For ease of implementation, i have not implemented exactly the same as paper.
The things presented below are implemented differently from the paper.

• Backbone network(Used Xception instead of network mentioned in the paper.)

• Learning rate schedule(Used tf.keras.optimizers.schedules.ExponentialDecay)

• Data augmentations

• Hyper parameters

• And so on ...

docker build -t ${ImageName}:${ImageTag} .

• Example

docker run -d -it --gpus all --shm-size=${ShmSize} ${ImageName}:${ImageTag} /bin/bash

Pascal VOC Dataset with TFDS

• Training Set: VOC2007 trainval + VOC2012 trainval (Total: 16551)
• Validation Set: VOC2007 test (Total: 4952)

Trained with default values of this repo. (Total Epoch: 105)

• Download pb file: <Google Drive Link>

pb file is uploaded as tar.gz. So, you have to decompress this file like below.

tar -zxvf yolo_voc_448x448.tar.gz

If you want to inference with this pb file, infer to inference_tutorial.ipynb

• GPU(GeForce RTX 3090): About 8 FPS
• CPU(AMD Ryzen 5 5600X 6-Core Processor): About 4 FPS

Path: ./voc_scripts/train_voc.py

python train_voc.py

Options

Default option values are ./configs/configs.py.
If the options are given, the default config values are overridden.

• --epochs: Number of training epochs
• --init_lr: Initial learning rate
• --lr_decay_rate: Learning rate decay rate
• --lr_decay_steps: Learning rate decay steps
• --batch_size: Training batch size
• --val_step: Validation interval during training
• --tb_img_max_outputs : Number of visualized prediction images in tensorboard
• --train_ds_sample_ratio: Training dataset sampling ratio
• --val_ds_sample_ratio: Validation dataset sampling ratio

Path: ./voc_scripts/eval_voc.py

Evaluation pretrained model with VOC2007 Test Dataset

python eval_voc.py

Options

• --batch_size: Evaluation batch size (Default: batch_size of ./configs/configs.py.)
• --pb_dir: Save pb directory path (Default: ./ckpts/voc_ckpts/yolo_voc_448x448)

You Only Look Once: Unified, Real-Time Object Detection <arxiv link>

@misc{redmon2016look,
      title={You Only Look Once: Unified, Real-Time Object Detection}, 
      author={Joseph Redmon and Santosh Divvala and Ross Girshick and Ali Farhadi},
      year={2016},
      eprint={1506.02640},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}