search

WongKinYiu / yolov7

Implementation of paper - YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors
GNU General Public License v3.0
13.09k stars 4.13k forks source link

Paper claims yolov7-tiny is similar to yolov4-tiny-3l #583

Open holger-prause opened 1 year ago

holger-prause commented 1 year ago

I trained a custom model with yolov4-tiny-3l and yolov7-tiny - both using the same hyper parameter(flip, disabled mixup, disabled ota) and the same dataset. Also both were trained by starting from pretrained weights. There is a huuuge difference in map - and this is before converting it to onnx or doing any reparametrization. The reason i am trying yolov7 is that yolov4 has really problems with large objects in my dataset(even when using custom anchors) and also yolov7 has a better technology stack(based on pytorch) from my point of view.

I dont really expect an answer but please still consider giving one. Am i missing something obvious? Maybe i should try training from random weights? If i am correct yolov7 uses coco pretrained weight while darknet uses imagenet?

mnurilmi commented 1 year ago

hi, is this problem solved? I only found the config for yolo-tiny, while in the paper there was yolo-tiny silu. Do you know where the config is?

mnurilmi commented 1 year ago

Which is better performance between for v4-tiny-3l and v7-tiny? based on your experiments?