YOLOv3_TensorFlow

1. Introduction

This is my implementation of YOLOv3 in pure TensorFlow. It contains the full pipeline of training and evaluation on your own dataset. The key features of this repo are:

Efficient tf.data pipeline
Weights converter (converting pretrained darknet weights on COCO dataset to TensorFlow checkpoint.)
Extremely fast GPU non maximum supression.
Full training and evaluation pipeline.
Kmeans algorithm to select prior anchor boxes.

2. Requirements

Python version: 2 or 3

Packages:

tensorflow >= 1.8.0 (theoretically any version that supports tf.data is ok)
opencv-python
tqdm

3. Weights convertion

The pretrained darknet weights file can be downloaded here. Place this weights file under directory ./data/darknet_weights/ and then run:

python convert_weight.py

Then the converted TensorFlow checkpoint file will be saved to ./data/darknet_weights/ directory.

You can also download the converted TensorFlow checkpoint file by me via [Google Drive link] or [Github Release] and then place it to the same directory.

4. Running demos

There are some demo images and videos under the ./data/demo_data/. You can run the demo by:

Single image test demo:

python test_single_image.py ./data/demo_data/messi.jpg

Video test demo:

python video_test.py ./data/demo_data/video.mp4

Some results:

Compare the kite detection results with TensorFlow's offical API result here.

(The kite detection result is under input image resolution 1344x896)

5. Inference speed

How fast is the inference speed? With images scaled to 416*416:

Backbone	GPU	Time(ms)
Darknet-53 (paper)	Titan X	29
Darknet-53 (my impl.)	Titan XP	~23

why is it so fast? Check the ImageNet classification result comparision from the paper:

6. Model architecture

For better understanding of the model architecture, you can refer to the following picture. With great thanks to Levio for your excellent work!

7. Training

7.1 Data preparation

(1) annotation file

Generate train.txt/val.txt/test.txt files under ./data/my_data/ directory. One line for one image, in the format like image_index image_absolute_path box_1 box_2 ... box_n. Box_x format: label_index x_min y_min x_max y_max. (The origin of coordinates is at the left top corner, left top => (xmin, ymin), right bottom => (xmax, ymax).) image_index is the line index which starts from zero. label_index is in range [0, class_num - 1].

For example:

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

(2) class_names file:

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

For example:

bird
person
bike
...

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

(3) prior anchor file:

Using the kmeans algorithm to get the prior anchors:

python get_kmeans.py

Then you will get 9 anchors and the average IoU. Save the anchors to a txt file.

The COCO dataset anchors offered by YOLO's author is placed at ./data/yolo_anchors.txt, you can use that one too.

NOTE: The yolo anchors computed by the kmeans script is on the original image scale. You may need to resize the anchors to your target training image size before training and write them to the anchors txt file. Then you should not modify the anchors later.

7.2 Training

Using train.py. The hyper-parameters and the corresponding annotations can be found in args.py:

CUDA_VISIBLE_DEVICES=GPU_ID python train.py

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

8. Evaluation

Using eval.py to evaluate the validation or test dataset. The parameters are as following:

$ python eval.py -h
usage: eval.py [-h] [--eval_file EVAL_FILE] 
               [--restore_path RESTORE_PATH]
               [--anchor_path ANCHOR_PATH] 
               [--class_name_path CLASS_NAME_PATH]
               [--batch_size BATCH_SIZE]
               [--img_size [IMG_SIZE [IMG_SIZE ...]]]
               [--num_threads NUM_THREADS]
               [--prefetech_buffer PREFETECH_BUFFER]
               [--nms_threshold NMS_THRESHOLD]
               [--score_threshold SCORE_THRESHOLD] 
               [--nms_topk NMS_TOPK]

Check the eval.py for more details. You should set the parameters yourself.

You will get the loss, recall, precision, average precision and mAP metrics results.

9. Some tricks

Here are some training tricks in my experiment:

(1) Apply the two-stage training strategy:

First stage: Restore darknet53_body part weights from COCO checkpoints, train the yolov3_head with big learning rate like 1e-3 until the loss reaches to a low level.

Second stage: Restore the weights from the first stage, then train the whole model with small learning rate like 1e-4 or smaller. At this stage remember to restore the optimizer parameters if you use optimizers like adam.

(2) Quick train:

If you want to obtain acceptable results in a short time like in 10 minutes. You can use the coco names but substitute several with real class names in your dataset. In this way you restore the whole pretrained COCO model and get a 80 class classification model, but you only care about the class names from your dataset.

(3) I've included many useful training strategies in args.py:

Cosine decay of lr (SGDR)
Multi-scale training
Label smoothing
Mix up data augmentation
Focal loss

These are all good strategies but it does not mean they will definitely improve the performance. You should choose the appropriate strategies for your own task.

This paper from gluon-cv has proved that data augmentation is critical to YOLO v3, which is completely in consistent with my own experiments. Some data augmentation strategies that seems reasonable may lead to poor performance. For example, after introducing random color jittering, the mAP on my own dataset drops heavily. Thus I hope you pay extra attention to the data augmentation.

(4) Loss nan? Setting a bigger warm_up_epoch number and try several more times.