royxue / darknet

Convolutional Neural Networks

#Darknet# Darknet is an open source neural network framework written in C and CUDA. It is fast, easy to install, and supports CPU and GPU computation.

For more information see the Darknet project website.

For questions or issues please use the Google Group.

---

This Repo is used training one specific class detection project.

Training Steps:

1. Prepare GPU machine(highly recommend): Using AWS g2/2x.large with some community AMI with Cuda, CuDNN installed, or you need to have them installed by yourself.

2. Change the Makefile, set GPU=1 and CUDNN=1.

3. Prepare your dataset, manualy label the data by BBox Label Tools, and convert to the format darknet accept. x, y, width, height are all relative values. <object-class> <x> <y> <width> <height>

4. Create labels names, like data/voc.names, list all your classes names.

5. Create data config, like cfg/voc.data

   classes= {NUM_CLASSES}
   train  = {DATA_PATH}/train.txt
   valid  = {DATA_PATH}/valid.txt
   names = data/xxx.names
   backup = {DATA_PATH}/backup/
   

   train.txt and valid.txt are just 2 lists of the image paths, darknet will automatically find labels in the image folder or it's parents folder. Suggest to have 2 folders /images and /labels.

6. Create your .cfg file, you can modify the existing ones, or create some brand new one. If you choose to modify existing ones, there are serval values you need to change:

   classes={NUM_CLASSES} filters=num*{NUM_CLASSES+coords+1} for the last convolutional layers.

7. (Optional) Download some pre-trained darknet weights.

8. Run the training by ./darknet detector train cfg/xxx.data cfg/xxx.cfg xxx.weights

9. Run the test, first you need to change the line 404 in src/darknet.c the data file to your training one, and then make and run detector with your test images.

10. If you want to use the model on android devices, please see darkflow for how to generate darknet protobuf.