simshineaicamera / magic

Magic tool for training deep learning models for specific object. Train your own model with less effort.

Magic

The project designed for simplifying training process as much as possible. We have used SOTA tracking and object detection models for automatic labeling.

Environments

1. Ubuntu 16.04/18.04
2. python > = 3.6
3. opencv >= 3.4
4. pyqt5
5. caffe-ssd (gpu version preferably)

Installation

1. Load the source of Magic project

git-lfs clone https://github.com/simshineaicamera/magic.git
cd magic

Note please use git-lfs, because yolov3 model is large, can't be downloaded without git-lfs

2. Install python package requirments:

python3.6 -m pip install -r requirments.txt

3. If you have already installed SIMCAM_SDK you can skip this step.
   Compile caffe-ssd on your system and set PYTHONPATH into your system path

Installation with Docker

1. The most easiest way to install the tool is using docker. Run following commands to upload and run the docker image

sudo docker run -it \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    -e DISPLAY=$DISPLAY \
    --privileged \
    simcam/magic:v1.0 bash
cd /home/Magic

How to use

1. Run the run.sh script

./run.sh

You will see GUI for auto labeling and training your own model

2. Let's say, you are going train your own cat detection model, press Upload Video button to upload a video , and please make sure that video contains desired object (in this case cat)

3. Choose object name
   Note, you can choose Other class as well, and give a label for the object for example mycat. In this option, the program use tracking method for automatic labeling

4. Press Generate Data button to generate data for training.

5. Data generating process takes 10-15 mins depends on your system CPU and GPU capability. After generating data, you will see message as below.

6. Press Train Model button to start training the model. Training speed depends on your system capability (CPU and GPU). You can see training process on the screen , epoch and accuracy.

You can stop training process early if you think accuracy of the model is enough. From the experiments, if model accuracy is higher than 0.86 value, you can test the model for deploy.

References

1. Caffe-SSD
2. MobileNet-SSD
3. SiamMask
4. MobileNet-YOLO
5. Yolov3