- 1. Introduction
- 2. Quick Start
- 3. Training, Evaluation and Inference
- 4. Model Compression
- 5. SHAS
- 6. Inference Deployment
This case provides a way for users to quickly build a lightweight, high-precision and practical classification model of human exists using PaddleClas PULC (Practical Ultra Lightweight image Classification). The model can be widely used in monitoring scenarios, personnel access control scenarios, massive data filtering scenarios, etc.
The following table lists the relevant indicators of the model. The first two lines means that using SwinTransformer_tiny and MobileNetV3_small_x0_35 as the backbone to training. The third to sixth lines means that the backbone is replaced by PPLCNet, additional use of EDA strategy and additional use of EDA strategy and SKL-UGI knowledge distillation strategy.
Backbone | Tpr(%) | Latency(ms) | Size(M) | Training Strategy |
---|---|---|---|---|
SwinTranformer_tiny | 95.69 | 95.30 | 111 | using ImageNet pretrained model |
MobileNetV3_small_x0_35 | 68.25 | 2.85 | 2.6 | using ImageNet pretrained model |
PPLCNet_x1_0 | 89.57 | 2.12 | 7.0 | using ImageNet pretrained model |
PPLCNet_x1_0 | 92.10 | 2.12 | 7.0 | using SSLD pretrained model |
PPLCNet_x1_0 | 93.43 | 2.12 | 7.0 | using SSLD pretrained model + EDA strategy |
PPLCNet_x1_0 | 96.23 | 2.12 | 7.0 | using SSLD pretrained model + EDA strategy + SKL-UGI knowledge distillation strategy |
It can be seen that high Tpr can be getted when backbone is SwinTranformer_tiny, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the Tpr will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0, the Tpr is higher more 20 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more than 20% faster. After additional using the SSLD pretrained model, the Tpr can be improved by about 2.6 percentage points without affecting the inference speed. Further, additional using the EDA strategy, the Tpr can be increased by 1.3 percentage points. Finally, after additional using the SKL-UGI knowledge distillation, the Tpr can be further improved by 2.8 percentage points. At this point, the Tpr is close to that of SwinTranformer_tiny, but the speed is more than 40 times faster. The training method and deployment instructions of PULC will be introduced in detail below.
Note:
- About
Tpr
metric, please refer to 3.2 section for more information . - The Latency is tested on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz. The MKLDNN is enabled and the number of threads is 10.
- About PP-LCNet, please refer to PP-LCNet Introduction and PP-LCNet Paper.
- Run the following command to install if CUDA9 or CUDA10 is available.
python3 -m pip install paddlepaddle-gpu -i https://mirror.baidu.com/pypi/simple
- Run the following command to install if GPU device is unavailable.
python3 -m pip install paddlepaddle -i https://mirror.baidu.com/pypi/simple
Please refer to PaddlePaddle Installation for more information about installation, for examples other versions.
The command of PaddleClas installation as bellow:
pip3 install paddleclas
First, please click here to download and unzip to get the test demo images.
- Prediction with CLI
paddleclas --model_name=person_exists --infer_imgs=pulc_demo_imgs/person_exists/objects365_01780782.jpg
Results:
>>> result
class_ids: [0], scores: [0.9955421453341842], label_names: ['nobody'], filename: pulc_demo_imgs/person_exists/objects365_01780782.jpg
Predict complete!
Note: If you want to test other images, only need to specify the --infer_imgs
argument, and the directory containing images is also supported.
- Prediction in Python
import paddleclas
model = paddleclas.PaddleClas(model_name="person_exists")
result = model.predict(input_data="pulc_demo_imgs/person_exists/objects365_01780782.jpg")
print(next(result))
Note: The result
returned by model.predict()
is a generator, so you need to use the next()
function to call it or for
loop to loop it. And it will predict with batch_size
size batch and return the prediction results when called. The default batch_size
is 1, and you also specify the batch_size
when instantiating, such as model = paddleclas.PaddleClas(model_name="person_exists", batch_size=2)
. The result of demo above:
>>> result
[{'class_ids': [0], 'scores': [0.9955421453341842], 'label_names': ['nobody'], 'filename': 'pulc_demo_imgs/person_exists/objects365_01780782.jpg'}]
Please refer to Installation to get the description about installation.
All datasets used in this case are open source data. Train data is the subset of MS-COCO training data. And the validation data is the subset of Object365 training data. ImageNet_val is ImageNet-1k validation data.
The data used in this case can be getted by processing the open source data. The detailed processes are as follows:
- Training data. This case deals with the annotation file of MS-COCO data training data. If a certain image contains the label of "person" and the area of this box is greater than 10% in the whole image, it is considered that the image contains human. If there is no label of "person" in a certain image, It is considered that the image does not contain human. After processing, 92964 pieces of available data were obtained, including 39813 images containing human and 53151 images without containing human.
- Validation data: randomly select a small part of data from object365 data, use the better model trained on MS-COCO to predict these data, take the intersection between the prediction results and the data annotation file, and filter the intersection results into the validation set according to the method of obtaining the training set. After processing, 27820 pieces of available data were obtained. There are 2255 pieces of data with human and 25565 pieces of data without human. The data visualization of the processed dataset is as follows:
Some image of the processed dataset is as follows:
And you can also download the data processed directly.
cd path_to_PaddleClas
Enter the dataset/
directory, download and unzip the dataset.
cd dataset
wget https://paddleclas.bj.bcebos.com/data/PULC/person_exists.tar
tar -xf person_exists.tar
cd ../
The datas under person_exists
directory:
├── 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
Where train/
and val/
are training set and validation set respectively. The train_list.txt
and val_list.txt
are label files of training data and validation data respectively. The file train_list.txt.debug
and val_list.txt.debug
are subset of train_list.txt
and val_list.txt
respectively. ImageNet_val/
is the validation data of ImageNet-1k, which will be used for SKL-UGI knowledge distillation, and its label file is train_list_for_distill.txt
.
Note:
- About the contents format of
train_list.txt
andval_list.txt
, please refer to Description about Classification Dataset in PaddleClas. - About the
train_list_for_distill.txt
, please refer to Knowledge Distillation Label.
The details of training config in ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml
. The command about training as follows:
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
The best metric of validation data is between 0.94
and 0.95
. There would be fluctuations because the data size is small.
Note:
- The metric Tpr, that describe the True Positive Rate when False Positive Rate is less than a certain threshold(1/1000 used in this case), is one of the commonly used metric for binary classification. About the details of Fpr and Tpr, please refer here.
- When evaluation, the best metric TprAtFpr will be printed that include
Fpr
,Tpr
and the currentthreshold
. TheTpr
means the Recall rate under the currentFpr
. TheTpr
higher, the model better. Thethreshold
would be used in deployment, which means the classification threshold under bestFpr
metric.
After training, you can use the following commands to evaluate the model.
python3 tools/eval.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml \
-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"
Among the above command, the argument -o Global.pretrained_model="output/PPLCNet_x1_0/best_model"
specify the path of the best model weight file. You can specify other path if needed.
After training, you can use the model that trained to infer. Command is as follow:
python3 tools/infer.py \
-c ./ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml \
-o Global.pretrained_model=output/PPLCNet_x1_0/best_model
The results:
[{'class_ids': [1], 'scores': [0.9999976], 'label_names': ['someone'], 'file_name': 'deploy/images/PULC/person_exists/objects365_02035329.jpg'}]
Note:
- Among the above command, argument
-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"
specify the path of the best model weight file. You can specify other path if needed. - The default test image is
deploy/images/PULC/person_exists/objects365_02035329.jpg
. And you can test other image, only need to specify the argument-o Infer.infer_imgs=path_to_test_image
. - The default threshold is
0.5
. If needed, you can specify the argumentInfer.PostProcess.threshold
, such as:-o Infer.PostProcess.threshold=0.9794
. And the argumentthreshold
is needed to be specified according by specific case. The0.9794
is the best threshold whenFpr
is less than1/1000
in this valuation dataset.
SKL-UGI is a simple but effective knowledge distillation algrithem proposed by PaddleClas.
Training the teacher model with hyperparameters specified in ppcls/configs/PULC/person_exists/PPLCNet/PPLCNet_x1_0.yaml
. The command is as follow:
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 \
-o Arch.name=ResNet101_vd
The best metric of validation data is between 0.96
and 0.98
. The best teacher model weight would be saved in file output/ResNet101_vd/best_model.pdparams
.
The training strategy, specified in training config file ppcls/configs/PULC/person_exists/PPLCNet_x1_0_distillation.yaml
, the teacher model is ResNet101_vd
, the student model is PPLCNet_x1_0
and the additional unlabeled training data is validation data of ImageNet1k. The command is as follow:
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
The best metric is between 0.95
and 0.97
. The best student model weight would be saved in file output/DistillationModel/best_model_student.pdparams
.
The hyperparameters used by 3.2 section and 4.1 section are according by Hyperparameters Searching
in PaddleClas. If you want to get better results on your own dataset, you can refer to Hyperparameters Searching to get better hyperparameters.
Note: This section is optional. Because the search process will take a long time, you can selectively run according to your specific. If not replace the dataset, you can ignore this section.
Paddle Inference is the original Inference Library of the PaddlePaddle, provides high-performance inference for server deployment. And compared with directly based on the pretrained model, Paddle Inference can use tools to accelerate prediction, so as to achieve better inference performance. Please refer to Paddle Inference for more information.
Paddle Inference need Paddle Inference Model to predict. Two process provided to get Paddle Inference Model. If want to use the provided by PaddleClas, you can download directly, click Downloading Inference Model.
The command about exporting Paddle Inference Model is as follow:
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
After running above command, the inference model files would be saved in deploy/models/PPLCNet_x1_0_person_exists_infer
, as shown below:
├── PPLCNet_x1_0_person_exists_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
Note: The best model is from knowledge distillation training. If knowledge distillation training is not used, the best model would be saved in output/PPLCNet_x1_0/best_model.pdparams
.
You can also download directly.
cd deploy/models
# download the inference model and decompression
wget https://paddleclas.bj.bcebos.com/models/PULC/person_exists_infer.tar && tar -xf person_exists_infer.tar
After decompression, the directory models
should be shown below.
├── person_exists_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
Return the directory deploy
:
cd ../
Run the following command to classify whether there are humans in the image ./images/PULC/person_exists/objects365_02035329.jpg
.
# Use the following command to predict with GPU.
python3.7 python/predict_cls.py -c configs/PULC/person_exists/inference_person_exists.yaml
# Use the following command to predict with CPU.
python3.7 python/predict_cls.py -c configs/PULC/person_exists/inference_person_exists.yaml -o Global.use_gpu=False
The prediction results:
objects365_02035329.jpg: class id(s): [1], score(s): [1.00], label_name(s): ['someone']
Note: The default threshold is 0.5
. If needed, you can specify the argument Infer.PostProcess.threshold
, such as: -o Infer.PostProcess.threshold=0.9794
. And the argument threshold
is needed to be specified according by specific case. The 0.9794
is the best threshold when Fpr
is less than 1/1000
in this valuation dataset. Please refer to 3.3 section for details.
If you want to predict images in directory, please specify the argument Global.infer_imgs
as directory path by -o Global.infer_imgs
. The command is as follow.
# Use the following command to predict with GPU. If want to replace with CPU, you can add argument -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/"
All prediction results will be printed, as shown below.
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']
Among the prediction results above, someone
means that there is a human in the image, nobody
means that there is no human in the image.
PaddleClas provides an example about how to deploy with C++. Please refer to Deployment with C++.
Paddle Serving is a flexible, high-performance carrier for machine learning models, and supports different protocol, such as RESTful, gRPC, bRPC and so on, which provides different deployment solutions for a variety of heterogeneous hardware and operating system environments. Please refer Paddle Serving for more information.
PaddleClas provides an example about how to deploy as service by Paddle Serving. Please refer to Paddle Serving Deployment.
Paddle-Lite is an open source deep learning framework that designed to make easy to perform inference on mobile, embeded, and IoT devices. Please refer to Paddle-Lite for more information.
PaddleClas provides an example of how to deploy on mobile by Paddle-Lite. Please refer to Paddle-Lite deployment.
Paddle2ONNX support convert Paddle Inference model to ONNX model. And you can deploy with ONNX model on different inference engine, such as TensorRT, OpenVINO, MNN/TNN, NCNN and so on. About Paddle2ONNX details, please refer to Paddle2ONNX.
PaddleClas provides an example of how to convert Paddle Inference model to ONNX model by paddle2onnx toolkit and predict by ONNX model. You can refer to paddle2onnx for deployment details.