search

ultralytics / yolov5

YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite
https://docs.ultralytics.com
GNU Affero General Public License v3.0
50.61k stars 16.32k forks source link

coco128 #3188

Closed www516717402 closed 3 years ago

www516717402 commented 3 years ago

❔Question

  1. I get poor result when I train coco128 without pretrained model and any data-augment. 128 images will fitness easy by yolov5s model.class loss is difficult to fit. The result as follows. Many object will recognize mistake to person class.

image

sezer-muhammed commented 3 years ago

Hi,

If there is no mistake about classes or training, the model can detect objects in training images. But for external images these results are normal. 128 images is very low for any yolov5 model.

If this image in already in training dataset, you may increase tranining time.

Or, try to change CUDA version. When I use 1660Ti I cannot train my model with CUDA 11.1. When I use RTX3070 I cannot train my model with CUDA 10.2

I dont know why but if your mAP value does not increase, you may try change your cuda version.

www516717402 commented 3 years ago

@sezer-muhammed Thank you for reply.

  1. I will check dataset again and increase trainning time by your advice.
  2. Whether raise error when cuda not match?
sezer-muhammed commented 3 years ago

@sezer-muhammed Thank you for reply.

  1. I will check dataset again and increase trainning time by your advice.
  2. Whether raise error when cuda not match?

You are welcome, Yes. If there is a CUDA problem in every epochs your mAP valuse will decrease.

www516717402 commented 3 years ago

Great help. I will try. Thank you again.