A complete TensorFlow implementation of YOLOv3/4(Tiny)

---

1. Introduction

This is a full implementation of YOLOv3, YOLOv4, YOLOv3-Tiny, YOLOv4-Tiny in pure TensorFlow.

Key features:

• K-means algorithm to select prior anchor boxes.
• Efficient tf.data multi-threading data pipeline
• Tensorlayer accelerated data augmentation, mixup, warm-up, label-smooth, focal-loss, multi-scale
• Weights conversion : .weights to ckpt, ckpt to .pb and demo usage of .pb
• Extremely fast GPU non-maximum-supression.

2. Requirements

• tensorflow1(with tf.data support)
• opencv-python
• tqdm
• tensorlayer1

3. Weights conversion

Weights can be downloaded here: link: https://pan.baidu.com/s/12Li_AZrZbGAs2642jjeOhw passwaord: nkd2

Place the weights file in the correspond directory in ./data/ and run the correspond conversion script, the converted ckpt file will be saved to the save directory.

4. Data preparation

1. annotation file

• Generate train.txt/val.txt files under ./data/my_data/ directory.

• One line for one image, in the format like index absolute_path width height box_1 box_2 ... box_n separate with a white space.

• Box_x format: label_index x_min y_min x_max y_max.

• index is the line index, starts from zero.

• label_index is the index of label in .names file, starts from zero.

0 xxx/xxx/a.jpg 1920 1080 0 453 369 473 391 1 588 245 608 268
1 xxx/xxx/b.jpg 1920 1080 1 466 403 485 422 2 793 300 809 320
...

2. class_name file:

• Generate the .names file under ./data/ directory. Each line represents a class name.

bird
person
bike
...

The COCO dataset class names file is placed at ./data/coco.names.

3. anchor file:

• Using the get_kmeans.py to get the prior anchors.

YOLO anchors is placed at ./data/yolo_anchors.txt, YOLO-Tiny anchors is placed at ./data/yolo_tiny_anchors.txt.

The yolo anchors computed by the k-means script is on the resized image scale. The default resize method is the letterbox resize, i.e., keep the original aspect ratio in the resized image.

4. VOC dataset:

   An example to parse VOC-like dataset is provided: write_list_voc.py, which can be used directly to parse VOCdevkit dataset.

5. Training

Run the correspond train.py file, continue training from a checkpoint is supported, you may refer to the tensorboard ouput in ./data/logs .

Check the args.py for more details. You may set the parameters yourself in your own specific task.

6. Freeze graph

The correspond freezegraph.py can generate .pb model, with or without weights.

The method to use .pb file can found in the correspond use_pb.py. You may refer to use_pb.py without suffix 'gpu' to for method to decode the yolo output purely on cpu.

---

Credits:

I referred to these fantastic repositories:

AlexeyAB/darknet

wizyoung/YOLOv3_TensorFlow