quarkdet

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #android #rust #deep-learning #linux #docker #windows #kubernetes #vue #pytorch #mcp

(★ 92)

QuarkDet lightweight object detection in PyTorch .Real-Time Object Detection on Mobile Devices.

quarkdet 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# QuarkDet implementation lightweight object detection based on PyTorch

[中文(有更多的详细资料)](./README_cn.md)  [English](./README.md)

Here we implement lightweight object detection 
backbone support mobilenetv3、shufflenetv2、ghostnet、efficientnet 
neck support FPN（cnn）,PAN（cnn）、FPN_Slim（non-cnn），PAN_Slim（non-cnn）、BiFPN 
head support gfl（Generalized Focal Loss）、gfl v2(custom)

# Test Environment

Ubuntu18.04 
PyTorch 1.7 
Python 3.6

Different devices have different performance on the mobile side, and models with different capabilities can be run according to different performance. The library provides models for low-power mobile devices and powerful mobile devices. 
If the model runs on a low-power mobile device, backbone can choose shufflenetv2 or moiblenetv3, and neck choosing FAN_Slim, is equivalent to an implementation of nanodet. 
If it is a mobile device with strong performance, backbone can choose to only write b2 and b3 types in the effcientnet, configuration, and neck can choose BiFPN, to be the equivalent of an effcientdet implementation.

Backbone support mobilenetv3、shufflenetv2、ghostnet、efficientnet 
neck support FPN（Convolution）,PAN（Convolution）、FPN_Slim（Non-convolution），PAN_Slim（Non-convolution）、BiFPN 
head support gfl（Generalized Focal Loss）、gfl v2(Custom version)

quarkdet can use the following permutations and combinations 
EfficientNet + BiFPN + GFL 
GhostNet + PAN + GFL 
GhostNet + BiFPN + GFL 
MobileNetV3 + PAN/PAN_Slim + GFLv2 
ShuffleNetV2 + PAN/PAN_Slim + GFL and so on. 
Just change the command line to a different configuration file during training， config file in the quarkdet/config folder

EfficientDet the original implementation is 
EfficientNet + BiFPN + Box/Class Head 
This place has been changed 
EfficientNet + BiFPN + GFL(Generalized Focal Loss)

## support mosaic Data Augmentation

```
load_mosaic=False,
mosaic_probability=0.2
mosaic_area=16,
```

Quarkdet support to configure in the file, the sample file config/ghostnet_slim640.yml and the original mosaic data enhancement is different, not randomly change the size of the picture, but 4 pictures of the same size, using a fixed center, that is, 4 pictures, equal size, support 320 * 320,416 * 416,640 * 640 equal width and height of the same size. 
Load_mosaic: indicates whether to start data enhancement. 
What percentage of mosaic_probability: data is enhanced by mosaic data 
If the mosaic_area:GT bbox size is less than this threshold, it will be filtered out.

## gfl v2 version

Slightly different from the implementation of the official website, it can rise a little, only after the decimal point. 
From class QuarkDetHead (GFLHead): # you can directly replace GFLHead with GFLHeadV2 
For the original GFocalV2 implementation, please [reference](https://github.com/implus/GFocalV2)。

# GhostNet is divided into full version and simplified version.

config file 
ghostnet_full.yml full network version
ghostnet_slim.yml simplified network version

GhostNet following streamlines have been made 
Remove all the layers in the stage5 whose expansion size is equal to 960. the removed layers also include .
Conv2d 1 × 1 the number of output channels equals 960 and 1280 layers, average pooling layer and last fully connected layer

## mobilenetv3small version

The network is intercepted from the beginning to hs2 (bn2)

## Train method

For Single-GPU config 
quarkdet.yml config example 
device: 
&emsp; gpu_ids: [0]

### For Single-GPU run

```
python tools/train.py config/quarkdet.yml
```

For Multi-GPU config 
quarkdet.yml config example 
device: 
&emsp; gpu_ids: [0,1]

### For Multi-GPU run

```
python -m torch.distributed.launch --nproc_per_node=2 --master_port 30001 tools/train.py config/quarkdet.yml
```

## Inference video

It can be used for demonstration

```
python ./demo/demo.py  'video' --path /media/ubuntu/data/1.mp4 --config config/efficientdet.yml --model ./workspace/efficientdet/model_best/model_best.pth
```

## Learning rate support ReduceLROnPlateau

When the continuous n times of the monitored index has not been improved, the learning rate is reduced, where n is patience in the configuration.

```
  lr_schedule:
    name: ReduceLROnPlateau
    mode: min
    factor: 0.1
    patience: 10
    verbose: True
    threshold: 0.00001
    threshold_mode: rel
    cooldown: 0
    min_lr: 0
    eps: 0.000000001 #1e-08
```

## ghostnet_full result

Resolution=320 * 320
epoch = 85

## ghostnet_slim result

Computational complexity: 0.56 GFLOPs 
Number of parameters: 1.77 M

If it is distributed training, it can be done in norm_cfg.
Dict (type='BN', momentum=0.01,eps=1e-3, requires_grad=True).
Type='BN' changed to type='SyncBN'
Because no judgment is made here whether it is distributed or not, BN is written in between.

## efficientdet

EfficientNet + BiFPN + GFL 
The original was a feature of 5 level, which was reduced to 3 level here. 
Automatic learning rate，epoch=190

Download 
[efficientdet-b2 download link](https://pan.baidu.com/s/1-_G5wWQwCPeHaahXbfarBQ) 
code：hl3o

## Muchas gracias.

https://github.com/huawei-noah/ghostnet 
https://github.com/xiaolai-sqlai/mobilenetv3 
https://github.com/RangiLyu/nanodet 
https://github.com/ultralytics/yolov5 
https://github.com/implus/GFocal 
https://github.com/implus/GFocalV2

quarkdet

quarkdet 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

quarkdet

quarkdet 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README