This is implementation of YOLO v2 with TensorFlow.
Clone YOLO_v2 repository
$ git clone https://github.com/leeyoshinari/YOLO_v2.git
$ cd YOLO_v2
Download Pascal VOC2007 dataset, and put the dataset into data/Pascal_voc
.
If you download other dataset, you also need to modify file paths.
Download weights file yolo_weights for COCO, and put weight file into data/output
.
Or you can also download my training weights file YOLO_v2 for VOC.
Modify configuration into yolo/config.py
.
Training
$ python train_val.py
Test
$ python test_val.py
For more information to wiki.
Darknet-19 has 19 convolutional layers, it's faster than yolo_v2. If you use darknet-19, you need some modifications. It's easy to modify.
Please download Darknet-19 weights file for VOC from darknet-19.
To train the model on your own dataset, you should need to modify:
Put all the images into the Images
folder, put all the labels into the Labels
folder. Select a part of the image for training, write this part of the image filenames into train.txt
, the remaining part of the image filenames written in test.txt
. Then put the Images
, Labels
, train.txt
and test.txt
into data/dataset
. Put weight file in data/output
.
config.py:
modify the CLASSES.
train.py:
replacefrom pascal_voc import Pascal_voc
with from preprocess import Data_preprocess
, and replace pre_data = Pascal_voc()
with pre_data = Data_preprocess()
.