cs-433 / ml_project2_reproducibility_challenge

EPFL Machine Learning Project 2

Reproducibility study of "Scaled-YOLOv4: Scaling Cross Stage Partial Network"

Second project for Machine Learning course (CS-443) at EPFL. This is the implementation of YOLOv4-CSP in "Scaled-YOLOv4: Scaling Cross Stage Partial Network" using PyTorch framework. The report explains the details of reproducibility study, and the python scripts are in the root folder.

For YOLOv4-tiny, please directly install and use Darknet.

• Getting started
  • Data
  • Folders and Files
• Running the code
• Contact us

Getting started

Data

The COCO 2017 dataset can be downloaded from COCO | Common objects in context, and the VOC 2007 dataset can be downloaded from The PASCAL object visual classes challenges.

Note that COCO and VOC labels should be converted to YOLO format first. For VOC 2007 labels, the user can use VOCtoYOLO.ipynb for conversion, and for COCO 2017 labels, the user can use COCOtoYOLO.py. Then, VOC image and label files should be sent to new folders named voc/images and voc/labels under the root folder respectively, and COCO image and label files should be sent to new folders named coco/images and coco/labels under the root folder respectively.

Folders and Files

train.py: the main script for YOLOv4-CSP model training,

test.py: the main script for model testing using the trained weights,

VOCtoYOLO.ipynb: converts VOC format label files to YOLO format,

COCOtoYOLO.py: converts COCO format label files to YOLO format,

data/coco.yaml: contains COCO class names and file path,

data/VOC07_YOLO.yaml: contains VOC class names and file path,

data/hyp.scratch.yaml: contains hyper-parameters for training,

utils: folder containing some util functions,

models/models.py: contains the Darknet-CSP backbone network algorithm,

models/yolov4-csp.cfg: contains the configuration information of Darknet-CSP,

report.pdf: project report explaining our reproducibility study in .pdf format written in Latex,

project2_description.pdf: assignment description given by EPFL,

requirements: environment used in the project.

Running the code

For training the VOC dataset, move to the root folder and execute:

python train.py --device 0 --batch-size 16 --data voc.yaml --cfg yolov4-csp.cfg --weights '' --name yolov4-csp

For testing,

python test.py --img 640 --conf 0.001 --iou 0.65 --batch 8 --device 0 --data voc.yaml --cfg models/yolov4-csp.cfg --weights weights/last.pt

Training the COCO dataset has similar procedure, but do not forget to change the data file path. The current code adopts CIoU loss function and Mish activation function. Detailed experiments for the algorithm is documented in report.pdf.

Contact us

Please don't hesitate to contact the authors about any questions about the project, data or algorithms in general:

• Tianyu Gu: [email protected]
• Xinyu Liu: [email protected]

@ 2022 Tianyu Gu