yolo3-tensorflow

TensorFlow implementation of yolo v3 objects detection.
Based: full or tiny, and cnn or mobilenets(mobilenet_v1, mobilenet_v2);
We can get 6 combination, but 1 of them has a little parameters and performs badly. So, you should build these 5 combination as folloing:

• cnn + full
• cnn + tiny
• mobilenet_v1 + full
• mobilenet_v2 + full
• mobilenet_v2 + tiny

These 5 frameworks are provided in this repository.

Dependence

python3
tensorflow >= 1.12
opencv

Quick start

• cnn full yolo3
  1. Download official yolov3.weights and put it on model_data floder of project.
  2. Run the command python util/convert.py full to convert weights to TensorFlow checkpoint file, which will locate in logs/cnn_full/ and named cnn_full_model.data-00000-of-00001
  3. Run the command python yolo.py or python yolo.py -w logs/cnn_full/cnn_full_model and input the image path to detect.
  4. Detect example:
     
• cnn tiny yolo3
  1. Download official yolov3-tiny.weights and put it on model_data floder of project.
  2. Run the command python util/convert.py tiny to convert weights to TensorFlow checkpoint file, which will locate in logs/cnn_tiny/ and named cnn_tiny_model.data-00000-of-00001
  3. Run the command python yolo.py -w logs/cnn_tiny/cnn_tiny_model and input the image path to detect.
  4. Detect example:
     

Train

1. Prepare Dataset
   Before training, you should generate your own annotation file and class names file. One row for one image
   Row format: image_file_path box1 box2 ... boxN
   Box format: x_min,y_min,x_max,y_max,class_id (no space)
   For VOC dataset, try python util/voc_annotation.py
   For your own dataset, you should change the voc_annotation.py
   Here is an example:

path/to/img1.jpg 50,100,150,200,0 30,50,200,120,3
path/to/img2.jpg 120,300,250,600,2
...

2. Prepare yolo anchors
   run python util/kmeans.py to generate anchors. Note that, anchor number should be 9 if you wang to train full yolo, else it should be 6.

3. Start to train
   The train arguments can be seen in config/train_config.py.

usage: train.py [-h] [-n NET_TYPE] [-t TINY] [-b BATCH_SIZE] [-e EPOCH]
                [-lr LEARN_RATE] [-pt PRETRAIN_PATH]
                [--anchor_path ANCHOR_PATH] [--train_path TRAIN_PATH]
                [--classes_path CLASSES_PATH] [-d DEBUG]

optional arguments:
  -h, --help            show this help message and exit
  -n NET_TYPE, --net_type NET_TYPE
                        net type: cnn, mobilenetv1 mobilenetv2 or mobilenetv3
  -t TINY, --tiny TINY  whether tiny yolo or not
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        batch_size
  -e EPOCH, --epoch EPOCH
                        epoch
  -lr LEARN_RATE, --learn_rate LEARN_RATE
                        learn_rate
  -pt PRETRAIN_PATH, --pretrain_path PRETRAIN_PATH
                        pretrain path
  --anchor_path ANCHOR_PATH
                        anchor path
  --train_path TRAIN_PATH
                        train file path
  --classes_path CLASSES_PATH
                        classes path
  -d DEBUG, --debug DEBUG
                        whether print per item loss

The dafault framework is cnn + full. If you want to train others, you can pass the -n (cnn, mobilenetv1 or mobilenetv2) and -t (True or False) arguments.

4. To be simple
   I have write scripts in shell folder. Just run CUDA_VISIBLE_DEVICES='0' sh ./shell/train_cnn_full.sh or CUDA_VISIBLE_DEVICES='0' nohup stdbuf -oL sh ./shell/train_cnn_full.sh > logs/cnn_full.txt & in background and the log will be write in cnn_full.txt.
   You can also change some other arguments such as batch_size and epoch and so on.
   If you want to use pretrain, you should pass the pretrain path. I will provide the pretrain weights later.
5. NOTE The mobilenet is converged more slower than cnn, you should train more epoch.
6. Tensorboard You can use Tensorboard to watch the training trend.
   Run Tensorboard --logdir ./ --host 127.0.0.1

Predict

The prediction arguments can be seen in config/pred_config.py.

usage: yolo.py [-h] [-i IMAGE] [-v VIDEO] [-w WEIGHT_PATH] [--score SCORE]
               [--classes_path CLASSES_PATH]

optional arguments:
  -h, --help            show this help message and exit
  -i IMAGE, --image IMAGE
                        image path
  -v VIDEO, --video VIDEO
                        video path
  -w WEIGHT_PATH, --weight_path WEIGHT_PATH
                        weight path
  --score SCORE         score threshold
  --classes_path CLASSES_PATH
                        classes path

Note that, the weights filename should be like cnn_full_model.xxx, cnn_tiny_model.xxx, or others. the framework will be built by the word 'cnn' and 'full' or 'cnn' and 'tiny'.
You can predict an image or video.
For example:
python yolo.py -w weight_path
python yolo.py -i imgage_path -w weight_path
python yolo.py -v video_path -w weight_path