You Only Look Once v4 with TensorFlow and DALI#

This example presents a sample implementation of a YOLOv4 network, based on the following paper – Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao, YOLOv4: Optimal Speed and Accuracy of Object Detection.

The implementation is based on the code available in the darknet repository.

To run training from scratch on all available GPUs with DALI, run the following command:

python src/main.py train /coco/train2017 /coco/annotations/instances_train2017.json \
  -b 8 -e 6 -s 1000 -o output.h5 \
  --pipeline dali-gpu --multigpu --use_mosaic

To save checkpoints each epoch, add the following flag to the command:

--ckpt_dir /ckpt

Resume the training by providing an appropriate path to the -w flag:

-w /ckpt/epoch_3.h5

To perform evaluation every second epoch, add the following flags:

--eval_file_root /coco/val2017 \
--eval_annotations /coco/annotations/instances_val2017.json \
--eval_frequency 2 --eval_steps 500

To evaluate trained model, run the following command:

python src/main.py eval /coco/val2017 /coco/annotations/instances_val2017.json \
  -w output.h5 -b 1 -s 5000

To perform an inference and display the results on screen, run the following command:

python src/main.py infer image.png -w output.h5 -c coco-labels.txt

Requirements#

  • This example was tested using python 3.8 and tensorflow 2.4.1 and should work on later versions.

  • The model requires COCO 2017 dataset for training.

  • Besides TensorFlow, the following python packages are also required:

    matplotlib tensorflow-addons pycocotools

Usage#

Training#

usage: main.py train [-h] file_root annotations
  [--batch_size BATCH_SIZE] [--epochs EPOCHS] [--steps STEPS] [--output OUTPUT]
  [--start_weights START_WEIGHTS] [--log_dir LOG_DIR] [--ckpt_dir CKPT_DIR]
  [--pipeline PIPELINE] [--multigpu] [--use_mosaic] [--learning_rate LEARNING_RATE]
  [--eval_file_root EVAL_FILE_ROOT] [--eval_annotations EVAL_ANNOTATIONS]
  [--eval_steps EVAL_STEPS] [--eval_frequency EVAL_FREQUENCY]
  [--seed SEED]

Argument

Description

-h, –help

show this help message and exit

file_root

path to folder containing train coco2017 images

annotations

path to instances_train2017.json file

–batch_size BATCH_SIZE, -b BATCH_SIZE

number of images per training step, default = 8

–epochs EPOCHS, -e EPOCHS

number of training epochs, default = 5

–steps STEPS, -s STEPS

number of training steps per epoch, default = 1000

–output OUTPUT, -o OUTPUT

path to a .h5 output file for trained model, default = output.h5

–start_weights START_WEIGHTS, -w START_WEIGHTS

initial weights file in h5 or YOLO format

–log_dir LOG_DIR

path to a directory for TensorBoard logs

–ckpt_dir CKPT_DIR

path to a directory for checkpoint files

–pipeline PIPELINE

either dali_gpu, dali_cpu or numpy

–multigpu

if present, training is run using all available GPUs

–use_mosaic

if present, mosaic data augmentation is used

–learning_rate LEARNING_RATE

learning rate for training, default = 1e-3

–eval_file_root EVAL_FILE_ROOT

path to folder containing val coco2017 images

–eval_annotations EVAL_ANNOTATIONS

path to instances_val2017.json file

–eval_steps EVAL_STEPS

number of images per evaluation step, default = 5000

–eval_frequency EVAL_FREQUENCY

number of training epochs between each evaluation, default = 5

–seed SEED

seed for DALI and TensorFlow

Inference#

usage: main.py infer [-h] image [--weights WEIGHTS] [--classes CLASSES]
                     [--output OUTPUT]

Argument

Description

-h, –help

show this help message and exit

image

path to an image to perform inference on

–weights WEIGHTS, -w WEIGHTS

path to a trained weights file in h5 or YOLO format

–classes CLASSES, -c CLASSES

path to a coco-labels.txt file

–output OUTPUT, -o OUTPUT

path to an output image

Evaluation#

usage: main.py eval [-h] file_root annotations [--weights WEIGHTS]
                    [--batch_size BATCH_SIZE] [--steps STEPS]

Argument

Description

-h, –help

show this help message and exit

file_root

path to folder containing val coco2017 images

annotations

path to instances_val2017.json file

–weights WEIGHTS, -w WEIGHTS

path to a trained weights file in h5 or YOLO format

–batch_size BATCH_SIZE, -b BATCH_SIZE

number of images per evaluation step, default = 1

–steps STEPS, -s STEPS

number of evaluation steps, default = 1000