修改说明

基于原版Nanodet修改，加入mosaic,mixup数据增强，删除contrast、brightness、saturation数据增强，添加gamma调整、hsv调整数据增强。关闭normalize，需要开启的可以自行修改nanodet/data/transform/color.py。在config目录下有修改后的配置文件。
修改了导出onnx逻辑，聚合了输出层。预测代码可参考 https://github.com/linghu8812/tensorrt_inference/tree/master/nanodet （注意：如果没有打开normalize，需要修改预处理部分代码）原repo #65 issue
原repo https://github.com/RangiLyu/nanodet

NanoDet

Super fast and lightweight anchor-free object detection model. Real-time on mobile devices.

⚡Super lightweight: Model file is only 1.8 MB.
⚡Super fast: 97fps(10.23ms) on mobile ARM CPU.
😎Training friendly: Much lower GPU memory cost than other models. Batch-size=80 is available on GTX1060 6G.
😎Easy to deploy: Provide C++ implementation and Android demo based on ncnn inference framework.

NEWS!!!

[2021.03.12] Apply the Transformer encoder to NanoDet! Introducing NanoDet-t, which replaces the PAN in NanoDet-m with a TAN(Transformer Attention Net), gets 21.7 mAP(+1.1) on COCO val 2017. Check nanodet-t.yml for more details.
[2021.03.03] Update Nanodet-m-416 COCO pretrained model. COCO mAP(0.5:0.95)=23.5. Download in Model Zoo.
[2021.02.03] Support EfficientNet-Lite and Rep-VGG backbone. Please check the config folder. Download models in Model Zoo
[2021.01.10] NanoDet-g with lower memory access cost, which designed for edge NPU or GPU, is now available! Check config/nanodet-g.yml and download in Model Zoo.

More...

[2020.12.19] MNN python and cpp demos are available.
[2020.12.05] Support voc .xml format dataset! Refer to config/nanodet_custom_xml_dataset.yml.
[2020.12.01] Great thanks to nihui, now you can try NanoDet running in web browser! 👉 https://nihui.github.io/ncnn-webassembly-nanodet/

Benchmarks

Model	Resolution	COCO mAP	Latency(ARM 4xCore)	FLOPS	Params	Model Size(ncnn fp16)
NanoDet-m	320*320	20.6	10.23ms	0.72B	0.95M	1.8MB
NanoDet-m	416*416	23.5	16.44ms	1.2B	0.95M	1.8MB
NanoDet-g	416*416	22.9	Not Designed For ARM	4.2B	3.81M	7.7MB
YoloV3-Tiny	416*416	16.6	37.6ms	5.62B	8.86M	33.7MB
YoloV4-Tiny	416*416	21.7	32.81ms	6.96B	6.06M	23.0MB

Note:

Performance is measured on Kirin 980(4xA76+4xA55) ARM CPU based on ncnn. You can test latency on your phone with ncnn_android_benchmark.
NanoDet mAP(0.5:0.95) is validated on COCO val2017 dataset with no testing time augmentation.
YOLO mAP refers from Scaled-YOLOv4: Scaling Cross Stage Partial Network.
NanoDet-g is designed for edge NPU, GPU or TPU with high parallel computing power but low memory bandwidth. It has much lower memory access cost than NanoDet-m.

NanoDet is a FCOS-style one-stage anchor-free object detection model which using ATSS for target sampling and using Generalized Focal Loss for classification and box regression. Please refer to these papers for more details.

Fcos: Fully convolutional one-stage object detection

ATSS:Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection

Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection

知乎中文介绍 | QQ交流群：908606542 (答案：炼丹)

Demo

Android demo

Android demo project is in demo_android_ncnn folder. Please refer to Android demo guide.

Here is a better implementation 👉 ncnn-android-nanodet

NCNN C++ demo

C++ demo based on ncnn is in demo_ncnn folder. Please refer to Cpp demo guide.

MNN demo

Inference using Alibaba's MNN framework is in demo_mnn folder. Including python and cpp inference code. Please refer to MNN demo guide.

Pytorch demo

First, install requirements and setup NanoDet following installation guide. Then download COCO pretrain weight from here

👉COCO pretrain weight (Google Drive)

Inference images

python demo/demo.py image --config CONFIG_PATH --model MODEL_PATH --path IMAGE_PATH

Inference video

python demo/demo.py video --config CONFIG_PATH --model MODEL_PATH --path VIDEO_PATH

Inference webcam

python demo/demo.py webcam --config CONFIG_PATH --model MODEL_PATH --camid YOUR_CAMERA_ID

Besides, We provide a notebook here to demonstrate how to make it work with PyTorch.

Install

Requirements

Linux or MacOS
CUDA >= 10.0
Python >= 3.6
Pytorch >= 1.6
experimental support Windows (Notice: Windows not support distributed training before pytorch1.7)

Step

Create a conda virtual environment and then activate it.

 conda create -n nanodet python=3.8 -y
 conda activate nanodet

Install pytorch

conda install pytorch torchvision cudatoolkit=11.1 -c pytorch -c conda-forge

Install requirements

pip install Cython termcolor numpy tensorboard pycocotools matplotlib pyaml opencv-python tqdm pytorch-lightning torchmetrics

Setup NanoDet

git clone https://github.com/RangiLyu/nanodet.git
cd nanodet
python setup.py develop

Model Zoo

NanoDet supports variety of backbones. Go to the config folder to see the sample training config files.

Model	Backbone	Resolution	COCO mAP	FLOPS	Params	Pre-train weight
NanoDet-m	ShuffleNetV2 1.0x	320*320	20.6	0.72B	0.95M	Download
NanoDet-m-416	ShuffleNetV2 1.0x	416*416	23.5	1.2B	0.95M	Download
NanoDet-t (NEW)	ShuffleNetV2 1.0x	320*320	21.7	0.96B	1.36M	Download
NanoDet-g	Custom CSP Net	416*416	22.9	4.2B	3.81M	Download
NanoDet-EfficientLite	EfficientNet-Lite0	320*320	24.7	1.72B	3.11M	Download
NanoDet-EfficientLite	EfficientNet-Lite1	416*416	30.3	4.06B	4.01M	Download
NanoDet-EfficientLite	EfficientNet-Lite2	512*512	32.6	7.12B	4.71M	Download
NanoDet-RepVGG	RepVGG-A0	416*416	27.8	11.3B	6.75M	Download

How to Train

Prepare dataset

If your dataset annotations are pascal voc xml format, refer to config/nanodet_custom_xml_dataset.yml

Or convert your dataset annotations to MS COCO format(COCO annotation format details).
Prepare config file

Copy and modify an example yml config file in config/ folder.

Change save_path to where you want to save model.

Change num_classes in model->arch->head.

Change image path and annotation path in both data->train and data->val.

Set gpu ids, num workers and batch size in device to fit your device.

Set total_epochs, lr and lr_schedule according to your dataset and batchsize.

If you want to modify network, data augmentation or other things, please refer to Config File Detail
Start training

NanoDet is now using pytorch lightning for training.

For both single-GPU or multiple-GPUs, run:
```
python tools/train.py CONFIG_FILE_PATH
```
Old training script is deprecated and will be deleted in next version. If you still want to use,
follow this...

For single GPU, run
```
python tools/deprecated/train.py CONFIG_FILE_PATH
```
For multi-GPU, NanoDet using distributed training. (Notice: Windows not support distributed training before pytorch1.7) Please run
```
python -m torch.distributed.launch --nproc_per_node=GPU_NUM --master_port 29501 tools/deprecated/train.py CONFIG_FILE_PATH
```
Visualize Logs

TensorBoard logs are saved in save_dir which you set in config file.

To visualize tensorboard logs, run:
```
cd <YOUR_SAVE_DIR>
tensorboard --logdir ./
```

How to Deploy

NanoDet provide C++ and Android demo based on ncnn library.

Convert model

To convert NanoDet pytorch model to ncnn, you can choose this way: pytorch->onnx->ncnn

To export onnx model, run tools/export_onnx.py.
```
python tools/export_onnx.py --cfg_path ${CONFIG_PATH} --model_path ${PYTORCH_MODEL_PATH}
```
Then using onnx-simplifier to simplify onnx structure.
```
python -m onnxsim ${INPUT_ONNX_MODEL} ${OUTPUT_ONNX_MODEL}
```
Run onnx2ncnn in ncnn tools to generate ncnn .param and .bin file.

After that, using ncnnoptimize to optimize ncnn model.

If you have quentions about converting ncnn model, refer to ncnn wiki. https://github.com/Tencent/ncnn/wiki
Run NanoDet model with C++

Please refer to demo_ncnn.
Run NanoDet on Android

Please refer to android_demo.