Yet-Another-YOLOv4-Pytorch

!!! For the jupyter notebook please install pytorch-lightning version 0.7.6

This is implementation of YOLOv4 object detection neural network on pytorch. I'll try to implement all features of original paper.

• [x] Model
• [x] Pretrained weights
• [x] Custom classes
• [x] CIoU
• [x] YOLO dataset
• [x] Letterbox for validation
• [x] HSV transforms for train
• [x] MOSAIC for train
• [x] Dropblock layers for training. One in each PAN layer, but you can easily add it to each layer. (Thanks to Evgenii Zheltonozhskii for pytorch implementation)
• [x] LARS optimizer
• [x] Pytorch lightning
• [x] Self adversial training with fgsm
• [x] SAM attention block from official YOLOv4 paper
• [x] ECA attention block from https://arxiv.org/abs/1910.03151 with fastglobalavgpool from https://arxiv.org/pdf/2003.13630.pdf
• [x] Weight standartization from https://arxiv.org/abs/1903.10520 (Do not suggest to use with pretrained, could lead to an input explosion, used with track_running_stats, otherwise explosion)
• [x] Notebook with guide
• [x] IoU Aware from https://arxiv.org/abs/2007.12099
• [x] NMS in Depth implementation (not connected)
• [x] Matrix NMS algorithm from https://arxiv.org/abs/2007.12099 (not connected)
• [ ] Deformable convolutions from https://arxiv.org/abs/2007.12099
• [x] Coord convolutions from https://arxiv.org/abs/2007.12099
• [x] Self adversial training with vanila grad
• [x] Hard mish
• [ ] Easy mAP for your DL
• [x] ASFF from https://arxiv.org/abs/1911.09516
• [x] RAdam optimizer from https://arxiv.org/abs/1908.03265
• [x] Ranger optimizer (RAdam + LookAhead) from https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
• [x] Repulsion Loss from https://arxiv.org/abs/1711.07752v2
• [ ] Soft IoU Loss from https://arxiv.org/abs/1904.00853v3
  • [x] Learning IoU (as in IoU, just use IoU aware)
  • [ ] EM-Merger (TODO: use https://github.com/eg4000/SKU110K_CVPR19 to do postprocessing util)
• [ ] Elastic from https://arxiv.org/abs/1812.05262
• [x] BN microbatching
• [x] BCN from https://arxiv.org/pdf/1903.10520.pdf
• [ ] AdamP from https://arxiv.org/abs/2006.08217v1
• [x] Channel-wise feature fusion by me)

What you can already do

You can use video_demo.py to take a look at the original weights realtime OD detection. (Have 9 fps on my GTX1060 laptop!!!)

You can train your own model with mosaic augmentation for training. Guides how to do this are written below. Borders of images on some datasets are even hard to find.

You can make inference, guide bellow.

Initialize NN

#YOU CAN USE TORCH HUB
m = torch.hub.load("VCasecnikovs/Yet-Another-YOLOv4-Pytorch", "yolov4", pretrained=True)

import model
#If you change n_classes from the pretrained, there will be caught one error, don't panic it is ok

#FROM SAVED WEIGHTS
m = model.YOLOv4(n_classes=1, weights_path="weights/yolov4.pth")

#AUTOMATICALLY DOWNLOAD PRETRAINED
m = model.YOLOv4(n_classes=1, pretrained=True)

Download weights

You can use torch hub or you can download weights using from this link: https://drive.google.com/open?id=12AaR4fvIQPZ468vhm0ZYZSLgWac2HBnq

Initialize dataset

import dataset
d = dataset.ListDataset("train.txt", img_dir='images', labels_dir='labels', img_extensions=['.JPG'], train=True)
path, img, bboxes = d[0]

!!! You can use SplitDataset.ipynb to create train.txt and valid.txt

"train.txt" is file which consists with filepaths to image (images\primula\DSC02542.JPG)

img_dir - Folder with images labels_dir - Folder with txt files for annotation img_extensions - extensions if images

If you set train=False -> uses letterboxes If you set train=True -> HSV augmentations and mosaic

dataset has collate_function

# collate func example
y1 = d[0]
y2 = d[1]
paths_b, xb, yb = d.collate_fn((y1, y2))
# yb has 6 columns

Y's format

Is a tensor of size (B, 6), where B is amount of boxes in all batch images.

1. Index of img to which this anchor belongs (if 1, then it belongs to x[1])
2. BBox class
3. x center
4. y center
5. width
6. height

Forward with loss

y_hat, loss = m(xb, yb)

!!! y_hat is already resized anchors to image size bboxes

Forward without loss

y_hat,  _ = m(img_batch) #_ is (0, 0, 0)

Check if bboxes are correct

import utils
from PIL import Image
path, img, bboxes = d[0]
img_with_bboxes = utils.get_img_with_bboxes(img, bboxes[:, 2:]) #Returns numpy array
Image.fromarray(img_with_bboxes)

Get predicted bboxes

anchors, loss = m(xb.cuda(), yb.cuda())
confidence_threshold = 0.05
iou_threshold = 0.5
bboxes, labels = utils.get_bboxes_from_anchors(anchors, confidence_threshold, iou_threshold, coco_dict) #COCO dict is id->class dictionary (f.e. 0->person)
#For first img
arr = utils.get_img_with_bboxes(xb[0].cpu(), bboxes[0].cpu(), resize=False, labels=labels[0])
Image.fromarray(arr)

References

In case if you missed:
Paper Yolo v4: https://arxiv.org/abs/2004.10934
Original repo: https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

@article{yolov4,
  title={YOLOv4: YOLOv4: Optimal Speed and Accuracy of Object Detection},
  author={Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao},
  journal = {arXiv},
  year={2020}
}