Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
YOLO for object detection tasks
| Date | Stars |
|---|---|
| 2026-07-24 | 430 |
| 2026-07-25 | 430 |
| 2026-07-28 | 429 |
| 2026-07-30 | 429 |
| 2026-08-06 | 429 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# [PYTORCH] YOLO (You Only Look Once)
## Introduction
Here is my pytorch implementation of the model described in the paper **YOLO9000: Better, Faster, Stronger** [paper](https://arxiv.org/abs/1612.08242).
<p align="center">
<img src="demo/video.gif"><br/>
<i>An example of my model's output.</i>
</p>
## How to use my code
With my code, you can:
* **Train your model from scratch**
* **Train your model with my trained model**
* **Evaluate test images with either my trained model or yours**
## Requirements:
* **python 3.6**
* **pytorch 0.4**
* **opencv (cv2)**
* **tensorboard**
* **tensorboardX** (This library could be skipped if you do not use SummaryWriter)
* **numpy**
## Datasets:
I used 4 different datases: VOC2007, VOC2012, COCO2014 and COCO2017. Statistics of datasets I used for experiments is shown below
| Dataset | Classes | #Train images/objects | #Validation images/objects |
|------------------------|:---------:|:-----------------------:|:----------------------------:|
| VOC2007 | 20 | 5011/12608 | 4952/- |
| VOC2012 | 20 | 5717/13609 | 5823/13841 |
| COCO2014 | 80 | 83k/- | 41k/- |
| COCO2017 | 80 | 118k/- | 5k/- |
Create a data folder under the repository,
```
cd {repo_root}
mkdir data
```
- **VOC**:
Download the voc images and annotations from [VOC2007](http://host.robots.ox.ac.uk/pascal/VOC/voc2007) or [VOC2012](http://host.robots.ox.ac.uk/pascal/VOC/voc2012). Make sure to put the files as the following structure:
```
VOCDevkit
├── VOC2007
│ ├── Annotations
│ ├── ImageSets
│ ├── JPEGImages
│ └── ...
└── VOC2012
├── Annotations
├── ImageSets
├── JPEGImages
└── ...
```
- **COCO**:
Download the coco images and annotations from [coco website](http://cocodataset.org/#download). Make sure to put the files as the following structure:
```
COCO
├── annotations
│ ├── instances_train2014.json
│ ├── instances_train2017.json
│ ├── instances_val2014.json
│ └── instances_val2017.json
│── images
│ ├── train2014
│ ├── train2017
│ ├── val2014
│ └── val2017
└── anno_pickle
├── COCO_train2014.pkl
├── COCO_val2014.pkl
├── COCO_train2017.pkl
└── COCO_val2017.pkl
```
## Setting:
* **Model structure**: In compared to the paper, I changed structure of top layers, to make it converge better. You could see the detail of my YoloNet in **src/yolo_net.py**.
* **Data augmentation**: I performed dataset augmentation, to make sure that you could re-trained my model with small dataset (~500 images). Techniques applied here includes HSV adjustment, crop, resize and flip with random probabilities
* **Loss**: The losses for object and non-objects are combined into a single loss in my implementation
* **Optimizer**: I used SGD optimizer and my learning rate schedule is as follows:
| Epoches | Learning rate |
|------------------------|:---------------:|
| 0-4 | 1e-5 |
| 5-79 | 1e-4 |
| 80-109 | 1e-5 |
| 110-end | 1e-6 |
* In my implementation, in every epoch, the model is saved only when its loss is the lowest one so far. You could also use early stopping, which could be triggered by specifying a positive integer value for parameter **es_patience**, to stop training process when validation loss has not been improved for **es_patience** epoches.
## Trained models
You could find all trained models I have trained in [YOLO trained models](https://drive.google.com/open?id=1Ee6FHQTGuJpNRYSa8DtHWzu4yWNyc7sp)
## Training
For each dataset, I provide 2 different pre-trained models, which I trained with corresresponding dataset:
- **whole_model_trained_yolo_xxx**: The wholeExcerpt of 6,717 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:ece900370973661a, topic:computer-vision, topic:object-detection, topic:yolo
matched fp:ece900370973661a, topic:deep-learning, topic:pytorch