YoloV2-Trainable

Trainable Tiny-YoloV2 implementation on Julia by Knet framework

The program loads pre-trained weights to layers except the last layer. It initalizes last layer (detection layer) randomly. It freezes other layers and trains last layer only by adam optimizer.

first download pre-trained weights by:

$ wget https://pjreddie.com/media/files/yolov2-tiny-voc.weights

If you want to train the model and see accuracy on Voc Dataset 2007, download the dataset by:

$ wget https://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar
  tar xf VOCtrainval_06-Nov-2007.tar

The repo has trained model which is trained.jld2. If you want to see accuracy for this model run:

$ julia YoloTrain.jl accuracy

If you want to display the image by using trained.jld2. Run:

$ julia YoloTrain.jl loadDisplay

To train model, run:

$ julia YoloTrain.jl train --batch_size 32 --epochs 20

Batch_size = 32 and 20 epochs is suggested. When it is trained, it saves trained model as trained1.jld2. If you want to use this model for accuracy and display, simply use the argument:

$ --choose true

Example Input and Output

Here is an example of input and output:

INPUT:

OUTPUT:

References

https://fairyonice.github.io/Part_4_Object_Detection_with_Yolo_using_VOC_2012_data_loss.html https://github.com/experiencor/keras-yolo2

To-Do List

-It will become a more generalized version to train different datasets.