。
<a name="4.3"></a>
### 4.3 使用 inference 模型进行推理
<a name="4.3.1"></a>
### 4.3.1 导出 inference 模型
通过导出 inference 模型,PaddlePaddle 支持使用预测引擎进行预测推理。接下来介绍如何用预测引擎进行推理:
首先,对训练好的模型进行转换:
```bash
python3 tools/export_model.py \
-c ./ppcls/configs/cls_demo/PULC/PPLCNet/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/DistillationModel/best_model_student \
-o Global.save_inference_dir=deploy/models/PPLCNet_x1_0_person
```
执行完该脚本后会在`deploy/models/`下生成`PPLCNet_x1_0_person`文件夹,该文件夹中的模型与 2.2 节下载的推理预测模型格式一致。
<a name="4.3.2"></a>
### 4.3.2 基于 inference 模型推理预测
推理预测的脚本为:
```
python3.7 python/predict_cls.py -c configs/PULC/person/inference_person_cls.yaml -o Global.inference_model_dir="models/PPLCNet_x1_0_person" -o PostProcess.ThreshOutput.threshold=0.9794
```
**备注:**
- 此处的`PostProcess.ThreshOutput.threshold`由eval时的最佳`threshold`来确定。
- 更多关于推理的细节,可以参考[2.2节](#2.2)
# PULC 有人/无人分类模型
------
## 目录
- [1. 模型和应用场景介绍](#1)
- [2. 模型快速体验](#2)
- [3. 模型训练、评估和预测](#3)
- [3.1 环境配置](#3.1)
- [3.2 数据准备](#3.2)
- [3.2.1 数据集来源](#3.2.1)
- [3.2.2 数据集获取](#3.2.2)
- [3.3 模型训练](#3.3)
- [3.4 模型评估](#3.4)
- [3.5 模型预测](#3.5)
- [4. 模型压缩](#4)
- [4.1 SKL-UGI 知识蒸馏](#4.1)
- [4.1.1 教师模型训练](#4.1.1)
- [4.1.2 蒸馏训练](#4.1.2)
- [5. 超参搜索](#5)
- [6. 模型推理部署](#6)
- [6.1 推理模型准备](#6.1)
- [6.1.1 基于训练得到的权重导出 inference 模型](#6.1.1)
- [6.1.2 直接下载 inference 模型](#6.1.2)
- [6.2 基于 Python 预测引擎推理](#6.2)
- [6.2.1 预测单张图像](#6.2.1)
- [6.2.2 基于文件夹的批量预测](#6.2.2)
- [6.3 基于 C++ 预测引擎推理](#6.3)
- [6.4 服务化部署](#6.4)
- [6.5 端侧部署](#6.5)
- [6.6 Paddle2ONNX 模型转换与预测](#6.6)
<a name="1"></a>
## 1. 模型和应用场景介绍
该案例提供了用户使用 PaddleClas 的超轻量图像分类方案(PULC,Practical Ultra Lightweight Classification)快速构建轻量级、高精度、可落地的有人/无人的分类模型。该模型可以广泛应用于如监控场景、人员进出管控场景、海量数据过滤场景等。
下表列出了判断图片中是否有人的二分类模型的相关指标,前两行展现了使用 SwinTranformer_tiny 和 MobileNetV3_large_x1_0 作为 backbone 训练得到的模型的相关指标,第三行至第六行依次展现了替换 backbone 为 PPLCNet_x1_0、使用 SSLD 预训练模型、使用 SSLD 预训练模型 + EDA 策略、使用 SSLD 预训练模型 + EDA 策略 + SKL-UGI 知识蒸馏策略训练得到的模型的相关指标。
| 模型 | Tpr(%) | 延时(ms) | 存储(M) | 策略 |
|-------|-----------|----------|---------------|---------------|
| SwinTranformer_tiny | 95.69 | 175.52 | 107 | 使用ImageNet预训练模型 |
| MobileNetV3_large_x1_0 | 91.97 | 4.70 | 17 | 使用ImageNet预训练模型 |
| PPLCNet_x1_0 | 89.57 | 2.36 | 6.5 | 使用ImageNet预训练模型 |
| PPLCNet_x1_0 | 92.10 | 2.36 | 6.5 | 使用SSLD预训练模型 |
| PPLCNet_x1_0 | 93.43 | 2.36 | 6.5 | 使用SSLD预训练模型+EDA策略|
| <b>PPLCNet_x1_0<b> | <b>95.60<b> | <b>2.36<b> | <b>6.5<b> | 使用SSLD预训练模型+EDA策略+SKL-UGI知识蒸馏策略|
从表中可以看出,backbone 为 SwinTranformer_tiny 时精度较高,但是推理速度较慢。将 backboone 替换为轻量级模型 MobileNetV3_large_x1_0 后,速度可以大幅提升,但是精度下降明显。将 backbone 替换为 PPLCNet_x1_0 时,精度较 MobileNetV3_large_x1_0 低两个多百分点,但是速度提升 2 倍左右。在此基础上,使用 SSLD 预训练模型后,在不改变推理速度的前提下,精度可以提升约 2.6 个百分点,进一步地,当融合EDA策略后,精度可以再提升 1.3 个百分点,最后,在使用 SKL-UGI 知识蒸馏后,精度可以继续提升 2.2 个百分点。此时,PPLCNet_x1_0 达到了 SwinTranformer_tiny 模型的精度,但是速度快 70+ 倍。关于 PULC 的训练方法和推理部署方法将在下面详细介绍。
**备注:**
* `Tpr`指标的介绍可以参考 [3.2 小节](#3.2)的备注部分,延时是基于 Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz 测试得到,开启MKLDNN加速策略,线程数为10。
* 关于PPLCNet的介绍可以参考[PPLCNet介绍](../models/PP-LCNet.md),相关论文可以查阅[PPLCNet paper](https://arxiv.org/abs/2109.15099)
<a name="2"></a>
## 2. 模型快速体验
(pip方式,待补充)
<a name="3"></a>
## 3. 模型训练、评估和预测
<a name="3.1"></a>
### 3.1 环境配置
* 安装:请先参考 [Paddle 安装教程](../installation/install_paddle.md) 以及 [PaddleClas 安装教程](../installation/install_paddleclas.md) 配置 PaddleClas 运行环境。
<a name="3.2"></a>
### 3.2 数据准备
<a name="3.2.1"></a>
#### 3.2.1 数据集来源
本案例中所使用的所有数据集均为开源数据,`train` 集合为[MS-COCO 数据](https://cocodataset.org/#overview)的训练集的子集,`val` 集合为[Object365 数据](https://www.objects365.org/overview.html)的训练集的子集,`ImageNet_val`[ImageNet-1k 数据](https://www.image-net.org/)的验证集。
<a name="3.2.2"></a>
#### 3.2.2 数据集获取
在公开数据集的基础上经过后处理即可得到本案例需要的数据,具体处理方法如下:
- 训练集合,本案例处理了 MS-COCO 数据训练集的标注文件,如果某张图含有“人”的标签,且这个框的面积在整张图中的比例大于 10%,即认为该张图中含有人,如果某张图中没有“人”的标签,则认为该张图中不含有人。经过处理后,得到 92964 条可用数据,其中有人的数据有 39813 条,无人的数据 53151 条。
- 验证集合,从 Object365 数据中随机抽取一小部分数据,使用在 MS-COCO 上训练得到的较好的模型预测这些数据,将预测结果和数据的标注文件取交集,将交集的结果按照得到训练集的方法筛选出验证集合。经过处理后,得到 27820 条可用数据。其中有人的数据有 2255 条,无人的数据有 25565 条。
处理后的数据集部分数据可视化如下:
![](../../images/PULC/docs/person_exists_data_demo.png)
此处提供了经过上述方法处理好的数据,可以直接下载得到。
进入 PaddleClas 目录。
```
cd path_to_PaddleClas
```
进入 `dataset/` 目录,下载并解压有人/无人场景的数据。
```shell
cd dataset
wget https://paddleclas.bj.bcebos.com/data/PULC/person_exists.tar
tar -xf person_exists.tar
cd ../
```
执行上述命令后,`dataset/` 下存在 `person_exists` 目录,该目录中具有以下数据:
```
├── train
│   ├── 000000000009.jpg
│   ├── 000000000025.jpg
...
├── val
│   ├── objects365_01780637.jpg
│   ├── objects365_01780640.jpg
...
├── ImageNet_val
│   ├── ILSVRC2012_val_00000001.JPEG
│   ├── ILSVRC2012_val_00000002.JPEG
...
├── train_list.txt
├── train_list.txt.debug
├── train_list_for_distill.txt
├── val_list.txt
└── val_list.txt.debug
```
其中 `train/``val/` 分别为训练集和验证集。`train_list.txt``val_list.txt` 分别为训练集和验证集的标签文件,`train_list.txt.debug``val_list.txt.debug` 分别为训练集和验证集的 `debug` 标签文件,其分别是 `train_list.txt``val_list.txt` 的子集,用该文件可以快速体验本案例的流程。`ImageNet_val/` 是 ImageNet-1k 的验证集,该集合和 `train` 集合的混合数据用于本案例的 `SKL-UGI知识蒸馏策略`,对应的训练标签文件为 `train_list_for_distill.txt`
**备注:**
* 关于 `train_list.txt``val_list.txt`的格式说明,可以参考[PaddleClas分类数据集格式说明](../data_preparation/classification_dataset.md#1-数据集格式说明)
* 关于如何得到蒸馏的标签文件可以参考[知识蒸馏标签获得方法](@ruoyu)
<a name="3.3"></a>
### 3.3 模型训练
`ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml` 中提供了基于该场景的训练配置,可以通过如下脚本启动训练:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml
```
验证集的最佳指标在 `0.94-0.95` 之间(数据集较小,容易造成波动)。
**备注:**
* 此时使用的指标为Tpr,该指标描述了在假正类率(Fpr)小于某一个指标时的真正类率(Tpr),是产业中二分类问题常用的指标之一。在本案例中,Fpr 为千分之一。关于 Fpr 和 Tpr 的更多介绍,可以参考[这里](https://baike.baidu.com/item/AUC/19282953)
* 在eval时,会打印出来当前最佳的 TprAtFpr 指标,具体地,其会打印当前的 `Fpr``Tpr` 值,以及当前的 `threshold`值,`Tpr` 值反映了在当前 `Fpr` 值下的召回率,该值越高,代表模型越好。`threshold` 表示当前最佳 `Fpr` 所对应的分类阈值,可用于后续模型部署落地等。
<a name="3.4"></a>
### 3.4 模型评估
训练好模型之后,可以通过以下命令实现对模型指标的评估。
```bash
python3 tools/eval.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml \
-o Global.pretrained_model="output/DPPLCNet_x1_0/best_model"
```
其中 `-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` 指定了当前最佳权重所在的路径,如果指定其他权重,只需替换对应的路径即可。
<a name="3.5"></a>
### 3.5 模型预测
模型训练完成之后,可以加载训练得到的预训练模型,进行模型预测。在模型库的 `tools/infer.py` 中提供了完整的示例,只需执行下述命令即可完成模型预测:
```python
python3 tools/infer.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/DistillationModel/best_model_student \
-o Global.pretrained_model=Infer.PostProcess.threshold=0.9794
```
输出结果如下:
```
[{'class_ids': [0], 'scores': [0.9878496769815683], 'label_names': ['nobody'], 'file_name': './dataset/person_exists/val/objects365_01780637.jpg'}]
```
**备注:**
* 这里`-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` 指定了当前最佳权重所在的路径,如果指定其他权重,只需替换对应的路径即可。
* 默认是对 `deploy/images/PULC/person_exists/objects365_02035329.jpg` 进行预测,此处也可以通过增加字段 `-o Infer.infer_imgs=xxx` 对其他图片预测。
* 这里的 `Infer.PostProcess.threshold` 的值需要根据实际场景来确定,此处的 `0.9794` 是在该场景中的 `val` 数据集在千分之一 Fpr 下得到的最佳 Tpr 所得到的。
<a name="4"></a>
## 4. 模型压缩
<a name="4.1"></a>
### 4.1 SKL-UGI 知识蒸馏
SKL-UGI 知识蒸馏是 PaddleClas 提出的一种简单有效的知识蒸馏方法,关于该方法的介绍,可以参考[SKL-UGI 知识蒸馏](@ruoyu)
<a name="4.1.1"></a>
#### 4.1.1 教师模型训练
复用 `ppcls/configs/PULC/person_exists/PPLCNet/PPLCNet_x1_0.yaml` 中的超参数,训练教师模型,训练脚本如下:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet/PPLCNet_x1_0.yaml \
-o Arch.name=ResNet101_vd
```
验证集的最佳指标为 `0.96-0.98` 之间,当前教师模型最好的权重保存在 `output/ResNet101_vd/best_model.pdparams`
<a name="4.1.2"></a>
#### 4.1.2 蒸馏训练
配置文件`ppcls/configs/PULC/person_exists/PPLCNet_x1_0_distillation.yaml`提供了`SKL-UGI知识蒸馏策略`的配置。该配置将`ResNet101_vd`当作教师模型,`PPLCNet_x1_0`当作学生模型,使用ImageNet数据集的验证集作为新增的无标签数据。训练脚本如下:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0_distillation.yaml \
-o Arch.models.0.Teacher.pretrained=output/ResNet101_vd/best_model
```
验证集的最佳指标为 `0.95-0.97` 之间,当前模型最好的权重保存在 `output/DistillationModel/best_model_student.pdparams`
<a name="5"></a>
## 5. 超参搜索
[3.2 节](#3.2)[4.1 节](#4.1)所使用的超参数是根据 PaddleClas 提供的 `SHAS 超参数搜索策略` 搜索得到的,如果希望在自己的数据集上得到更好的结果,可以参考[SHAS 超参数搜索策略](#TODO)来获得更好的训练超参数。
**备注:** 此部分内容是可选内容,搜索过程需要较长的时间,您可以根据自己的硬件情况来选择执行。如果没有更换数据集,可以忽略此节内容。
<a name="6"></a>
## 6. 模型推理部署
<a name="6.1"></a>
### 6.1 推理模型准备
Paddle Inference 是飞桨的原生推理库, 作用于服务器端和云端,提供高性能的推理能力。相比于直接基于预训练模型进行预测,Paddle Inference可使用MKLDNN、CUDNN、TensorRT 进行预测加速,从而实现更优的推理性能。更多关于Paddle Inference推理引擎的介绍,可以参考[Paddle Inference官网教程](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/infer/inference/inference_cn.html)
当使用 Paddle Inference 推理时,加载的模型类型为 inference 模型。本案例提供了两种获得 inference 模型的方法,如果希望得到和文档相同的结果,请选择[直接下载 inference 模型](#6.1.2)的方式。
<a name="6.1.1"></a>
### 6.1.1 基于训练得到的权重导出 inference 模型
此处,我们提供了将权重和模型转换的脚本,执行该脚本可以得到对应的 inference 模型:
```bash
python3 tools/export_model.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/DistillationModel/best_model_student \
-o Global.save_inference_dir=deploy/models/PPLCNet_x1_0_person_exists_infer
```
执行完该脚本后会在 `deploy/models/` 下生成 `PPLCNet_x1_0_person_exists_infer` 文件夹,`models` 文件夹下应有如下文件结构:
```
├── PPLCNet_x1_0_person_exists_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
**备注:** 此处的最佳权重是经过知识蒸馏后的权重路径,如果没有执行知识蒸馏的步骤,最佳模型保存在`output/PPLCNet_x1_0/best_model.pdparams`中。
<a name="6.1.2"></a>
### 6.1.2 直接下载 inference 模型
[6.1.1 小节](#6.1.1)提供了导出 inference 模型的方法,此处也提供了该场景可以下载的 inference 模型,可以直接下载体验。
```
cd deploy/models
# 下载 inference 模型并解压
wget https://paddleclas.bj.bcebos.com/models/PULC/person_exists_infer.tar && tar -xf person_exists_infer.tar
```
解压完毕后,`models` 文件夹下应有如下文件结构:
```
├── person_exists_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
<a name="6.2"></a>
### 6.2 基于 Python 预测引擎推理
<a name="6.2.1"></a>
#### 6.2.1 预测单张图像
返回 `deploy` 目录:
```
cd ../
```
运行下面的命令,对图像 `./images/PULC/person_exists/objects365_02035329.jpg` 进行有人/无人分类。
```shell
# 使用下面的命令使用 GPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/person_exists/inference_person_exists.yaml -o PostProcess.ThreshOutput.threshold=0.9794
# 使用下面的命令使用 CPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/person_exists/inference_person_exists.yaml -o PostProcess.ThreshOutput.threshold=0.9794 -o Global.use_gpu=False
```
输出结果如下。
```
objects365_02035329.jpg: class id(s): [1], score(s): [1.00], label_name(s): ['someone']
```
**备注:** 真实场景中往往需要在假正类率(Fpr)小于某一个指标下求真正类率(Tpr),该场景中的 `val` 数据集在千分之一 Fpr 下得到的最佳 Tpr 所得到的阈值为 `0.9794`,故此处的 `threshold``0.9794`。该阈值的确定方法可以参考[3.2节](#3.2)备注部分。
<a name="6.2.2"></a>
#### 6.2.2 基于文件夹的批量预测
如果希望预测文件夹内的图像,可以直接修改配置文件中的 `Global.infer_imgs` 字段,也可以通过下面的 `-o` 参数修改对应的配置。
```shell
# 使用下面的命令使用 GPU 进行预测,如果希望使用 CPU 预测,可以在命令后面添加 -o Global.use_gpu=False
python3.7 python/predict_cls.py -c configs/PULC/person_exists/inference_person_exists.yaml -o Global.infer_imgs="./images/PULC/person_exists/"
```
终端中会输出该文件夹内所有图像的分类结果,如下所示。
```
objects365_01780782.jpg: class id(s): [0], score(s): [1.00], label_name(s): ['nobody']
objects365_02035329.jpg: class id(s): [1], score(s): [1.00], label_name(s): ['someone']
```
其中,`someone` 表示该图里存在人,`nobody` 表示该图里不存在人。
<a name="6.3"></a>
### 6.3 基于 C++ 预测引擎推理
PaddleClas 提供了基于 C++ 预测引擎推理的示例,您可以参考[服务器端 C++ 预测](../inference_deployment/cpp_deploy.md)来完成相应的推理部署。如果您使用的是 Windows 平台,可以参考[基于 Visual Studio 2019 Community CMake 编译指南](../inference_deployment/cpp_deploy_on_windows.md)完成相应的预测库编译和模型预测工作。
<a name="6.4"></a>
### 6.4 服务化部署
Paddle Serving 提供高性能、灵活易用的工业级在线推理服务。Paddle Serving 支持 RESTful、gRPC、bRPC 等多种协议,提供多种异构硬件和多种操作系统环境下推理解决方案。更多关于Paddle Serving 的介绍,可以参考[Paddle Serving 代码仓库](https://github.com/PaddlePaddle/Serving)
PaddleClas 提供了基于 Paddle Serving 来完成模型服务化部署的示例,您可以参考[模型服务化部署](../inference_deployment/paddle_serving_deploy.md)来完成相应的部署工作。
<a name="6.5"></a>
### 6.5 端侧部署
Paddle Lite 是一个高性能、轻量级、灵活性强且易于扩展的深度学习推理框架,定位于支持包括移动端、嵌入式以及服务器端在内的多硬件平台。更多关于 Paddle Lite 的介绍,可以参考[Paddle Lite 代码仓库](https://github.com/PaddlePaddle/Paddle-Lite)
PaddleClas 提供了基于 Paddle Lite 来完成模型端侧部署的示例,您可以参考[端侧部署](../inference_deployment/paddle_lite_deploy.md)来完成相应的部署工作。
<a name="6.6"></a>
### 6.6 Paddle2ONNX 模型转换与预测
Paddle2ONNX 支持将 PaddlePaddle 模型格式转化到 ONNX 模型格式。通过 ONNX 可以完成将 Paddle 模型到多种推理引擎的部署,包括TensorRT/OpenVINO/MNN/TNN/NCNN,以及其它对 ONNX 开源格式进行支持的推理引擎或硬件。更多关于 Paddle2ONNX 的介绍,可以参考[Paddle2ONNX 代码仓库](https://github.com/PaddlePaddle/Paddle2ONNX)
PaddleClas 提供了基于 Paddle2ONNX 来完成 inference 模型转换 ONNX 模型并作推理预测的示例,您可以参考[Paddle2ONNX 模型转换与预测](@shuilong)来完成相应的部署工作。
# PULC 交通标志分类模型
------
## 目录
- [1. 模型和应用场景介绍](#1)
- [2. 模型快速体验](#2)
- [3. 模型训练、评估和预测](#3)
- [3.1 环境配置](#3.1)
- [3.2 数据准备](#3.2)
- [3.2.1 数据集来源](#3.2.1)
- [3.2.2 数据集获取](#3.2.2)
- [3.3 模型训练](#3.3)
- [3.4 模型评估](#3.4)
- [3.5 模型预测](#3.5)
- [4. 模型压缩](#4)
- [4.1 SKL-UGI 知识蒸馏](#4.1)
- [4.1.1 教师模型训练](#4.1.1)
- [4.1.2 蒸馏训练](#4.1.2)
- [5. 超参搜索](#5)
- [6. 模型推理部署](#6)
- [6.1 推理模型准备](#6.1)
- [6.1.1 基于训练得到的权重导出 inference 模型](#6.1.1)
- [6.1.2 直接下载 inference 模型](#6.1.2)
- [6.2 基于 Python 预测引擎推理](#6.2)
- [6.2.1 预测单张图像](#6.2.1)
- [6.2.2 基于文件夹的批量预测](#6.2.2)
- [6.3 基于 C++ 预测引擎推理](#6.3)
- [6.4 服务化部署](#6.4)
- [6.5 端侧部署](#6.5)
- [6.6 Paddle2ONNX 模型转换与预测](#6.6)
<a name="1"></a>
## 1. 模型和应用场景介绍
该案例提供了用户使用 PaddleClas 的超轻量图像分类方案(PULC,Practical Ultra Lightweight Classification)快速构建轻量级、高精度、可落地的交通标志分类模型。该模型可以广泛应用于自动驾驶、道路监控等场景。
下表列出了不同交通标志分类模型的相关指标,前两行展现了使用 SwinTranformer_tiny 和 MobileNetV3_large_x1_0 作为 backbone 训练得到的模型的相关指标,第三行至第六行依次展现了替换 backbone 为 PPLCNet_x1_0、使用 SSLD 预训练模型、使用 SSLD 预训练模型 + EDA 策略、使用 SSLD 预训练模型 + EDA 策略 + SKL-UGI 知识蒸馏策略训练得到的模型的相关指标。
| 模型 | Top-1 Acc(%) | 延时(ms) | 存储(M) | 策略 |
|-------|-----------|----------|---------------|---------------|
| SwinTranformer_tiny | 98.11 | 87.19 | 111 | 使用ImageNet预训练模型 |
| MobileNetV3_large_x1_0 | 97.79 | 5.59 | 23 | 使用ImageNet预训练模型 |
| PPLCNet_x1_0 | 97.78 | 2.67 | 8.2 | 使用ImageNet预训练模型 |
| PPLCNet_x1_0 | 97.84 | 2.67 | 8.2 | 使用SSLD预训练模型 |
| PPLCNet_x1_0 | 98.14 | 2.67 | 8.2 | 使用SSLD预训练模型+EDA策略|
| <b>PPLCNet_x1_0<b> | <b>98.35<b> | <b>2.67<b> | <b>8.2<b> | 使用SSLD预训练模型+EDA策略+SKL-UGI知识蒸馏策略|
从表中可以看出,backbone 为 SwinTranformer_tiny 时精度较高,但是推理速度较慢。将 backbone 替换为轻量级模型 MobileNetV3_large_x1_0 后,速度可以大幅提升,但是精度下降明显。将 backbone 替换为 PPLCNet_x1_0 时,精度低0.01%,但是速度提升 2 倍左右。在此基础上,使用 SSLD 预训练模型后,在不改变推理速度的前提下,精度可以提升约 0.06%,进一步地,当融合EDA策略后,精度可以再提升 0.3%,最后,在使用 SKL-UGI 知识蒸馏后,精度可以继续提升 0.21%。此时,PPLCNet_x1_0 的精度超越了SwinTranformer_tiny,速度快32倍。关于 PULC 的训练方法和推理部署方法将在下面详细介绍。
**备注:**
* 关于PPLCNet的介绍可以参考[PPLCNet介绍](../models/PP-LCNet.md),相关论文可以查阅[PPLCNet paper](https://arxiv.org/abs/2109.15099)
<a name="2"></a>
## 2. 模型快速体验
(pip方式,待补充)
<a name="3"></a>
## 3. 模型训练、评估和预测
<a name="3.1"></a>
### 3.1 环境配置
* 安装:请先参考 [Paddle 安装教程](../installation/install_paddle.md) 以及 [PaddleClas 安装教程](../installation/install_paddleclas.md) 配置 PaddleClas 运行环境。
<a name="3.2"></a>
### 3.2 数据准备
<a name="3.2.1"></a>
#### 3.2.1 数据集来源
本案例中所使用的数据为[Tsinghua-Tencent 100K dataset (CC-BY-NC license)](https://cg.cs.tsinghua.edu.cn/traffic-sign/),在使用的过程中,对交通标志检测框进行随机扩充与裁剪,从而得到用于训练与测试的图像,下面简称该数据集为`TT100K`数据集。
<a name="3.2.2"></a>
#### 3.2.2 数据集获取
在TT00K数据集上,对交通标志检测框进行随机扩充与裁剪,从而得到用于训练与测试的图像。随机扩充检测框的逻辑如下所示。
```python
def get_random_crop_box(xmin, ymin, xmax, ymax, img_height, img_width, ratio=1.0):
h = ymax - ymin
w = ymax - ymin
xmin_diff = random.random() * ratio * min(w, xmin/ratio)
ymin_diff = random.random() * ratio * min(h, ymin/ratio)
xmax_diff = random.random() * ratio * min(w, (img_width-xmin-1)/ratio)
ymax_diff = random.random() * ratio * min(h, (img_height-ymin-1)/ratio)
new_xmin = round(xmin - xmin_diff)
new_ymin = round(ymin - ymin_diff)
new_xmax = round(xmax + xmax_diff)
new_ymax = round(ymax + ymax_diff)
return new_xmin, new_ymin, new_xmax, new_ymax
```
完整的预处理逻辑,可以参考下载好的数据集文件夹中的`deal.py`文件。
处理后的数据集部分数据可视化如下。
<div align="center">
<img src="../../images/PULC/docs/traffic_sign_data_demo.png" width = "500" />
</div>
此处提供了经过上述方法处理好的数据,可以直接下载得到。
进入 PaddleClas 目录。
```
cd path_to_PaddleClas
```
进入 `dataset/` 目录,下载并解压交通标志分类场景的数据。
```shell
cd dataset
wget https://paddleclas.bj.bcebos.com/data/PULC/traffic_sign.tar
tar -xf traffic_sign.tar
cd ../
```
执行上述命令后,`dataset/`下存在`traffic_sign`目录,该目录中具有以下数据:
```
traffic_sign
├── train
│ ├── 0_62627.jpg
│ ├── 100000_89031.jpg
│ ├── 100001_89031.jpg
...
├── test
│ ├── 100423_2315.jpg
│ ├── 100424_2315.jpg
│ ├── 100425_2315.jpg
...
├── other
│ ├── 100603_3422.jpg
│ ├── 100604_3422.jpg
...
├── label_list_train.txt
├── label_list_test.txt
├── label_list_other.txt
├── label_list_train_for_distillation.txt
├── label_list_train.txt.debug
├── label_list_test.txt.debug
├── label_name_id.txt
├── deal.py
```
其中`train/``test/`分别为训练集和验证集。`label_list_train.txt``label_list_test.txt`分别为训练集和验证集的标签文件,`label_list_train.txt.debug``label_list_test.txt.debug`分别为训练集和验证集的`debug`标签文件,其分别是`label_list_train.txt``label_list_test.txt`的子集,用该文件可以快速体验本案例的流程。`train``other`的混合数据用于本案例的`SKL-UGI知识蒸馏策略`,对应的训练标签文件为`label_list_train_for_distillation.txt`
**备注:**
* 关于 `label_list_train.txt``label_list_test.txt`的格式说明,可以参考[PaddleClas分类数据集格式说明](../data_preparation/classification_dataset.md#1-数据集格式说明)
* 关于如何得到蒸馏的标签文件可以参考[知识蒸馏标签获得方法](@ruoyu)
<a name="3.3"></a>
### 3.3 模型训练
`ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0.yaml` 中提供了基于该场景的训练配置,可以通过如下脚本启动训练:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0.yaml
```
验证集的最佳指标在 `98.14%` 左右(数据集较小,一般有0.1%左右的波动)。
<a name="3.4"></a>
### 3.4 模型评估
训练好模型之后,可以通过以下命令实现对模型指标的评估。
```bash
python3 tools/eval.py \
-c ./ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0.yaml \
-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"
```
其中 `-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` 指定了当前最佳权重所在的路径,如果指定其他权重,只需替换对应的路径即可。
<a name="3.5"></a>
### 3.5 模型预测
模型训练完成之后,可以加载训练得到的预训练模型,进行模型预测。在模型库的 `tools/infer.py` 中提供了完整的示例,只需执行下述命令即可完成模型预测:
```bash
python3 tools/infer.py \
-c ./ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/DistillationModel/best_model
```
输出结果如下:
```
99603_17806.jpg: class id(s): [216, 145, 49, 207, 169], score(s): [1.00, 0.00, 0.00, 0.00, 0.00], label_name(s): ['pm20', 'pm30', 'pm40', 'pl25', 'pm15']
```
**备注:**
* 这里`-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` 指定了当前最佳权重所在的路径,如果指定其他权重,只需替换对应的路径即可。
* 默认是对 `deploy/images/PULC/traffic_sign/99603_17806.jpg` 进行预测,此处也可以通过增加字段 `-o Infer.infer_imgs=xxx` 对其他图片预测。
<a name="4"></a>
## 4. 模型压缩
<a name="4.1"></a>
### 4.1 SKL-UGI 知识蒸馏
SKL-UGI 知识蒸馏是 PaddleClas 提出的一种简单有效的知识蒸馏方法,关于该方法的介绍,可以参考[SKL-UGI 知识蒸馏](@ruoyu)
<a name="4.1.1"></a>
#### 4.1.1 教师模型训练
复用 `ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0.yaml` 中的超参数,训练教师模型,训练脚本如下:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0.yaml \
-o Arch.name=ResNet101_vd
```
验证集的最佳指标为 `98.59%` 左右,当前教师模型最好的权重保存在 `output/ResNet101_vd/best_model.pdparams`
<a name="4.1.2"></a>
#### 4.1.2 蒸馏训练
配置文件`ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0_distillation.yaml`提供了`SKL-UGI知识蒸馏策略`的配置。该配置将`ResNet101_vd`当作教师模型,`PPLCNet_x1_0`当作学生模型,使用ImageNet数据集的验证集作为新增的无标签数据。训练脚本如下:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0_distillation.yaml \
-o Arch.models.0.Teacher.pretrained=output/ResNet101_vd/best_model
```
验证集的最佳指标为 `98.35%` 左右,当前模型最好的权重保存在 `output/DistillationModel/best_model_student.pdparams`
<a name="5"></a>
## 5. 超参搜索
[3.2 节](#3.2)[4.1 节](#4.1)所使用的超参数是根据 PaddleClas 提供的 `SHAS 超参数搜索策略` 搜索得到的,如果希望在自己的数据集上得到更好的结果,可以参考[SHAS 超参数搜索策略](#TODO)来获得更好的训练超参数。
**备注:** 此部分内容是可选内容,搜索过程需要较长的时间,您可以根据自己的硬件情况来选择执行。如果没有更换数据集,可以忽略此节内容。
<a name="6"></a>
## 6. 模型推理部署
<a name="6.1"></a>
### 6.1 推理模型准备
Paddle Inference 是飞桨的原生推理库, 作用于服务器端和云端,提供高性能的推理能力。相比于直接基于预训练模型进行预测,Paddle Inference可使用MKLDNN、CUDNN、TensorRT 进行预测加速,从而实现更优的推理性能。更多关于Paddle Inference推理引擎的介绍,可以参考[Paddle Inference官网教程](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/infer/inference/inference_cn.html)
当使用 Paddle Inference 推理时,加载的模型类型为 inference 模型。本案例提供了两种获得 inference 模型的方法,如果希望得到和文档相同的结果,请选择[直接下载 inference 模型](#6.1.2)的方式。
<a name="6.1.1"></a>
### 6.1.1 基于训练得到的权重导出 inference 模型
此处,我们提供了将权重和模型转换的脚本,执行该脚本可以得到对应的 inference 模型:
```bash
python3 tools/export_model.py \
-c ./ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/DistillationModel/best_model_student \
-o Global.save_inference_dir=deploy/models/PPLCNet_x1_0_traffic_sign_infer
```
执行完该脚本后会在 `deploy/models/` 下生成 `PPLCNet_x1_0_traffic_sign_infer` 文件夹,`models` 文件夹下应有如下文件结构:
```
├── PPLCNet_x1_0_traffic_sign_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
**备注:** 此处的最佳权重是经过知识蒸馏后的权重路径,如果没有执行知识蒸馏的步骤,最佳模型保存在`output/PPLCNet_x1_0/best_model.pdparams`中。
<a name="6.1.2"></a>
### 6.1.2 直接下载 inference 模型
[6.1.1 小节](#6.1.1)提供了导出 inference 模型的方法,此处也提供了该场景可以下载的 inference 模型,可以直接下载体验。
```
cd deploy/models
# 下载 inference 模型并解压
wget https://paddleclas.bj.bcebos.com/models/PULC/traffic_sign_infer.tar && tar -xf traffic_sign_infer.tar
```
解压完毕后,`models` 文件夹下应有如下文件结构:
```
├── traffic_sign_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
<a name="6.2"></a>
### 6.2 基于 Python 预测引擎推理
<a name="6.2.1"></a>
#### 6.2.1 预测单张图像
返回 `deploy` 目录:
```
cd ../
```
运行下面的命令,对图像 `./images/PULC/traffic_sign/99603_17806.jpg` 进行交通标志分类。
```shell
# 使用下面的命令使用 GPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/traffic_sign/inference_traffic_sign.yaml
# 使用下面的命令使用 CPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/traffic_sign/inference_traffic_sign.yaml -o Global.use_gpu=False
```
输出结果如下。
```
99603_17806.jpg: class id(s): [216, 145, 49, 207, 169], score(s): [1.00, 0.00, 0.00, 0.00, 0.00], label_name(s): ['pm20', 'pm30', 'pm40', 'pl25', 'pm15']
```
<a name="6.2.2"></a>
#### 6.2.2 基于文件夹的批量预测
如果希望预测文件夹内的图像,可以直接修改配置文件中的 `Global.infer_imgs` 字段,也可以通过下面的 `-o` 参数修改对应的配置。
```shell
# 使用下面的命令使用 GPU 进行预测,如果希望使用 CPU 预测,可以在命令后面添加 -o Global.use_gpu=False
python3.7 python/predict_cls.py -c configs/PULC/traffic_sign/inference_traffic_sign.yaml -o Global.infer_imgs="./images/PULC/traffic_sign/"
```
终端中会输出该文件夹内所有图像的分类结果,如下所示。
```
100999_83928.jpg: class id(s): [182, 179, 162, 128, 24], score(s): [0.99, 0.01, 0.00, 0.00, 0.00], label_name(s): ['pl110', 'pl100', 'pl120', 'p26', 'pm10']
99603_17806.jpg: class id(s): [216, 145, 49, 24, 169], score(s): [1.00, 0.00, 0.00, 0.00, 0.00], label_name(s): ['pm20', 'pm30', 'pm40', 'pm10', 'pm15']
```
输出的 `label_name`可以从`dataset/traffic_sign/report.pdf`文件中查阅对应的图片。
<a name="6.3"></a>
### 6.3 基于 C++ 预测引擎推理
PaddleClas 提供了基于 C++ 预测引擎推理的示例,您可以参考[服务器端 C++ 预测](../inference_deployment/cpp_deploy.md)来完成相应的推理部署。如果您使用的是 Windows 平台,可以参考[基于 Visual Studio 2019 Community CMake 编译指南](../inference_deployment/cpp_deploy_on_windows.md)完成相应的预测库编译和模型预测工作。
<a name="6.4"></a>
### 6.4 服务化部署
Paddle Serving 提供高性能、灵活易用的工业级在线推理服务。Paddle Serving 支持 RESTful、gRPC、bRPC 等多种协议,提供多种异构硬件和多种操作系统环境下推理解决方案。更多关于Paddle Serving 的介绍,可以参考[Paddle Serving 代码仓库](https://github.com/PaddlePaddle/Serving)
PaddleClas 提供了基于 Paddle Serving 来完成模型服务化部署的示例,您可以参考[模型服务化部署](../inference_deployment/paddle_serving_deploy.md)来完成相应的部署工作。
<a name="6.5"></a>
### 6.5 端侧部署
Paddle Lite 是一个高性能、轻量级、灵活性强且易于扩展的深度学习推理框架,定位于支持包括移动端、嵌入式以及服务器端在内的多硬件平台。更多关于 Paddle Lite 的介绍,可以参考[Paddle Lite 代码仓库](https://github.com/PaddlePaddle/Paddle-Lite)
PaddleClas 提供了基于 Paddle Lite 来完成模型端侧部署的示例,您可以参考[端侧部署](../inference_deployment/paddle_lite_deploy.md)来完成相应的部署工作。
<a name="6.6"></a>
### 6.6 Paddle2ONNX 模型转换与预测
Paddle2ONNX 支持将 PaddlePaddle 模型格式转化到 ONNX 模型格式。通过 ONNX 可以完成将 Paddle 模型到多种推理引擎的部署,包括TensorRT/OpenVINO/MNN/TNN/NCNN,以及其它对 ONNX 开源格式进行支持的推理引擎或硬件。更多关于 Paddle2ONNX 的介绍,可以参考[Paddle2ONNX 代码仓库](https://github.com/PaddlePaddle/Paddle2ONNX)
PaddleClas 提供了基于 Paddle2ONNX 来完成 inference 模型转换 ONNX 模型并作推理预测的示例,您可以参考[Paddle2ONNX 模型转换与预测](@shuilong)来完成相应的部署工作。
# PULC 车辆属性识别模型
------
## 目录
- [1. 模型和应用场景介绍](#1)
- [2. 模型快速体验](#2)
- [3. 模型训练、评估和预测](#3)
- [3.1 环境配置](#3.1)
- [3.2 数据准备](#3.2)
- [3.2.1 数据集来源](#3.2.1)
- [3.2.2 数据集获取](#3.2.2)
- [3.3 模型训练](#3.3)
- [3.4 模型评估](#3.4)
- [3.5 模型预测](#3.5)
- [4. 模型压缩](#4)
- [4.1 SKL-UGI 知识蒸馏](#4.1)
- [4.1.1 教师模型训练](#4.1.1)
- [4.1.2 蒸馏训练](#4.1.2)
- [5. 超参搜索](#5)
- [6. 模型推理部署](#6)
- [6.1 推理模型准备](#6.1)
- [6.1.1 基于训练得到的权重导出 inference 模型](#6.1.1)
- [6.1.2 直接下载 inference 模型](#6.1.2)
- [6.2 基于 Python 预测引擎推理](#6.2)
- [6.2.1 预测单张图像](#6.2.1)
- [6.2.2 基于文件夹的批量预测](#6.2.2)
- [6.3 基于 C++ 预测引擎推理](#6.3)
- [6.4 服务化部署](#6.4)
- [6.5 端侧部署](#6.5)
- [6.6 Paddle2ONNX 模型转换与预测](#6.6)
<a name="1"></a>
## 1. 模型和应用场景介绍
该案例提供了用户使用 PaddleClas 的超轻量图像分类方案(PULC,Practical Ultra Lightweight Classification)快速构建轻量级、高精度、可落地的车辆属性识别模型。该模型可以广泛应用于车辆识别、道路监控等场景。
下表列出了不同车辆属性识别模型的相关指标,前两行展现了使用 Res2Net200_vd_26w_4s 和 MobileNetV3_large_x1_0 作为 backbone 训练得到的模型的相关指标,第三行至第六行依次展现了替换 backbone 为 PPLCNet_x1_0、使用 SSLD 预训练模型、使用 SSLD 预训练模型 + EDA 策略、使用 SSLD 预训练模型 + EDA 策略 + SKL-UGI 知识蒸馏策略训练得到的模型的相关指标。
| 模型 | ma(%) | 延时(ms) | 存储(M) | 策略 |
|-------|-----------|----------|---------------|---------------|
| Res2Net200_vd_26w_4s | 91.36 | 66.58 | 293 | 使用ImageNet预训练模型 |
| ResNet50 | 89.98 | 12.74 | 92 | 使用ImageNet预训练模型 |
| MobileNetV3_large_x1_0 | 89.77 | 5.59 | 23 | 使用ImageNet预训练模型 |
| PPLCNet_x1_0 | 89.57 | 2.56 | 8.2 | 使用ImageNet预训练模型 |
| PPLCNet_x1_0 | 90.07 | 2.56 | 8.2 | 使用SSLD预训练模型 |
| PPLCNet_x1_0 | 90.59 | 2.56 | 8.2 | 使用SSLD预训练模型+EDA策略|
| <b>PPLCNet_x1_0<b> | <b>90.81<b> | <b>2.56<b> | <b>8.2<b> | 使用SSLD预训练模型+EDA策略+SKL-UGI知识蒸馏策略|
从表中可以看出,backbone 为 Res2Net200_vd_26w_4s 时精度较高,但是推理速度较慢。将 backbone 替换为轻量级模型 MobileNetV3_large_x1_0 后,速度可以大幅提升,但是精度下降明显。将 backbone 替换为 PPLCNet_x1_0 时,精度低0.2%,但是速度提升 2 倍左右。在此基础上,使用 SSLD 预训练模型后,在不改变推理速度的前提下,精度可以提升约 0.5%,进一步地,当融合EDA策略后,精度可以再提升 0.52%,最后,在使用 SKL-UGI 知识蒸馏后,精度可以继续提升 0.23%。此时,PPLCNet_x1_0 的精度与 Res2Net200_vd_26w_4s 仅相差0.55%,但是速度快26倍。关于 PULC 的训练方法和推理部署方法将在下面详细介绍。
**备注:**
* 关于PPLCNet的介绍可以参考[PPLCNet介绍](../models/PP-LCNet.md),相关论文可以查阅[PPLCNet paper](https://arxiv.org/abs/2109.15099)
<a name="2"></a>
## 2. 模型快速体验
```
(pip方式,待补充)
```
<a name="3"></a>
## 3. 模型训练、评估和预测
<a name="3.1"></a>
### 3.1 环境配置
* 安装:请先参考 [Paddle 安装教程](../installation/install_paddle.md) 以及 [PaddleClas 安装教程](../installation/install_paddleclas.md) 配置 PaddleClas 运行环境。
<a name="3.2"></a>
### 3.2 数据准备
<a name="3.2.1"></a>
#### 3.2.1 数据集来源
本案例中所使用的数据为[VeRi 数据集](https://www.v7labs.com/open-datasets/veri-dataset)
<a name="3.2.2"></a>
#### 3.2.2 数据集获取
部分数据可视化如下所示。
<div align="center">
<img src="../../images/PULC/docs/vehicle_attr_data_demo.png" width = "500" />
</div>
首先从[VeRi数据集官网](https://www.v7labs.com/open-datasets/veri-dataset)中申请并下载数据,放在PaddleClas的`dataset`目录下,数据集目录名为`VeRi`,使用下面的命令进入该文件夹。
```shell
cd PaddleClas/dataset/VeRi/
```
然后使用下面的代码转换label(可以在python终端中执行下面的命令,也可以将其写入一个文件,然后使用`python3 convert.py`的方式运行该文件)。
```python
import os
from xml.dom.minidom import parse
vehicleids = []
def convert_annotation(input_fp, output_fp):
in_file = open(input_fp)
list_file = open(output_fp, 'w')
tree = parse(in_file)
root = tree.documentElement
for item in root.getElementsByTagName("Item"):
label = ['0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0']
if item.hasAttribute("imageName"):
name = item.getAttribute("imageName")
if item.hasAttribute("vehicleID"):
vehicleid = item.getAttribute("vehicleID")
if vehicleid not in vehicleids :
vehicleids.append(vehicleid)
vid = vehicleids.index(vehicleid)
if item.hasAttribute("colorID"):
colorid = int (item.getAttribute("colorID"))
label[colorid-1] = '1'
if item.hasAttribute("typeID"):
typeid = int (item.getAttribute("typeID"))
label[typeid+9] = '1'
label = ','.join(label)
list_file.write(os.path.join('image_train', name) + "\t" + label + "\n")
list_file.close()
convert_annotation('train_label.xml', 'train_list.txt') #imagename vehiclenum colorid typeid
convert_annotation('test_label.xml', 'test_list.txt')
```
执行上述命令后,`VeRi`目录中具有以下数据:
```
VeRi
├── image_train
│ ├── 0001_c001_00016450_0.jpg
│ ├── 0001_c001_00016460_0.jpg
│ ├── 0001_c001_00016470_0.jpg
...
├── image_test
│ ├── 0002_c002_00030600_0.jpg
│ ├── 0002_c002_00030605_1.jpg
│ ├── 0002_c002_00030615_1.jpg
...
...
├── train_list.txt
├── test_list.txt
├── train_label.xml
├── test_label.xml
```
其中`train/``test/`分别为训练集和验证集。`train_list.txt``test_list.txt`分别为训练集和验证集的转换后用于训练的标签文件。
<a name="3.3"></a>
### 3.3 模型训练
`ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0.yaml` 中提供了基于该场景的训练配置,可以通过如下脚本启动训练:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0.yaml
```
验证集的最佳指标在 `90.07%` 左右(数据集较小,一般有0.3%左右的波动)。
<a name="3.4"></a>
### 3.4 模型评估
训练好模型之后,可以通过以下命令实现对模型指标的评估。
```bash
python3 tools/eval.py \
-c ./ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0.yaml \
-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"
```
其中 `-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` 指定了当前最佳权重所在的路径,如果指定其他权重,只需替换对应的路径即可。
<a name="3.5"></a>
### 3.5 模型预测
模型训练完成之后,可以加载训练得到的预训练模型,进行模型预测。在模型库的 `tools/infer.py` 中提供了完整的示例,只需执行下述命令即可完成模型预测:
```bash
python3 tools/infer.py \
-c ./ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/DistillationModel/best_model
```
输出结果如下:
```
[{'attr': 'Color: (yellow, prob: 0.9893478155136108), Type: (hatchback, prob: 0.9734100103378296)', 'pred': [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0], 'file_name': './deploy/images/PULC/vehicle_attr/0002_c002_00030670_0.jpg'}]
```
**备注:**
* 这里`-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` 指定了当前最佳权重所在的路径,如果指定其他权重,只需替换对应的路径即可。
* 默认是对 `./deploy/images/PULC/vehicle_attr/0002_c002_00030670_0.jpg` 进行预测,此处也可以通过增加字段 `-o Infer.infer_imgs=xxx` 对其他图片预测。
<a name="4"></a>
## 4. 模型压缩
<a name="4.1"></a>
### 4.1 SKL-UGI 知识蒸馏
SKL-UGI 知识蒸馏是 PaddleClas 提出的一种简单有效的知识蒸馏方法,关于该方法的介绍,可以参考[SKL-UGI 知识蒸馏](@ruoyu)
<a name="4.1.1"></a>
#### 4.1.1 教师模型训练
复用 `ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0.yaml` 中的超参数,训练教师模型,训练脚本如下:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0.yaml \
-o Arch.name=ResNet101_vd
```
验证集的最佳指标为 `91.60%` 左右,当前教师模型最好的权重保存在 `output/ResNet101_vd/best_model.pdparams`
<a name="4.1.2"></a>
#### 4.1.2 蒸馏训练
配置文件`ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0_distillation.yaml`提供了`SKL-UGI知识蒸馏策略`的配置。该配置将`ResNet101_vd`当作教师模型,`PPLCNet_x1_0`当作学生模型。训练脚本如下:
```shell
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0_distillation.yaml \
-o Arch.models.0.Teacher.pretrained=output/ResNet101_vd/best_model
```
验证集的最佳指标为 `90.81%` 左右,当前模型最好的权重保存在 `output/DistillationModel/best_model_student.pdparams`
<a name="5"></a>
## 5. 超参搜索
[3.2 节](#3.2)[4.1 节](#4.1)所使用的超参数是根据 PaddleClas 提供的 `SHAS 超参数搜索策略` 搜索得到的,如果希望在自己的数据集上得到更好的结果,可以参考[SHAS 超参数搜索策略](#TODO)来获得更好的训练超参数。
**备注:** 此部分内容是可选内容,搜索过程需要较长的时间,您可以根据自己的硬件情况来选择执行。如果没有更换数据集,可以忽略此节内容。
<a name="6"></a>
## 6. 模型推理部署
<a name="6.1"></a>
### 6.1 推理模型准备
Paddle Inference 是飞桨的原生推理库, 作用于服务器端和云端,提供高性能的推理能力。相比于直接基于预训练模型进行预测,Paddle Inference可使用MKLDNN、CUDNN、TensorRT 进行预测加速,从而实现更优的推理性能。更多关于Paddle Inference推理引擎的介绍,可以参考[Paddle Inference官网教程](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/infer/inference/inference_cn.html)
当使用 Paddle Inference 推理时,加载的模型类型为 inference 模型。本案例提供了两种获得 inference 模型的方法,如果希望得到和文档相同的结果,请选择[直接下载 inference 模型](#6.1.2)的方式。
<a name="6.1.1"></a>
### 6.1.1 基于训练得到的权重导出 inference 模型
此处,我们提供了将权重和模型转换的脚本,执行该脚本可以得到对应的 inference 模型:
```bash
python3 tools/export_model.py \
-c ./ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/DistillationModel/best_model_student \
-o Global.save_inference_dir=deploy/models/PPLCNet_x1_0_vehicle_attr_infer
```
执行完该脚本后会在 `deploy/models/` 下生成 `PPLCNet_x1_0_vehicle_attr_infer` 文件夹,`models` 文件夹下应有如下文件结构:
```
├── PPLCNet_x1_0_vehicle_attr_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
**备注:** 此处的最佳权重是经过知识蒸馏后的权重路径,如果没有执行知识蒸馏的步骤,最佳模型保存在`output/PPLCNet_x1_0/best_model.pdparams`中。
<a name="6.1.2"></a>
### 6.1.2 直接下载 inference 模型
[6.1.1 小节](#6.1.1)提供了导出 inference 模型的方法,此处也提供了该场景可以下载的 inference 模型,可以直接下载体验。
```
cd deploy/models
# 下载 inference 模型并解压
wget https://paddleclas.bj.bcebos.com/models/PULC/vehicle_attr_infer.tar && tar -xf vehicle_attr_infer.tar
```
解压完毕后,`models` 文件夹下应有如下文件结构:
```
├── vehicle_attr_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
<a name="6.2"></a>
### 6.2 基于 Python 预测引擎推理
<a name="6.2.1"></a>
#### 6.2.1 预测单张图像
返回 `deploy` 目录:
```
cd ../
```
运行下面的命令,对图像 `./images/PULC/vehicle_attr/0002_c002_00030670_0.jpg` 进行车辆属性识别。
```shell
# 使用下面的命令使用 GPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/vehicle_attr/inference_vehicle_attr.yaml -o Global.use_gpu=True
# 使用下面的命令使用 CPU 进行预测
python3.7 python/predict_cls.py -c configs/PULC/vehicle_attr/inference_vehicle_attr.yaml -o Global.use_gpu=False
```
输出结果如下。
```
0002_c002_00030670_0.jpg: attributes: Color: (yellow, prob: 0.9893478155136108), Type: (hatchback, prob: 0.97340989112854),
predict output: [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0]
```
<a name="6.2.2"></a>
#### 6.2.2 基于文件夹的批量预测
如果希望预测文件夹内的图像,可以直接修改配置文件中的 `Global.infer_imgs` 字段,也可以通过下面的 `-o` 参数修改对应的配置。
```shell
# 使用下面的命令使用 GPU 进行预测,如果希望使用 CPU 预测,可以在命令后面添加 -o Global.use_gpu=False
python3.7 python/predict_cls.py -c configs/PULC/vehicle_attr/inference_vehicle_attr.yaml -o Global.infer_imgs="./images/PULC/vehicle_attr/"
```
终端中会输出该文件夹内所有图像的属性识别结果,如下所示。
```
0002_c002_00030670_0.jpg: attributes: Color: (yellow, prob: 0.9893478155136108), Type: (hatchback, prob: 0.97340989112854),
predict output: [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0]
0014_c012_00040750_0.jpg: attributes: Color: (red, prob: 0.9998721480369568), Type: (sedan, prob: 0.999976634979248),
predict output: [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0]
```
<a name="6.3"></a>
### 6.3 基于 C++ 预测引擎推理
PaddleClas 提供了基于 C++ 预测引擎推理的示例,您可以参考[服务器端 C++ 预测](../inference_deployment/cpp_deploy.md)来完成相应的推理部署。如果您使用的是 Windows 平台,可以参考[基于 Visual Studio 2019 Community CMake 编译指南](../inference_deployment/cpp_deploy_on_windows.md)完成相应的预测库编译和模型预测工作。
<a name="6.4"></a>
### 6.4 服务化部署
Paddle Serving 提供高性能、灵活易用的工业级在线推理服务。Paddle Serving 支持 RESTful、gRPC、bRPC 等多种协议,提供多种异构硬件和多种操作系统环境下推理解决方案。更多关于Paddle Serving 的介绍,可以参考[Paddle Serving 代码仓库](https://github.com/PaddlePaddle/Serving)
PaddleClas 提供了基于 Paddle Serving 来完成模型服务化部署的示例,您可以参考[模型服务化部署](../inference_deployment/paddle_serving_deploy.md)来完成相应的部署工作。
<a name="6.5"></a>
### 6.5 端侧部署
Paddle Lite 是一个高性能、轻量级、灵活性强且易于扩展的深度学习推理框架,定位于支持包括移动端、嵌入式以及服务器端在内的多硬件平台。更多关于 Paddle Lite 的介绍,可以参考[Paddle Lite 代码仓库](https://github.com/PaddlePaddle/Paddle-Lite)
PaddleClas 提供了基于 Paddle Lite 来完成模型端侧部署的示例,您可以参考[端侧部署](../inference_deployment/paddle_lite_deploy.md)来完成相应的部署工作。
<a name="6.6"></a>
### 6.6 Paddle2ONNX 模型转换与预测
Paddle2ONNX 支持将 PaddlePaddle 模型格式转化到 ONNX 模型格式。通过 ONNX 可以完成将 Paddle 模型到多种推理引擎的部署,包括TensorRT/OpenVINO/MNN/TNN/NCNN,以及其它对 ONNX 开源格式进行支持的推理引擎或硬件。更多关于 Paddle2ONNX 的介绍,可以参考[Paddle2ONNX 代码仓库](https://github.com/PaddlePaddle/Paddle2ONNX)
PaddleClas 提供了基于 Paddle2ONNX 来完成 inference 模型转换 ONNX 模型并作推理预测的示例,您可以参考[Paddle2ONNX 模型转换与预测](@shuilong)来完成相应的部署工作。
# 分布式训练
## 1. 简介
* 分布式训练指的是将训练任务按照一定方法拆分到多个计算节点进行计算,再按照一定的方法对拆分后计算得到的梯度等信息进行聚合与更新。飞桨分布式训练技术源自百度的业务实践,在自然语言处理、计算机视觉、搜索和推荐等领域经过超大规模业务检验。分布式训练的高性能,是飞桨的核心优势技术之一,在图像分类等任务上,分布式训练可以达到几乎线性的加速比。图像分类训练任务中往往包含大量训练数据,以ImageNet为例,ImageNet22k数据集中包含1400W张图像,如果使用单卡训练,会非常耗时。因此PaddleClas中使用分布式训练接口完成训练任务,同时支持单机训练与多机训练。更多关于分布式训练的方法与文档可以参考:[分布式训练快速开始教程](https://fleet-x.readthedocs.io/en/latest/paddle_fleet_rst/parameter_server/ps_quick_start.html)
## 2. 使用方法
### 2.1 单机训练
* 以识别为例,本地准备好数据之后,使用`paddle.distributed.launch`的接口启动训练任务即可。下面为运行代码示例。
```shell
python3 -m paddle.distributed.launch \
--log_dir=./log/ \
--gpus "0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/ImageNet/ResNet/ResNet50.yaml
```
### 2.2 多机训练
* 相比单机训练,多机训练时,只需要添加`--ips`的参数,该参数表示需要参与分布式训练的机器的ip列表,不同机器的ip用逗号隔开。下面为运行代码示例。
```shell
ip_list="192.168.0.1,192.168.0.2"
python3 -m paddle.distributed.launch \
--log_dir=./log/ \
--ips="${ip_list}" \
--gpus="0,1,2,3" \
tools/train.py \
-c ./ppcls/configs/ImageNet/ResNet/ResNet50.yaml
```
**注:**
* 不同机器的ip信息需要用逗号隔开,可以通过`ifconfig`或者`ipconfig`查看。
* 不同机器之间需要做免密设置,且可以直接ping通,否则无法完成通信。
* 不同机器之间的代码、数据与运行命令或脚本需要保持一致,且所有的机器上都需要运行设置好的训练命令或者脚本。最终`ip_list`中的第一台机器的第一块设备是trainer0,以此类推。
* 不同机器的起始端口可能不同,建议在启动多机任务前,在不同的机器中设置相同的多机运行起始端口,命令为`export FLAGS_START_PORT=17000`,端口值建议在`10000~20000`之间。
## 3. 性能效果测试
* 在4机8卡V100的机器上,基于[SSLD知识蒸馏训练策略](../advanced_tutorials/knowledge_distillation.md)(数据量500W)进行模型训练,不同模型的训练耗时以及多机加速比情况如下所示。
| 模型 | 精度 | 单机8卡耗时 | 4机8卡耗时 | 加速比 |
|:---------:|:--------:|:--------:|:--------:|:------:|
| PPHGNet-base_ssld | 85.00% | 15.74d | 4.86d | **3.23** |
| PPLCNetv2-base_ssld | 80.10% | 6.4d | 1.67d | **3.83** |
| PPLCNet_x0_25_ssld | 53.43% | 6.2d | 1.78d | **3.48** |
...@@ -154,7 +154,8 @@ class MobileNetV3(TheseusLayer): ...@@ -154,7 +154,8 @@ class MobileNetV3(TheseusLayer):
class_expand=LAST_CONV, class_expand=LAST_CONV,
dropout_prob=0.2, dropout_prob=0.2,
return_patterns=None, return_patterns=None,
return_stages=None): return_stages=None,
**kwargs):
super().__init__() super().__init__()
self.cfg = config self.cfg = config
......
...@@ -94,13 +94,16 @@ class ConvBNLayer(TheseusLayer): ...@@ -94,13 +94,16 @@ class ConvBNLayer(TheseusLayer):
stride=stride, stride=stride,
padding=(filter_size - 1) // 2, padding=(filter_size - 1) // 2,
groups=num_groups, groups=num_groups,
weight_attr=ParamAttr(initializer=KaimingNormal(), learning_rate=lr_mult), weight_attr=ParamAttr(
initializer=KaimingNormal(), learning_rate=lr_mult),
bias_attr=False) bias_attr=False)
self.bn = BatchNorm2D( self.bn = BatchNorm2D(
num_filters, num_filters,
weight_attr=ParamAttr(regularizer=L2Decay(0.0), learning_rate=lr_mult), weight_attr=ParamAttr(
bias_attr=ParamAttr(regularizer=L2Decay(0.0), learning_rate=lr_mult)) regularizer=L2Decay(0.0), learning_rate=lr_mult),
bias_attr=ParamAttr(
regularizer=L2Decay(0.0), learning_rate=lr_mult))
self.hardswish = nn.Hardswish() self.hardswish = nn.Hardswish()
def forward(self, x): def forward(self, x):
...@@ -128,8 +131,7 @@ class DepthwiseSeparable(TheseusLayer): ...@@ -128,8 +131,7 @@ class DepthwiseSeparable(TheseusLayer):
num_groups=num_channels, num_groups=num_channels,
lr_mult=lr_mult) lr_mult=lr_mult)
if use_se: if use_se:
self.se = SEModule(num_channels, self.se = SEModule(num_channels, lr_mult=lr_mult)
lr_mult=lr_mult)
self.pw_conv = ConvBNLayer( self.pw_conv = ConvBNLayer(
num_channels=num_channels, num_channels=num_channels,
filter_size=1, filter_size=1,
...@@ -189,7 +191,8 @@ class PPLCNet(TheseusLayer): ...@@ -189,7 +191,8 @@ class PPLCNet(TheseusLayer):
lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0], lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
use_last_conv=True, use_last_conv=True,
return_patterns=None, return_patterns=None,
return_stages=None): return_stages=None,
**kwargs):
super().__init__() super().__init__()
self.scale = scale self.scale = scale
self.class_expand = class_expand self.class_expand = class_expand
...@@ -271,7 +274,8 @@ class PPLCNet(TheseusLayer): ...@@ -271,7 +274,8 @@ class PPLCNet(TheseusLayer):
self.avg_pool = AdaptiveAvgPool2D(1) self.avg_pool = AdaptiveAvgPool2D(1)
if self.use_last_conv: if self.use_last_conv:
self.last_conv = Conv2D( self.last_conv = Conv2D(
in_channels=make_divisible(NET_CONFIG["blocks6"][-1][2] * scale), in_channels=make_divisible(NET_CONFIG["blocks6"][-1][2] *
scale),
out_channels=self.class_expand, out_channels=self.class_expand,
kernel_size=1, kernel_size=1,
stride=1, stride=1,
...@@ -282,7 +286,8 @@ class PPLCNet(TheseusLayer): ...@@ -282,7 +286,8 @@ class PPLCNet(TheseusLayer):
else: else:
self.last_conv = None self.last_conv = None
self.flatten = nn.Flatten(start_axis=1, stop_axis=-1) self.flatten = nn.Flatten(start_axis=1, stop_axis=-1)
self.fc = Linear(self.class_expand if self.use_last_conv else NET_CONFIG["blocks6"][-1][2], class_num) self.fc = Linear(self.class_expand if self.use_last_conv else
NET_CONFIG["blocks6"][-1][2], class_num)
super().init_res( super().init_res(
stages_pattern, stages_pattern,
......
...@@ -165,7 +165,8 @@ class BottleneckBlock(nn.Layer): ...@@ -165,7 +165,8 @@ class BottleneckBlock(nn.Layer):
class Res2Net_vd(nn.Layer): class Res2Net_vd(nn.Layer):
def __init__(self, layers=50, scales=4, width=26, class_num=1000): def __init__(self, layers=50, scales=4, width=26, class_num=1000,
**kwargs):
super(Res2Net_vd, self).__init__() super(Res2Net_vd, self).__init__()
self.layers = layers self.layers = layers
......
...@@ -113,7 +113,7 @@ DataLoader: ...@@ -113,7 +113,7 @@ DataLoader:
use_shared_memory: True use_shared_memory: True
Infer: Infer:
infer_imgs: docs/images/inference_deployment/whl_demo.jpg infer_imgs: deploy/images/PULC/person_exists/objects365_02035329.jpg
batch_size: 10 batch_size: 10
transforms: transforms:
- DecodeImage: - DecodeImage:
......
...@@ -119,7 +119,7 @@ DataLoader: ...@@ -119,7 +119,7 @@ DataLoader:
use_shared_memory: True use_shared_memory: True
Infer: Infer:
infer_imgs: docs/images/inference_deployment/whl_demo.jpg infer_imgs: deploy/images/PULC/person_exists/objects365_02035329.jpg
batch_size: 10 batch_size: 10
transforms: transforms:
- DecodeImage: - DecodeImage:
......
...@@ -135,7 +135,7 @@ DataLoader: ...@@ -135,7 +135,7 @@ DataLoader:
use_shared_memory: True use_shared_memory: True
Infer: Infer:
infer_imgs: docs/images/inference_deployment/whl_demo.jpg infer_imgs: deploy/images/PULC/person_exists/objects365_02035329.jpg
batch_size: 10 batch_size: 10
transforms: transforms:
- DecodeImage: - DecodeImage:
......
...@@ -119,7 +119,7 @@ DataLoader: ...@@ -119,7 +119,7 @@ DataLoader:
use_shared_memory: True use_shared_memory: True
Infer: Infer:
infer_imgs: docs/images/inference_deployment/whl_demo.jpg infer_imgs: deploy/images/PULC/person_exists/objects365_02035329.jpg
batch_size: 10 batch_size: 10
transforms: transforms:
- DecodeImage: - DecodeImage:
......
...@@ -136,7 +136,7 @@ DataLoader: ...@@ -136,7 +136,7 @@ DataLoader:
use_shared_memory: True use_shared_memory: True
Infer: Infer:
infer_imgs: docs/images/inference_deployment/whl_demo.jpg infer_imgs: deploy/images/PULC/person_exists/objects365_02035329.jpg
batch_size: 10 batch_size: 10
transforms: transforms:
- DecodeImage: - DecodeImage:
......
base_config_file: ppcls/configs/PULC/person_exists/PPLCNet_x1_0_search.yaml
distill_config_file: ppcls/configs/PULC/person_exists/PPLCNet_x1_0_distillation.yaml
gpus: 0,1,2,3
output_dir: output/search_person_cls
search_times: 1
search_dict:
- search_key: lrs
replace_config:
- Optimizer.lr.learning_rate
search_values: [0.0075, 0.01, 0.0125]
- search_key: resolutions
replace_config:
- DataLoader.Train.dataset.transform_ops.1.RandCropImage.size
- DataLoader.Train.dataset.transform_ops.3.TimmAutoAugment.img_size
search_values: [176, 192, 224]
- search_key: ra_probs
replace_config:
- DataLoader.Train.dataset.transform_ops.3.TimmAutoAugment.prob
search_values: [0.0, 0.1, 0.5]
- search_key: re_probs
replace_config:
- DataLoader.Train.dataset.transform_ops.5.RandomErasing.EPSILON
search_values: [0.0, 0.1, 0.5]
- search_key: lr_mult_list
replace_config:
- Arch.lr_mult_list
search_values:
- [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]
- [0.0, 0.4, 0.4, 0.8, 0.8, 1.0]
- [1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
teacher:
rm_keys:
- Arch.lr_mult_list
search_values:
- ResNet101_vd
- ResNet50_vd
final_replace:
Arch.lr_mult_list: Arch.models.1.Student.lr_mult_list
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
eval_interval: 1
epochs: 10
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# model architecture
Arch:
name: MobileNetV3_large_x1_0
class_num: 232
pretrained: True
# loss function config for traing/eval process
Loss:
Train:
- CELoss:
weight: 1.0
epsilon: 0.1
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00002
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/traffic_sign/label_list_train.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/traffic_sign/label_list_test.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Infer:
infer_imgs: docs/images/inference_deployment/whl_demo.jpg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
PostProcess:
name: Topk
topk: 5
class_id_map_file: dataset/traffic_sign/label_name_id.txt
Metric:
Train:
- TopkAcc:
topk: [1, 5]
Eval:
- TopkAcc:
topk: [1, 5]
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
eval_interval: 1
start_eval_epoch: 0
epochs: 10
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# training model under @to_static
to_static: False
use_dali: False
# model architecture
Arch:
name: PPLCNet_x1_0
class_num: 232
pretrained: True
use_ssld: True
# loss function config for traing/eval process
Loss:
Train:
- CELoss:
weight: 1.0
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.02
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00004
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/tt100k_clas_v2/label_list_train.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 224
- TimmAutoAugment:
prob: 0.5
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.0
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/traffic_sign/label_list_test.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Infer:
infer_imgs: deploy/images/PULC/traffic_sign/99603_17806.jpg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
PostProcess:
name: Topk
topk: 5
class_id_map_file: dataset/traffic_sign/label_name_id.txt
Metric:
Train:
- TopkAcc:
topk: [1, 5]
Eval:
- TopkAcc:
topk: [1, 5]
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
eval_interval: 1
epochs: 10
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# training model under @to_static
to_static: False
use_dali: False
# mixed precision training
AMP:
scale_loss: 128.0
use_dynamic_loss_scaling: True
# O1: mixed fp16
level: O1
# model architecture
Arch:
name: "DistillationModel"
class_num: &class_num 232
# if not null, its lengths should be same as models
pretrained_list:
# if not null, its lengths should be same as models
freeze_params_list:
- True
- False
models:
- Teacher:
name: ResNet101_vd
class_num: *class_num
pretrained: False
- Student:
name: PPLCNet_x1_0
class_num: *class_num
pretrained: True
use_ssld: True
infer_model_name: "Student"
# loss function config for traing/eval process
Loss:
Train:
- DistillationDMLLoss:
weight: 1.0
model_name_pairs:
- ["Student", "Teacher"]
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00004
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/traffic_sign/label_list_train_for_distillation.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 224
- TimmAutoAugment:
prob: 0.0
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.0
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/traffic_sign/label_list_test.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Infer:
infer_imgs: docs/images/inference_deployment/whl_demo.jpg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
PostProcess:
name: Topk
topk: 5
class_id_map_file: dataset/traffic_sign/label_name_id.txt
Metric:
Train:
- DistillationTopkAcc:
model_key: "Student"
topk: [1, 5]
Eval:
- TopkAcc:
topk: [1, 5]
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
eval_interval: 1
start_eval_epoch: 0
epochs: 10
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# training model under @to_static
to_static: False
use_dali: False
# model architecture
Arch:
name: PPLCNet_x1_0
class_num: 232
pretrained: True
# use_ssld: True
# loss function config for traing/eval process
Loss:
Train:
- CELoss:
weight: 1.0
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00004
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/traffic_sign/label_list_train.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 224
- TimmAutoAugment:
prob: 0.0
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.0
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/traffic_sign/label_list_test.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Infer:
# infer_imgs: dataset/traffic_sign_demo/
infer_imgs: dataset/tt100k_clas_v2/test/
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
PostProcess:
name: Topk
topk: 5
class_id_map_file: dataset/traffic_sign/label_name_id.txt
Metric:
Train:
- TopkAcc:
topk: [1, 5]
Eval:
- TopkAcc:
topk: [1, 5]
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
device: gpu
save_interval: 1
eval_during_train: True
eval_interval: 1
start_eval_epoch: 0
epochs: 10
print_batch_step: 10
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
save_inference_dir: ./inference
# training model under @to_static
to_static: False
use_dali: False
# mixed precision training
AMP:
scale_loss: 128.0
use_dynamic_loss_scaling: True
# O1: mixed fp16
level: O1
# model architecture
Arch:
name: SwinTransformer_tiny_patch4_window7_224
class_num: 232
pretrained: True
# loss function config for traing/eval process
Loss:
Train:
- CELoss:
weight: 1.0
epsilon: 0.1
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: AdamW
beta1: 0.9
beta2: 0.999
epsilon: 1e-8
weight_decay: 0.05
no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm
one_dim_param_no_weight_decay: True
lr:
name: Cosine
learning_rate: 2e-4
eta_min: 2e-6
warmup_epoch: 5
warmup_start_lr: 2e-7
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/traffic_sign/label_list_train.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 224
interpolation: bicubic
backend: pil
- RandFlipImage:
flip_code: 1
- TimmAutoAugment:
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.25
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
batch_transform_ops:
- OpSampler:
MixupOperator:
alpha: 0.8
prob: 0.5
CutmixOperator:
alpha: 1.0
prob: 0.5
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: ./dataset/
cls_label_path: ./dataset/tt100k_clas_v2/label_list_test.txt
delimiter: "\t"
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Infer:
infer_imgs: docs/images/inference_deployment/whl_demo.jpg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
PostProcess:
name: Topk
topk: 5
class_id_map_file: dataset/traffic_sign/label_name_id.txt
Metric:
Train:
- TopkAcc:
topk: [1, 5]
Eval:
- TopkAcc:
topk: [1, 5]
base_config_file: ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0_search.yaml
distill_config_file: ppcls/configs/PULC/traffic_sign/PPLCNet_x1_0_distillation.yaml
gpus: 0,1,2,3
output_dir: output/search_traffic_sign
search_times: 1
search_dict:
- search_key: lrs
replace_config:
- Optimizer.lr.learning_rate
search_values: [0.0075, 0.01, 0.0125]
- search_key: resolutions
replace_config:
- DataLoader.Train.dataset.transform_ops.1.RandCropImage.size
- DataLoader.Train.dataset.transform_ops.2.TimmAutoAugment.img_size
search_values: [176, 192, 224]
- search_key: ra_probs
replace_config:
- DataLoader.Train.dataset.transform_ops.2.TimmAutoAugment.prob
search_values: [0.0, 0.1, 0.5]
- search_key: re_probs
replace_config:
- DataLoader.Train.dataset.transform_ops.4.RandomErasing.EPSILON
search_values: [0.0, 0.1, 0.5]
- search_key: lr_mult_list
replace_config:
- Arch.lr_mult_list
search_values:
- [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]
- [0.0, 0.4, 0.4, 0.8, 0.8, 1.0]
- [1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
teacher:
algorithm: "skl-ugi"
rm_keys:
- Arch.lr_mult_list
search_values:
- ResNet101_vd
- ResNet50_vd
final_replace:
Arch.lr_mult_list: Arch.models.1.Student.lr_mult_list
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: "./output/"
device: "gpu"
save_interval: 5
eval_during_train: True
eval_interval: 1
epochs: 30
print_batch_step: 20
use_visualdl: False
# used for static mode and model export
image_shape: [3, 192, 256]
save_inference_dir: "./inference"
use_multilabel: True
# model architecture
Arch:
name: "MobileNetV3_large_x1_0"
pretrained: True
class_num: 19
infer_add_softmax: False
# loss function config for traing/eval process
Loss:
Train:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Eval:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.0005
# data loader for train and eval
DataLoader:
Train:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/train_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- Padv2:
size: [276, 212]
pad_mode: 1
fill_value: 0
- RandomCropImage:
size: [256, 192]
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: True
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/test_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Metric:
Eval:
- ATTRMetric:
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: "./output/"
device: "gpu"
save_interval: 5
eval_during_train: True
eval_interval: 1
epochs: 30
print_batch_step: 20
use_visualdl: False
# used for static mode and model export
image_shape: [3, 192, 256]
save_inference_dir: "./inference"
use_multilabel: True
# model architecture
Arch:
name: "PPLCNet_x1_0"
pretrained: True
class_num: 19
use_ssld: True
lr_mult_list: [1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
infer_add_softmax: False
# loss function config for traing/eval process
Loss:
Train:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Eval:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.0125
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.0005
# data loader for train and eval
DataLoader:
Train:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/train_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- TimmAutoAugment:
prob: 0.0
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: [256, 192]
- Padv2:
size: [276, 212]
pad_mode: 1
fill_value: 0
- RandomCropImage:
size: [256, 192]
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.5
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: True
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/test_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Infer:
infer_imgs: ./deploy/images/PULC/vehicle_attr/0002_c002_00030670_0.jpg
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
PostProcess:
name: VehicleAttribute
color_threshold: 0.5
type_threshold: 0.5
Metric:
Eval:
- ATTRMetric:
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: "./output/"
device: "gpu"
save_interval: 5
eval_during_train: True
eval_interval: 1
epochs: 30
print_batch_step: 20
use_visualdl: False
# used for static mode and model export
image_shape: [3, 192, 256]
save_inference_dir: "./inference"
use_multilabel: True
# model architecture
Arch:
name: "DistillationModel"
class_num: &class_num 19
# if not null, its lengths should be same as models
pretrained_list:
# if not null, its lengths should be same as models
freeze_params_list:
- True
- False
use_ssld: True
models:
- Teacher:
name: ResNet101_vd
class_num: *class_num
- Student:
name: PPLCNet_x1_0
class_num: *class_num
pretrained: True
use_ssld: True
# loss function config for traing/eval process
Loss:
Train:
- DistillationMultiLabelLoss:
weight: 1.0
model_names: ["Student"]
weight_ratio: True
size_sum: True
- DistillationDMLLoss:
weight: 1.0
weight_ratio: True
sum_across_class_dim: False
model_name_pairs:
- ["Student", "Teacher"]
Eval:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.0005
# data loader for train and eval
DataLoader:
Train:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/train_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- TimmAutoAugment:
prob: 0.0
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: [256, 192]
- Padv2:
size: [276, 212]
pad_mode: 1
fill_value: 0
- RandomCropImage:
size: [256, 192]
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.0
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: True
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/test_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Metric:
Eval:
- ATTRMetric:
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: "./output/"
device: "gpu"
save_interval: 5
eval_during_train: True
eval_interval: 1
epochs: 30
print_batch_step: 20
use_visualdl: False
# used for static mode and model export
image_shape: [3, 192, 256]
save_inference_dir: "./inference"
use_multilabel: True
# model architecture
Arch:
name: "PPLCNet_x1_0"
pretrained: True
use_ssld: True
class_num: 19
infer_add_softmax: False
# loss function config for traing/eval process
Loss:
Train:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Eval:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.0005
# data loader for train and eval
DataLoader:
Train:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/train_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- TimmAutoAugment:
prob: 0.0
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: [256, 192]
- Padv2:
size: [276, 212]
pad_mode: 1
fill_value: 0
- RandomCropImage:
size: [256, 192]
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 0.0
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: True
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/test_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 128
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Metric:
Eval:
- ATTRMetric:
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: "./output/mo"
device: "gpu"
save_interval: 5
eval_during_train: True
eval_interval: 1
epochs: 30
print_batch_step: 20
use_visualdl: False
# used for static mode and model export
image_shape: [3, 192, 256]
save_inference_dir: "./inference"
use_multilabel: True
# mixed precision training
AMP:
scale_loss: 128.0
use_dynamic_loss_scaling: True
# O1: mixed fp16
level: O1
# model architecture
Arch:
name: "Res2Net200_vd_26w_4s"
pretrained: True
class_num: 19
infer_add_softmax: False
# loss function config for traing/eval process
Loss:
Train:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Eval:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.0005
# data loader for train and eval
DataLoader:
Train:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/train_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- Padv2:
size: [276, 212]
pad_mode: 1
fill_value: 0
- RandomCropImage:
size: [256, 192]
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: True
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/test_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Metric:
Eval:
- ATTRMetric:
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: "./output/"
device: "gpu"
save_interval: 5
eval_during_train: True
eval_interval: 1
epochs: 30
print_batch_step: 20
use_visualdl: False
# used for static mode and model export
image_shape: [3, 192, 256]
save_inference_dir: "./inference"
use_multilabel: True
# model architecture
Arch:
name: "ResNet50"
pretrained: True
class_num: 19
infer_add_softmax: False
# loss function config for traing/eval process
Loss:
Train:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Eval:
- MultiLabelLoss:
weight: 1.0
weight_ratio: True
size_sum: True
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.01
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.0005
# data loader for train and eval
DataLoader:
Train:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/train_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- Padv2:
size: [276, 212]
pad_mode: 1
fill_value: 0
- RandomCropImage:
size: [256, 192]
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: True
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
Eval:
dataset:
name: MultiLabelDataset
image_root: "dataset/VeRi/"
cls_label_path: "dataset/VeRi/test_list.txt"
label_ratio: True
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
size: [256, 192]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 64
drop_last: False
shuffle: False
loader:
num_workers: 8
use_shared_memory: True
Metric:
Eval:
- ATTRMetric:
base_config_file: ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0_search.yaml
distill_config_file: ppcls/configs/PULC/vehicle_attr/PPLCNet_x1_0_distillation.yaml
gpus: 0,1,2,3
output_dir: output/search_vehicle_attr
search_times: 1
search_dict:
- search_key: lrs
replace_config:
- Optimizer.lr.learning_rate
search_values: [0.0075, 0.01, 0.0125]
- search_key: ra_probs
replace_config:
- DataLoader.Train.dataset.transform_ops.2.TimmAutoAugment.prob
search_values: [0.0, 0.1, 0.5]
- search_key: re_probs
replace_config:
- DataLoader.Train.dataset.transform_ops.7.RandomErasing.EPSILON
search_values: [0.0, 0.1, 0.5]
- search_key: lr_mult_list
replace_config:
- Arch.lr_mult_list
search_values:
- [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]
- [0.0, 0.4, 0.4, 0.8, 0.8, 1.0]
- [1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
teacher:
algorithm: "skl-ugi"
rm_keys:
- Arch.lr_mult_list
search_values:
- ResNet101_vd
- ResNet50_vd
final_replace:
Arch.lr_mult_list: Arch.models.1.Student.lr_mult_list
...@@ -18,6 +18,7 @@ from . import topk, threshoutput ...@@ -18,6 +18,7 @@ from . import topk, threshoutput
from .topk import Topk, MultiLabelTopk from .topk import Topk, MultiLabelTopk
from .threshoutput import ThreshOutput from .threshoutput import ThreshOutput
from .attr_rec import VehicleAttribute
def build_postprocess(config): def build_postprocess(config):
......
# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import numpy as np
import paddle
import paddle.nn.functional as F
class VehicleAttribute(object):
def __init__(self, color_threshold=0.5, type_threshold=0.5):
self.color_threshold = color_threshold
self.type_threshold = type_threshold
self.color_list = [
"yellow", "orange", "green", "gray", "red", "blue", "white",
"golden", "brown", "black"
]
self.type_list = [
"sedan", "suv", "van", "hatchback", "mpv", "pickup", "bus",
"truck", "estate"
]
def __call__(self, x, file_names=None):
if isinstance(x, dict):
x = x['logits']
assert isinstance(x, paddle.Tensor)
if file_names is not None:
assert x.shape[0] == len(file_names)
x = F.sigmoid(x).numpy()
# postprocess output of predictor
batch_res = []
for idx, res in enumerate(x):
res = res.tolist()
label_res = []
color_idx = np.argmax(res[:10])
type_idx = np.argmax(res[10:])
print(color_idx, type_idx)
if res[color_idx] >= self.color_threshold:
color_info = f"Color: ({self.color_list[color_idx]}, prob: {res[color_idx]})"
else:
color_info = "Color unknown"
if res[type_idx + 10] >= self.type_threshold:
type_info = f"Type: ({self.type_list[type_idx]}, prob: {res[type_idx + 10]})"
else:
type_info = "Type unknown"
label_res = f"{color_info}, {type_info}"
threshold_list = [self.color_threshold
] * 10 + [self.type_threshold] * 9
pred_res = (np.array(res) > np.array(threshold_list)
).astype(np.int8).tolist()
batch_res.append({
"attr": label_res,
"pred": pred_res,
"file_name": file_names[idx]
})
return batch_res
...@@ -21,9 +21,9 @@ import paddle.nn.functional as F ...@@ -21,9 +21,9 @@ import paddle.nn.functional as F
class Topk(object): class Topk(object):
def __init__(self, topk=1, class_id_map_file=None, delimiter=None): def __init__(self, topk=1, class_id_map_file=None, delimiter=None):
assert isinstance(topk, (int, )) assert isinstance(topk, (int, ))
self.class_id_map = self.parse_class_id_map(class_id_map_file)
self.topk = topk self.topk = topk
self.delimiter = delimiter if delimiter is not None else " " self.delimiter = delimiter if delimiter is not None else " "
self.class_id_map = self.parse_class_id_map(class_id_map_file)
def parse_class_id_map(self, class_id_map_file): def parse_class_id_map(self, class_id_map_file):
if class_id_map_file is None: if class_id_map_file is None:
......
...@@ -365,9 +365,6 @@ class RandomCropImage(object): ...@@ -365,9 +365,6 @@ class RandomCropImage(object):
j = random.randint(0, w - tw) j = random.randint(0, w - tw)
img = img[i:i + th, j:j + tw, :] img = img[i:i + th, j:j + tw, :]
if img.shape[0] != 256 or img.shape[1] != 192:
raise ValueError('sample: ', h, w, i, j, th, tw, img.shape)
return img return img
......
...@@ -152,39 +152,33 @@ class Engine(object): ...@@ -152,39 +152,33 @@ class Engine(object):
self.eval_loss_func = None self.eval_loss_func = None
# build metric # build metric
if self.mode == 'train': if self.mode == 'train' and "Metric" in self.config and "Train" in self.config[
metric_config = self.config.get("Metric") "Metric"] and self.config["Metric"]["Train"]:
if metric_config is not None: metric_config = self.config["Metric"]["Train"]
metric_config = metric_config.get("Train") if hasattr(self.train_dataloader, "collate_fn"
if metric_config is not None: ) and self.train_dataloader.collate_fn is not None:
if hasattr( for m_idx, m in enumerate(metric_config):
self.train_dataloader, "collate_fn" if "TopkAcc" in m:
) and self.train_dataloader.collate_fn is not None: msg = f"Unable to calculate accuracy when using \"batch_transform_ops\". The metric \"{m}\" has been removed."
for m_idx, m in enumerate(metric_config): logger.warning(msg)
if "TopkAcc" in m:
msg = f"'TopkAcc' metric can not be used when setting 'batch_transform_ops' in config. The 'TopkAcc' metric has been removed."
logger.warning(msg)
break
metric_config.pop(m_idx) metric_config.pop(m_idx)
self.train_metric_func = build_metrics(metric_config) self.train_metric_func = build_metrics(metric_config)
else:
self.train_metric_func = None
else: else:
self.train_metric_func = None self.train_metric_func = None
if self.mode == "eval" or (self.mode == "train" and if self.mode == "eval" or (self.mode == "train" and
self.config["Global"]["eval_during_train"]): self.config["Global"]["eval_during_train"]):
metric_config = self.config.get("Metric")
if self.eval_mode == "classification": if self.eval_mode == "classification":
if metric_config is not None: if "Metric" in self.config and "Eval" in self.config["Metric"]:
metric_config = metric_config.get("Eval") self.eval_metric_func = build_metrics(self.config["Metric"]
if metric_config is not None: ["Eval"])
self.eval_metric_func = build_metrics(metric_config) else:
self.eval_metric_func = None
elif self.eval_mode == "retrieval": elif self.eval_mode == "retrieval":
if metric_config is None: if "Metric" in self.config and "Eval" in self.config["Metric"]:
metric_config = [{"name": "Recallk", "topk": (1, 5)}] metric_config = self.config["Metric"]["Eval"]
else: else:
metric_config = metric_config["Eval"] metric_config = [{"name": "Recallk", "topk": (1, 5)}]
self.eval_metric_func = build_metrics(metric_config) self.eval_metric_func = build_metrics(metric_config)
else: else:
self.eval_metric_func = None self.eval_metric_func = None
...@@ -222,7 +216,7 @@ class Engine(object): ...@@ -222,7 +216,7 @@ class Engine(object):
AMP_RELATED_FLAGS_SETTING.update({ AMP_RELATED_FLAGS_SETTING.update({
'FLAGS_cudnn_batchnorm_spatial_persistent': 1 'FLAGS_cudnn_batchnorm_spatial_persistent': 1
}) })
paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING) paddle.set_flags(AMP_RELATED_FLAGS_SETTING)
self.scale_loss = self.config["AMP"].get("scale_loss", 1.0) self.scale_loss = self.config["AMP"].get("scale_loss", 1.0)
self.use_dynamic_loss_scaling = self.config["AMP"].get( self.use_dynamic_loss_scaling = self.config["AMP"].get(
...@@ -460,7 +454,7 @@ class Engine(object): ...@@ -460,7 +454,7 @@ class Engine(object):
assert self.mode == "export" assert self.mode == "export"
use_multilabel = self.config["Global"].get( use_multilabel = self.config["Global"].get(
"use_multilabel", "use_multilabel",
False) and not "ATTRMetric" in self.config["Metric"]["Eval"][0] False) and "ATTRMetric" in self.config["Metric"]["Eval"][0]
model = ExportModel(self.config["Arch"], self.model, use_multilabel) model = ExportModel(self.config["Arch"], self.model, use_multilabel)
if self.config["Global"]["pretrained_model"] is not None: if self.config["Global"]["pretrained_model"] is not None:
load_dygraph_pretrain(model.base_model, load_dygraph_pretrain(model.base_model,
......
...@@ -34,7 +34,6 @@ def classification_eval(engine, epoch_id=0): ...@@ -34,7 +34,6 @@ def classification_eval(engine, epoch_id=0):
} }
print_batch_step = engine.config["Global"]["print_batch_step"] print_batch_step = engine.config["Global"]["print_batch_step"]
metric_key = None
tic = time.time() tic = time.time()
accum_samples = 0 accum_samples = 0
total_samples = len( total_samples = len(
......
...@@ -24,6 +24,7 @@ from .distillationloss import DistillationDistanceLoss ...@@ -24,6 +24,7 @@ from .distillationloss import DistillationDistanceLoss
from .distillationloss import DistillationRKDLoss from .distillationloss import DistillationRKDLoss
from .distillationloss import DistillationKLDivLoss from .distillationloss import DistillationKLDivLoss
from .distillationloss import DistillationDKDLoss from .distillationloss import DistillationDKDLoss
from .distillationloss import DistillationMultiLabelLoss
from .multilabelloss import MultiLabelLoss from .multilabelloss import MultiLabelLoss
from .afdloss import AFDLoss from .afdloss import AFDLoss
......
...@@ -22,6 +22,7 @@ from .distanceloss import DistanceLoss ...@@ -22,6 +22,7 @@ from .distanceloss import DistanceLoss
from .rkdloss import RKdAngle, RkdDistance from .rkdloss import RKdAngle, RkdDistance
from .kldivloss import KLDivLoss from .kldivloss import KLDivLoss
from .dkdloss import DKDLoss from .dkdloss import DKDLoss
from .multilabelloss import MultiLabelLoss
class DistillationCELoss(CELoss): class DistillationCELoss(CELoss):
...@@ -89,13 +90,16 @@ class DistillationDMLLoss(DMLLoss): ...@@ -89,13 +90,16 @@ class DistillationDMLLoss(DMLLoss):
def __init__(self, def __init__(self,
model_name_pairs=[], model_name_pairs=[],
act="softmax", act="softmax",
weight_ratio=False,
sum_across_class_dim=False,
key=None, key=None,
name="loss_dml"): name="loss_dml"):
super().__init__(act=act) super().__init__(act=act, sum_across_class_dim=sum_across_class_dim)
assert isinstance(model_name_pairs, list) assert isinstance(model_name_pairs, list)
self.key = key self.key = key
self.model_name_pairs = model_name_pairs self.model_name_pairs = model_name_pairs
self.name = name self.name = name
self.weight_ratio = weight_ratio
def forward(self, predicts, batch): def forward(self, predicts, batch):
loss_dict = dict() loss_dict = dict()
...@@ -105,7 +109,10 @@ class DistillationDMLLoss(DMLLoss): ...@@ -105,7 +109,10 @@ class DistillationDMLLoss(DMLLoss):
if self.key is not None: if self.key is not None:
out1 = out1[self.key] out1 = out1[self.key]
out2 = out2[self.key] out2 = out2[self.key]
loss = super().forward(out1, out2) if self.weight_ratio is True:
loss = super().forward(out1, out2, batch)
else:
loss = super().forward(out1, out2)
if isinstance(loss, dict): if isinstance(loss, dict):
for key in loss: for key in loss:
loss_dict["{}_{}_{}_{}".format(key, pair[0], pair[1], loss_dict["{}_{}_{}_{}".format(key, pair[0], pair[1],
...@@ -122,6 +129,7 @@ class DistillationDistanceLoss(DistanceLoss): ...@@ -122,6 +129,7 @@ class DistillationDistanceLoss(DistanceLoss):
def __init__(self, def __init__(self,
mode="l2", mode="l2",
model_name_pairs=[], model_name_pairs=[],
act=None,
key=None, key=None,
name="loss_", name="loss_",
**kargs): **kargs):
...@@ -130,6 +138,13 @@ class DistillationDistanceLoss(DistanceLoss): ...@@ -130,6 +138,13 @@ class DistillationDistanceLoss(DistanceLoss):
self.key = key self.key = key
self.model_name_pairs = model_name_pairs self.model_name_pairs = model_name_pairs
self.name = name + mode self.name = name + mode
assert act in [None, "sigmoid", "softmax"]
if act == "sigmoid":
self.act = nn.Sigmoid()
elif act == "softmax":
self.act = nn.Softmax(axis=-1)
else:
self.act = None
def forward(self, predicts, batch): def forward(self, predicts, batch):
loss_dict = dict() loss_dict = dict()
...@@ -139,6 +154,9 @@ class DistillationDistanceLoss(DistanceLoss): ...@@ -139,6 +154,9 @@ class DistillationDistanceLoss(DistanceLoss):
if self.key is not None: if self.key is not None:
out1 = out1[self.key] out1 = out1[self.key]
out2 = out2[self.key] out2 = out2[self.key]
if self.act is not None:
out1 = self.act(out1)
out2 = self.act(out2)
loss = super().forward(out1, out2) loss = super().forward(out1, out2)
for key in loss: for key in loss:
loss_dict["{}_{}_{}".format(self.name, key, idx)] = loss[key] loss_dict["{}_{}_{}".format(self.name, key, idx)] = loss[key]
...@@ -235,3 +253,34 @@ class DistillationDKDLoss(DKDLoss): ...@@ -235,3 +253,34 @@ class DistillationDKDLoss(DKDLoss):
loss = super().forward(out1, out2, batch) loss = super().forward(out1, out2, batch)
loss_dict[f"{self.name}_{pair[0]}_{pair[1]}"] = loss loss_dict[f"{self.name}_{pair[0]}_{pair[1]}"] = loss
return loss_dict return loss_dict
class DistillationMultiLabelLoss(MultiLabelLoss):
"""
DistillationMultiLabelLoss
"""
def __init__(self,
model_names=[],
epsilon=None,
size_sum=False,
weight_ratio=False,
key=None,
name="loss_mll"):
super().__init__(
epsilon=epsilon, size_sum=size_sum, weight_ratio=weight_ratio)
assert isinstance(model_names, list)
self.key = key
self.model_names = model_names
self.name = name
def forward(self, predicts, batch):
loss_dict = dict()
for name in self.model_names:
out = predicts[name]
if self.key is not None:
out = out[self.key]
loss = super().forward(out, batch)
for key in loss:
loss_dict["{}_{}".format(key, name)] = loss[key]
return loss_dict
...@@ -16,13 +16,15 @@ import paddle ...@@ -16,13 +16,15 @@ import paddle
import paddle.nn as nn import paddle.nn as nn
import paddle.nn.functional as F import paddle.nn.functional as F
from ppcls.loss.multilabelloss import ratio2weight
class DMLLoss(nn.Layer): class DMLLoss(nn.Layer):
""" """
DMLLoss DMLLoss
""" """
def __init__(self, act="softmax", eps=1e-12): def __init__(self, act="softmax", sum_across_class_dim=False, eps=1e-12):
super().__init__() super().__init__()
if act is not None: if act is not None:
assert act in ["softmax", "sigmoid"] assert act in ["softmax", "sigmoid"]
...@@ -33,6 +35,7 @@ class DMLLoss(nn.Layer): ...@@ -33,6 +35,7 @@ class DMLLoss(nn.Layer):
else: else:
self.act = None self.act = None
self.eps = eps self.eps = eps
self.sum_across_class_dim = sum_across_class_dim
def _kldiv(self, x, target): def _kldiv(self, x, target):
class_num = x.shape[-1] class_num = x.shape[-1]
...@@ -40,11 +43,20 @@ class DMLLoss(nn.Layer): ...@@ -40,11 +43,20 @@ class DMLLoss(nn.Layer):
(target + self.eps) / (x + self.eps)) * class_num (target + self.eps) / (x + self.eps)) * class_num
return cost return cost
def forward(self, x, target): def forward(self, x, target, gt_label=None):
if self.act is not None: if self.act is not None:
x = self.act(x) x = self.act(x)
target = self.act(target) target = self.act(target)
loss = self._kldiv(x, target) + self._kldiv(target, x) loss = self._kldiv(x, target) + self._kldiv(target, x)
loss = loss / 2 loss = loss / 2
loss = paddle.mean(loss)
# for multi-label dml loss
if gt_label is not None:
gt_label, label_ratio = gt_label[:, 0, :], gt_label[:, 1, :]
targets_mask = paddle.cast(gt_label > 0.5, 'float32')
weight = ratio2weight(targets_mask, paddle.to_tensor(label_ratio))
weight = weight * (gt_label > -1)
loss = loss * weight
loss = loss.sum(1).mean() if self.sum_across_class_dim else loss.mean()
return {"DMLLoss": loss} return {"DMLLoss": loss}
...@@ -26,6 +26,7 @@ from easydict import EasyDict ...@@ -26,6 +26,7 @@ from easydict import EasyDict
from ppcls.metric.avg_metrics import AvgMetrics from ppcls.metric.avg_metrics import AvgMetrics
from ppcls.utils.misc import AverageMeter, AttrMeter from ppcls.utils.misc import AverageMeter, AttrMeter
from ppcls.utils import logger
class TopkAcc(AvgMetrics): class TopkAcc(AvgMetrics):
...@@ -39,7 +40,7 @@ class TopkAcc(AvgMetrics): ...@@ -39,7 +40,7 @@ class TopkAcc(AvgMetrics):
def reset(self): def reset(self):
self.avg_meters = { self.avg_meters = {
"top{}".format(k): AverageMeter("top{}".format(k)) f"top{k}": AverageMeter(f"top{k}")
for k in self.topk for k in self.topk
} }
...@@ -47,11 +48,21 @@ class TopkAcc(AvgMetrics): ...@@ -47,11 +48,21 @@ class TopkAcc(AvgMetrics):
if isinstance(x, dict): if isinstance(x, dict):
x = x["logits"] x = x["logits"]
output_dims = x.shape[-1]
metric_dict = dict() metric_dict = dict()
for k in self.topk: for idx, k in enumerate(self.topk):
metric_dict["top{}".format(k)] = paddle.metric.accuracy( if output_dims < k:
x, label, k=k) msg = f"The output dims({output_dims}) is less than k({k}), and the argument {k} of Topk has been removed."
self.avg_meters["top{}".format(k)].update(metric_dict["top{}".format(k)], x.shape[0]) logger.warning(msg)
self.avg_meters.pop(f"top{k}")
continue
metric_dict[f"top{k}"] = paddle.metric.accuracy(x, label, k=k)
self.avg_meters[f"top{k}"].update(metric_dict[f"top{k}"],
x.shape[0])
self.topk = list(filter(lambda k: k <= output_dims, self.topk))
return metric_dict return metric_dict
......
...@@ -62,8 +62,8 @@ def load_params(exe, prog, path, ignore_params=None): ...@@ -62,8 +62,8 @@ def load_params(exe, prog, path, ignore_params=None):
""" """
Load model from the given path. Load model from the given path.
Args: Args:
exe (fluid.Executor): The fluid.Executor object. exe (paddle.static.Executor): The paddle.static.Executor object.
prog (fluid.Program): load weight to which Program object. prog (paddle.static.Program): load weight to which Program object.
path (string): URL string or loca model path. path (string): URL string or loca model path.
ignore_params (list): ignore variable to load when finetuning. ignore_params (list): ignore variable to load when finetuning.
It can be specified by finetune_exclude_pretrained_params It can be specified by finetune_exclude_pretrained_params
......
...@@ -87,7 +87,7 @@ def main(args): ...@@ -87,7 +87,7 @@ def main(args):
'FLAGS_max_inplace_grad_add': 8, 'FLAGS_max_inplace_grad_add': 8,
} }
os.environ['FLAGS_cudnn_batchnorm_spatial_persistent'] = '1' os.environ['FLAGS_cudnn_batchnorm_spatial_persistent'] = '1'
paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING) paddle.set_flags(AMP_RELATED_FLAGS_SETTING)
use_xpu = global_config.get("use_xpu", False) use_xpu = global_config.get("use_xpu", False)
use_npu = global_config.get("use_npu", False) use_npu = global_config.get("use_npu", False)
......
...@@ -112,7 +112,7 @@ def get_path_from_url(url, ...@@ -112,7 +112,7 @@ def get_path_from_url(url,
str: a local path to save downloaded models & weights & datasets. str: a local path to save downloaded models & weights & datasets.
""" """
from paddle.fluid.dygraph.parallel import ParallelEnv from paddle.distributed import ParallelEnv
assert is_url(url), "downloading from {} not a url".format(url) assert is_url(url), "downloading from {} not a url".format(url)
# parse path after download to decompress under root_dir # parse path after download to decompress under root_dir
......
...@@ -35,18 +35,23 @@ ...@@ -35,18 +35,23 @@
│ ├── MobileNetV3 # MobileNetV3系列模型测试配置文件目录 │ ├── MobileNetV3 # MobileNetV3系列模型测试配置文件目录
│ │ ├── MobileNetV3_large_x1_0_train_infer_python.txt #基础训练预测配置文件 │ │ ├── MobileNetV3_large_x1_0_train_infer_python.txt #基础训练预测配置文件
│ │ ├── MobileNetV3_large_x1_0_train_linux_gpu_fleet_amp_infer_python_linux_gpu_cpu.txt #多机多卡训练预测配置文件 │ │ ├── MobileNetV3_large_x1_0_train_linux_gpu_fleet_amp_infer_python_linux_gpu_cpu.txt #多机多卡训练预测配置文件
│ │ └── MobileNetV3_large_x1_0_train_linux_gpu_normal_amp_infer_python_linux_gpu_cpu.txt #混合精度训练预测配置文件 │ │ ├── MobileNetV3_large_x1_0_train_linux_gpu_normal_amp_infer_python_linux_gpu_cpu.txt #混合精度训练预测配置文件
│ └── ResNet # ResNet系列模型测试配置文件目录 │ │ ├── MobileNetV3_large_x1_0_paddle2onnx_infer_python.txt #paddle2onnx推理测试配置文件
│ ├── ResNet50_vd_train_infer_python.txt #基础训练预测配置文件 │ │ └── ......
│ ├── ResNet50_vd_train_linux_gpu_fleet_amp_infer_python_linux_gpu_cpu.txt #多机多卡训练预测配置文件 │ ├──ResNet # ResNet系列模型测试配置文件目录
│ └── ResNet50_vd_train_linux_gpu_normal_amp_infer_python_linux_gpu_cpu.txt #混合精度训练预测配置文件 │ │ ├── ResNet50_vd_train_infer_python.txt #基础训练预测配置文件
| ...... │ │ ├── ResNet50_vd_train_linux_gpu_fleet_amp_infer_python_linux_gpu_cpu.txt #多机多卡训练预测配置文件
│ │ ├── ResNet50_vd_train_linux_gpu_normal_amp_infer_python_linux_gpu_cpu.txt #混合精度训练预测配置文件
│ │ ├── ResNet50_vd_paddle2onnx_infer_python.txt #paddle2onnx推理测试配置文件
│ │ └── ......
│ └── ......
├── docs ├── docs
│ ├── guide.png │ ├── guide.png
│ └── test.png │ └── test.png
├── prepare.sh # 完成test_*.sh运行所需要的数据和模型下载 ├── prepare.sh # 完成test_*.sh运行所需要的数据和模型下载
├── README.md # 使用文档 ├── README.md # 使用文档
├── results # 预先保存的预测结果,用于和实际预测结果进行精读比对 ├── results # 预先保存的预测结果,用于和实际预测结果进行精读比对
├── test_paddle2onnx.sh # 测试paddle2onnx推理预测的主程序
└── test_train_inference_python.sh # 测试python训练预测的主程序 └── test_train_inference_python.sh # 测试python训练预测的主程序
``` ```
...@@ -99,10 +104,12 @@ bash test_tipc/test_train_inference_python.sh ./test_tipc/configs/MobileNetV3/Mo ...@@ -99,10 +104,12 @@ bash test_tipc/test_train_inference_python.sh ./test_tipc/configs/MobileNetV3/Mo
## 4 开始测试 ## 4 开始测试
各功能测试中涉及混合精度、裁剪、量化等训练相关,及mkldnn、Tensorrt等多种预测相关参数配置,请点击下方相应链接了解更多细节和使用教程: 各功能测试中涉及混合精度、裁剪、量化等训练相关,及mkldnn、Tensorrt等多种预测相关参数配置,请点击下方相应链接了解更多细节和使用教程:
- [test_train_inference_python 使用](docs/test_train_inference_python.md):测试基于Python的模型训练、评估、推理等基本功能,包括裁剪、量化、蒸馏。 - [test_train_inference_python 使用](docs/test_train_inference_python.md):测试基于Python的模型训练、评估、推理等基本功能,包括裁剪、量化、蒸馏。
- [test_inference_cpp 使用](docs/test_inference_cpp.md) :测试基于C++的模型推理。 - [test_inference_cpp 使用](docs/test_inference_cpp.md) :测试基于C++的模型推理。
- [test_serving 使用](docs/test_serving.md) :测试基于Paddle Serving的服务化部署功能。 - [test_serving 使用](docs/test_serving.md) :测试基于Paddle Serving的服务化部署功能。
- [test_lite_arm_cpu_cpp 使用](docs/test_lite_arm_cpu_cpp.md): 测试基于Paddle-Lite的ARM CPU端c++预测部署功能. - [test_lite_arm_cpu_cpp 使用](docs/test_lite_arm_cpu_cpp.md): 测试基于Paddle-Lite的ARM CPU端c++预测部署功能.
- [test_paddle2onnx 使用](docs/test_paddle2onnx.md):测试Paddle2ONNX的模型转化功能,并验证正确性。 - [test_paddle2onnx 使用](docs/test_paddle2onnx.md):测试Paddle2ONNX的模型转化功能,并验证正确性。
- [test_serving_infer_python 使用](docs/test_serving_infer_python.md):测试python serving功能。
- [test_train_fleet_inference_python 使用](./docs/test_train_fleet_inference_python.md):测试基于Python的多机多卡训练与推理等基本功能。
===========================train_params===========================
model_name:GeneralRecognition_PPLCNet_x2_5
python:python3.7
gpu_list:192.168.0.1,192.168.0.2;0,1
-o Global.device:gpu
-o Global.auto_cast:null
-o Global.epochs:lite_train_lite_infer=2|whole_train_whole_infer=120
-o Global.output_dir:./output/
-o DataLoader.Train.sampler.batch_size:8
-o Global.pretrained_model:null
train_model_name:latest
train_infer_img_dir:./dataset/ILSVRC2012/val
null:null
##
trainer:norm_train
norm_train:tools/train.py -c ppcls/configs/GeneralRecognition/GeneralRecognition_PPLCNet_x2_5.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False
pact_train:null
fpgm_train:null
distill_train:null
null:null
null:null
##
===========================eval_params===========================
eval:tools/eval.py -c ppcls/configs/GeneralRecognition/GeneralRecognition_PPLCNet_x2_5.yaml
null:null
##
===========================infer_params==========================
-o Global.save_inference_dir:./inference
-o Global.pretrained_model:
norm_export:tools/export_model.py -c ppcls/configs/GeneralRecognition/GeneralRecognition_PPLCNet_x2_5.yaml
quant_export:null
fpgm_export:null
distill_export:null
kl_quant:null
export2:null
pretrained_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/pretrain/general_PPLCNet_x2_5_pretrained_v1.0.pdparams
infer_model:../inference/
infer_export:True
infer_quant:Fasle
inference:python/predict_rec.py -c configs/inference_rec.yaml
-o Global.use_gpu:True|False
-o Global.enable_mkldnn:True|False
-o Global.cpu_num_threads:1|6
-o Global.batch_size:1|16
-o Global.use_tensorrt:True|False
-o Global.use_fp16:True|False
-o Global.rec_inference_model_dir:../inference
-o Global.infer_imgs:../dataset/Aliproduct/demo_test/
-o Global.save_log_path:null
-o Global.benchmark:True
null:null
null:null
===========================infer_benchmark_params==========================
random_infer_input:[{float32,[3,224,224]}]
\ No newline at end of file
===========================cpp_infer_params===========================
model_name:MobileNetV3_large_x1_0
cpp_infer_type:cls
cls_inference_model_dir:./MobileNetV3_large_x1_0_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/MobileNetV3_large_x1_0_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:MobileNetV3_large_x1_0
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/MobileNetV3_large_x1_0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/MobileNetV3_large_x1_0_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/MobileNetV3_large_x1_0_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/MobileNetV3_large_x1_0_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:MobileNetV3_large_x1_0
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/MobileNetV3_large_x1_0_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/MobileNetV3_large_x1_0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/MobileNetV3_large_x1_0_serving/
--serving_client:./deploy/paddleserving/MobileNetV3_large_x1_0_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================paddle2onnx_params===========================
model_name:PP-ShiTu_general_rec
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/general_PPLCNet_x2_5_lite_v1.0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/general_PPLCNet_x2_5_lite_v1.0_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/general_PPLCNet_x2_5_lite_v1.0_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/general_PPLCNet_x2_5_lite_v1.0_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================cpp_infer_params=========================== ===========================cpp_infer_params===========================
model_name:PPShiTu model_name:PPShiTu
cpp_infer_type:shitu cpp_infer_type:shitu
feature_inference_model_dir:./feature_inference/ feature_inference_model_dir:./general_PPLCNet_x2_5_lite_v1.0_infer/
det_inference_model_dir:./det_inference det_inference_model_dir:./picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer/
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/general_PPLCNet_x2_5_lite_v1.0_infer.tar cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/general_PPLCNet_x2_5_lite_v1.0_infer.tar
det_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer.tar det_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer.tar
infer_quant:False infer_quant:False
inference_cmd:./deploy/cpp_shitu/build/pp_shitu -c inference_drink.yaml inference_cmd:./deploy/cpp_shitu/build/pp_shitu -c inference_drink.yaml
use_gpu:True|False use_gpu:True|False
enable_mkldnn:True|False enable_mkldnn:False
cpu_threads:1|6 cpu_threads:1
batch_size:1 batch_size:1
use_tensorrt:False|True use_tensorrt:False
precision:fp32|fp16 precision:fp32
data_dir:./dataset/drink_dataset_v1.0 data_dir:./dataset/drink_dataset_v1.0
benchmark:True benchmark:True
generate_yaml_cmd:python3 test_tipc/generate_cpp_yaml.py generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
transform_index_cmd:python3 deploy/cpp_shitu/tools/transform_id_map.py -c inference_drink.yaml transform_index_cmd:python3.7 deploy/cpp_shitu/tools/transform_id_map.py -c inference_drink.yaml
===========================serving_params===========================
model_name:PPShiTu
python:python3.7
cls_inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/general_PPLCNet_x2_5_lite_v1.0_infer.tar
det_inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./models/general_PPLCNet_x2_5_lite_v1.0_infer/
--dirname:./models/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./models/general_PPLCNet_x2_5_lite_v1.0_serving/
--serving_client:./models/general_PPLCNet_x2_5_lite_v1.0_client/
--serving_server:./models/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_serving/
--serving_client:./models/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_client/
serving_dir:./paddleserving/recognition
web_service:recognition_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================paddle2onnx_params===========================
model_name:PP-ShiTu_mainbody_det
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================cpp_infer_params===========================
model_name:PPHGNet_small
cpp_infer_type:cls
cls_inference_model_dir:./PPHGNet_small_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPHGNet_small_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PPHGNet_small
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/PPHGNet_small_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/PPHGNet_small_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/PPHGNet_small_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPHGNet_small
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPHGNet_small_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPHGNet_small_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPHGNet_small_serving/
--serving_client:./deploy/paddleserving/PPHGNet_small_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================train_params===========================
model_name:PPHGNet_small
python:python3.7
gpu_list:192.168.0.1,192.168.0.2;0,1
-o Global.device:gpu
-o Global.auto_cast:null
-o Global.epochs:lite_train_lite_infer=2|whole_train_whole_infer=120
-o Global.output_dir:./output/
-o DataLoader.Train.sampler.batch_size:8
-o Global.pretrained_model:null
train_model_name:latest
train_infer_img_dir:./dataset/ILSVRC2012/val
null:null
##
trainer:norm_train
norm_train:tools/train.py -c ppcls/configs/ImageNet/PPHGNet/PPHGNet_small.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False
pact_train:null
fpgm_train:null
distill_train:null
null:null
null:null
##
===========================eval_params===========================
eval:tools/eval.py -c ppcls/configs/ImageNet/PPHGNet/PPHGNet_small.yaml
null:null
##
===========================infer_params==========================
-o Global.save_inference_dir:./inference
-o Global.pretrained_model:
norm_export:tools/export_model.py -c ppcls/configs/ImageNet/PPHGNet/PPHGNet_small.yaml
quant_export:null
fpgm_export:null
distill_export:null
kl_quant:null
export2:null
pretrained_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/legendary_models/PPHGNet_small_pretrained.pdparams
infer_model:../inference/
infer_export:True
infer_quant:Fasle
inference:python/predict_cls.py -c configs/inference_cls.yaml -o PreProcess.transform_ops.0.ResizeImage.resize_short=236
-o Global.use_gpu:True|False
-o Global.enable_mkldnn:True|False
-o Global.cpu_num_threads:1|6
-o Global.batch_size:1|16
-o Global.use_tensorrt:True|False
-o Global.use_fp16:True|False
-o Global.inference_model_dir:../inference
-o Global.infer_imgs:../dataset/ILSVRC2012/val
-o Global.save_log_path:null
-o Global.benchmark:True
null:null
===========================infer_benchmark_params==========================
random_infer_input:[{float32,[3,224,224]}]
===========================cpp_infer_params===========================
model_name:PPHGNet_tiny
cpp_infer_type:cls
cls_inference_model_dir:./PPHGNet_tiny_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPHGNet_tiny_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PPHGNet_tiny
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/PPHGNet_tiny_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/PPHGNet_tiny_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/PPHGNet_tiny_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPHGNet_tiny
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPHGNet_tiny_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPHGNet_tiny_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPHGNet_tiny_serving/
--serving_client:./deploy/paddleserving/PPHGNet_tiny_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================cpp_infer_params===========================
model_name:PPLCNet_x0_25
cpp_infer_type:cls
cls_inference_model_dir:./PPLCNet_x0_25_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_25_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PPLCNet_x0_25
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/PPLCNet_x0_25_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/PPLCNet_x0_25_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/PPLCNet_x0_25_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPLCNet_x0_25
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_25_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPLCNet_x0_25_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPLCNet_x0_25_serving/
--serving_client:./deploy/paddleserving/PPLCNet_x0_25_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================cpp_infer_params===========================
model_name:PPLCNet_x0_35
cpp_infer_type:cls
cls_inference_model_dir:./PPLCNet_x0_35_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_35_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PPLCNet_x0_25
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/PPLCNet_x0_25_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/PPLCNet_x0_25_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_25_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/PPLCNet_x0_25_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPLCNet_x0_35
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_35_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPLCNet_x0_35_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPLCNet_x0_35_serving/
--serving_client:./deploy/paddleserving/PPLCNet_x0_35_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================cpp_infer_params===========================
model_name:PPLCNet_x0_5
cpp_infer_type:cls
cls_inference_model_dir:./PPLCNet_x0_5_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_5_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PP-ShiTu_mainbody_det
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/rec/models/inference/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/picodet_PPLCNet_x2_5_mainbody_lite_v1.0_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPLCNet_x0_5
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_5_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPLCNet_x0_5_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPLCNet_x0_5_serving/
--serving_client:./deploy/paddleserving/PPLCNet_x0_5_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
\ No newline at end of file
===========================cpp_infer_params===========================
model_name:PPLCNet_x0_75
cpp_infer_type:cls
cls_inference_model_dir:./PPLCNet_x0_75_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_75_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PPLCNet_x0_75
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/PPLCNet_x0_75_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/PPLCNet_x0_75_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_75_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/PPLCNet_x0_75_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPLCNet_x0_75
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x0_75_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPLCNet_x0_75_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPLCNet_x0_75_serving/
--serving_client:./deploy/paddleserving/PPLCNet_x0_75_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================cpp_infer_params===========================
model_name:PPLCNet_x1_0
cpp_infer_type:cls
cls_inference_model_dir:./PPLCNet_x1_0_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x1_0_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PPLCNet_x1_0
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/PPLCNet_x1_0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/PPLCNet_x1_0_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x1_0_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/PPLCNet_x1_0_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPLCNet_x1_0
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x1_0_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPLCNet_x1_0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPLCNet_x1_0_serving/
--serving_client:./deploy/paddleserving/PPLCNet_x1_0_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================train_params===========================
model_name:PPLCNet_x1_0
python:python3.7
gpu_list:192.168.0.1,192.168.0.2;0,1
-o Global.device:gpu
-o Global.auto_cast:null
-o Global.epochs:lite_train_lite_infer=2|whole_train_whole_infer=120
-o Global.output_dir:./output/
-o DataLoader.Train.sampler.batch_size:8
-o Global.pretrained_model:null
train_model_name:latest
train_infer_img_dir:./dataset/ILSVRC2012/val
null:null
##
trainer:norm_train
norm_train:tools/train.py -c ppcls/configs/ImageNet/PPLCNet/PPLCNet_x1_0.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False
pact_train:null
fpgm_train:null
distill_train:null
null:null
null:null
##
===========================eval_params===========================
eval:tools/eval.py -c ppcls/configs/ImageNet/PPLCNet/PPLCNet_x1_0.yaml
null:null
##
===========================infer_params==========================
-o Global.save_inference_dir:./inference
-o Global.pretrained_model:
norm_export:tools/export_model.py -c ppcls/configs/ImageNet/PPLCNet/PPLCNet_x1_0.yaml
quant_export:null
fpgm_export:null
distill_export:null
kl_quant:null
export2:null
pretrained_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/legendary_models/PPLCNet_x1_0_pretrained.pdparams
infer_model:../inference/
infer_export:True
infer_quant:Fasle
inference:python/predict_cls.py -c configs/inference_cls.yaml
-o Global.use_gpu:True|False
-o Global.enable_mkldnn:True|False
-o Global.cpu_num_threads:1|6
-o Global.batch_size:1|16
-o Global.use_tensorrt:True|False
-o Global.use_fp16:True|False
-o Global.inference_model_dir:../inference
-o Global.infer_imgs:../dataset/ILSVRC2012/val
-o Global.save_log_path:null
-o Global.benchmark:True
null:null
===========================infer_benchmark_params==========================
random_infer_input:[{float32,[3,224,224]}]
\ No newline at end of file
===========================cpp_infer_params===========================
model_name:PPLCNet_x1_5
cpp_infer_type:cls
cls_inference_model_dir:./PPLCNet_x1_5_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x1_5_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PPLCNet_x1_5
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/PPLCNet_x1_5_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/PPLCNet_x1_5_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x1_5_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/PPLCNet_x1_5_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPLCNet_x1_5
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x1_5_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPLCNet_x1_5_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPLCNet_x1_5_serving/
--serving_client:./deploy/paddleserving/PPLCNet_x1_5_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================cpp_infer_params===========================
model_name:PPLCNet_x2_0
cpp_infer_type:cls
cls_inference_model_dir:./PPLCNet_x2_0_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x2_0_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PPLCNet_x2_0
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/PPLCNet_x2_0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/PPLCNet_x2_0_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x2_0_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/PPLCNet_x2_0_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPLCNet_x2_0
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x2_0_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPLCNet_x2_0_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPLCNet_x2_0_serving/
--serving_client:./deploy/paddleserving/PPLCNet_x2_0_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================cpp_infer_params===========================
model_name:PPLCNet_x2_5
cpp_infer_type:cls
cls_inference_model_dir:./PPLCNet_x2_5_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x2_5_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
===========================paddle2onnx_params===========================
model_name:PPLCNet_x2_5
python:python3.7
2onnx: paddle2onnx
--model_dir:./deploy/models/PPLCNet_x2_5_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--save_file:./deploy/models/PPLCNet_x2_5_infer/inference.onnx
--opset_version:10
--enable_onnx_checker:True
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x2_5_infer.tar
inference:./python/predict_cls.py
Global.use_onnx:True
Global.inference_model_dir:./models/PPLCNet_x2_5_infer
Global.use_gpu:False
-c:configs/inference_cls.yaml
\ No newline at end of file
===========================serving_params===========================
model_name:PPLCNet_x2_5
python:python3.7
inference_model_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNet_x2_5_infer.tar
trans_model:-m paddle_serving_client.convert
--dirname:./deploy/paddleserving/PPLCNet_x2_5_infer/
--model_filename:inference.pdmodel
--params_filename:inference.pdiparams
--serving_server:./deploy/paddleserving/PPLCNet_x2_5_serving/
--serving_client:./deploy/paddleserving/PPLCNet_x2_5_client/
serving_dir:./deploy/paddleserving
web_service:classification_web_service.py
--use_gpu:0|null
pipline:pipeline_http_client.py
===========================cpp_infer_params===========================
model_name:PPLCNetV2_base
cpp_infer_type:cls
cls_inference_model_dir:./PPLCNetV2_base_infer/
det_inference_model_dir:
cls_inference_url:https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/inference/PPLCNetV2_base_infer.tar
det_inference_url:
infer_quant:False
inference_cmd:./deploy/cpp/build/clas_system -c inference_cls.yaml
use_gpu:True|False
enable_mkldnn:False
cpu_threads:1
batch_size:1
use_tensorrt:False
precision:fp32
image_dir:./dataset/ILSVRC2012/val/ILSVRC2012_val_00000001.JPEG
benchmark:False
generate_yaml_cmd:python3.7 test_tipc/generate_cpp_yaml.py
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册