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 pipeline.
- Kmeans algorithm to select prior anchor boxes.
- Multi-GPU training with sync batch norm. (on working)
- tensorflow >= 1.8.0 (lower versions may work too)
3. Weights convertion
The pretrained darknet weights file can be downloaded here. Place this weights file under directory
./data/darknet_weights/ and then run:
Then the converted TensorFlow checkpoint file will be saved to
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
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:
|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!
7.1 Data preparation
(1) annotation file
train.txt/val.txt/test.txt files under
./data/my_data/ directory. One line for one image, in the format like
image_absolute_path box_1 box_2 ... box_n. Box_format:
label_index x_min y_min x_max y_max.(The origin of coordinates is at the left top corner.)
xxx/xxx/1.jpg 0 453 369 473 391 1 588 245 608 268 xxx/xxx/2.jpg 1 466 403 485 422 2 793 300 809 320 ...
NOTE: You should leave a blank line at the end of each txt file.
(2) class_names file:
data.names file under
./data/my_data/ directory. Each line represents a class name.
bird person bike ...
The COCO dataset class names file is placed at
(3) prior anchor file:
Using the kmeans algorithm to get the prior anchors:
Then you will get 9 anchors and the average IOU. Save the anchors to a txt file.
The COCO dataset anchors offered by YOLO v3 author is placed at
./data/yolo_anchors.txt, you can use that one too.
NOTE: The yolo anchors should be scaled to the rescaled new image size. Suppose your image size is [W, H], and the image will be rescale to 416*416 as input, for each generated anchor [anchor_w, anchor_h], you should apply the transformation anchor_w = anchor_w / W * 416, anchor_h = anchor_g / H * 416.
train.py. The parameters are as following:
$ python train.py -h usage: train.py [-h] [--train_file TRAIN_FILE] [--val_file VAL_FILE] [--restore_path RESTORE_PATH] [--save_dir SAVE_DIR] [--log_dir LOG_DIR] [--progress_log_path PROGRESS_LOG_PATH] [--anchor_path ANCHOR_PATH] [--class_name_path CLASS_NAME_PATH] [--batch_size BATCH_SIZE] [--img_size [IMG_SIZE [IMG_SIZE ...]]] [--total_epoches TOTAL_EPOCHES] [--train_evaluation_freq TRAIN_EVALUATION_FREQ] [--val_evaluation_freq VAL_EVALUATION_FREQ] [--save_freq SAVE_FREQ] [--num_threads NUM_THREADS] [--prefetech_buffer PREFETECH_BUFFER] [--optimizer_name OPTIMIZER_NAME] [--save_optimizer SAVE_OPTIMIZER] [--learning_rate_init LEARNING_RATE_INIT] [--lr_type LR_TYPE] [--lr_decay_freq LR_DECAY_FREQ] [--lr_decay_factor LR_DECAY_FACTOR] [--lr_lower_bound LR_LOWER_BOUND] [--restore_part [RESTORE_PART [RESTORE_PART ...]]] [--update_part [UPDATE_PART [UPDATE_PART ...]]] [--update_part [UPDATE_PART [UPDATE_PART ...]]] [--use_warm_up USE_WARM_UP] [--warm_up_lr WARM_UP_LR] [--warm_up_epoch WARM_UP_EPOCH]
train.py for more details. You should set the parameters yourself.
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, like less than 1.
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 good 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 the class names from your dataset.
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]
eval.py for more details. You should set the parameters yourself.
You will get the loss, recall and precision metrics results, like:
recall: 0.927, precision: 0.945 total_loss: 0.210, loss_xy: 0.010, loss_wh: 0.025, loss_conf: 0.125, loss_class: 0.050
9. Other tricks
There are many skills you can try during training:
(1) Data augmentation: You can implement your data augmentation like color jittering under
data_augmentation method in
(2) Mixed up and label smoothing like what Gluon-CV does.
(3) Normalizations like L2 norm.
(4) Mutil-scale training: You can change the input image scales (i.e. different input resolutions) periodically like the author does in the original paper.
I refer to many fantastic repos during the implementation: