search

erikguo / yolov3

Integrate SlimYoloV3 in YOLOv3 in PyTorch in Branch slimyolov3
GNU General Public License v3.0
26 stars 8 forks source link

issue about the precision #14

Closed ASONG0506 closed 4 years ago

ASONG0506 commented 4 years ago

Hi, erikguo,Thanks for your excellent job! I trained the yolov3-spp model on coco dataset, but got really bad performance, whose precision is really low but the recall is very high. One sample of test result is shown below. image I referenced some issue such as #2 , #295 , #188 and #71, and I modify the hypermeter "iou_t" but the mAP drop a lot, so did you trained the model in public dataset such as coco? How can you balance betweem the precision and recall? Thank you!

zihaozhang9 commented 4 years ago

Hello! How can I use the train.py code? I am now receiving the following error: AssertionError: No labels found. Recommend correcting image and label paths.

What are the input parameters of train.py?

ASONG0506 commented 4 years ago

obviously you need to check your dataset

---Original--- From: "zihaozhang9"<notifications@github.com> Date: Wed, Oct 23, 2019 22:14 PM To: "erikguo/yolov3"<yolov3@noreply.github.com>; Cc: "Asong"<1113461781@qq.com>;"Author"<author@noreply.github.com>; Subject: Re: [erikguo/yolov3] issue about the precision (#14)

Hello! How can I use the train.py code? I am now receiving the following error: AssertionError: No labels found. Recommend correcting image and label paths.

What are the input parameters of train.py?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

ASONG0506 commented 4 years ago

Finally I solved this problem by using "--evolve " in the training process to find out the best heper-parameters.

MuhammadAsadJaved commented 4 years ago

Hello! How can I use the train.py code? I am now receiving the following error: AssertionError: No labels found. Recommend correcting image and label paths.

What are the input parameters of train.py?

Have you resolved your problem? if no then show me your train.lists and val.lists

marmikp commented 4 years ago

I am training in yolov3.weights model with my own single class data. F1 score increased to 0.88 in the first 50 epochs but now it is constant for all the next epochs with e-5 learning rate.