/design_contest_yolo

YOLOv3-tiny PyTorch Implementation and some tools

Primary LanguagePython

FPGA デザインコンテスト 物体認識

  • アルゴリズム : YOLOv3-tiny

リポジトリのクローン (任意のディレクトリの下で)

$ git clone https://github.com/yarakigit/design_contest_yolo.git

Python仮想環境の有効化 (DL-Boxで、ユーザー : dl-box)

$ source ~/venv_design_contest/.venv/bin/activate

train

  • 学習結果は/output下
  • 0から学習させるとき
$ make train
  • 学習済みの重みを初期値として学習させるとき
$ make train_continue

detect

  • 推論結果はoutput.jpgとして表示
make detect

sample_output

詳しくはMakefileを参照してください。

YOLOv3-tiny PyTorch implementation

Features

  • Train network
  • Test network
  • Inference on an image
  • Inference on a video
  • Inference on a camera movie

Usage

Train

python train.py [Arguments]

[Arguments (all are optional)]

  • weights : pretrained weights used when transfer learning or fine tuning (default : None)

  • model : specify 'sep' if you use separable model (default : None)

  • data_root : path to dataset (train.txt must be placed, default : $HOME/datasets/COCO/2014)

  • num_classes : num of classification

  • class_names : path to file of class name (default : namefiles/coco.names)

  • output_model : file name of trained model (default : yolo-tiny.pt)

  • batchsize : batch size

  • lr : initial value of learning rate

  • epochs : training epochs

  • decay : weights decay

  • trans : transfer learning switch (default : False, set to True)

  • finetune : fine tuning switch (default : False, set to True)

  • novalid : don't do validation after every epoch (default : False, set to True)

  • nosave : don't save training result (default : False, set to True)

Run train.py like above. This training program saves the trained model into 'results/{date of end of training}/'. That directory includes model (*.pt), parameter list that have used in training (train_parameters.txt), learning rate transition (lr.png), mAP transition (mAP.png), and loss transition (loss.png).

Test

python test.py [Arguments]

[Arguments (all are optional)]

  • weights : trained model (default : weights/yolov3-tiny.pt)
  • model : specify 'sep' if you use separable model (default : None)
  • data_root : path to dataset (test.txt must be placed, default : $HOME/datasets/COCO/2014)
  • num_classes : num of classification
  • class_names : path to file of class name (default : namefiles/coco.names)
  • quant : quantization switch (default : False, set to True)
  • nogpu : specify if you don't want to use GPU (default : False, set to True)

Inference

python detect.py [Arguments]

[Arguments (all are optional)]

  • weights : trained model (default : weights/yolov3-tiny.pt)

  • model : specify 'sep' if you use separable model (default : None)

  • image : path to an image for inference (default : images/car.jpg)

  • output_image : output file name of inference result (default : output.jpg)

  • num_classes : num of classification

  • class_names : path to file of class name (default : namefiles/car.names)

  • quant : quantization switch (default : False, set to True)

  • nogpu : specify if you don't want to use GPU (default : False, set to True)

  • conf_thres : confidence threshold for NMS

  • nms_thres : nms threshold for NMS (IoU threshold)

Note

  • The directory 'tools/' is under development.
  • Some codes are based on eriklindernoren/PyTorch-YOLOv3