Your solutions for this challenge must be able to function in a resource-limited setting i.e. it should run on a low-resource smartphone. As such, we are imposing the following restrictions on resources:
T4 GPU, maximum 9h training, maximum 3h inference
Model frameworks must be appropriate for use on edge devices (e.g. ONNX, TensorFlow Lite)
I think yolo11l and m fits in well for this description. It can be used on onnx and isn't that heavy around 50mb. If you break the restrictions that's when your model doesn't fit the description.
That's the next thing I'm going to do. Honestly I thought the classes are well balanced so I didn't do any sort of class balancing. But now I'm suffering to improve the score, so yeah why not 😅
What's your CV? Maybe there might be data leakage. Some models perform better than other models though. I know there is some kind of gap when I also post my submissions.
I haven't reached the 80 score yet but you can literally reach 78 easily by:
Perhaps try kfolds or reduce the test size if you are using yolo11.
Selecting a different model variant maybe (s, m, or even l)
Tune hyperparameters especially epochs (30-100), batch size.
For post processing the normal way, reduced confidence level (0 or 0.001 - might bring false positives though) , tuned IOU and image size.
For preprocessing, I didn't touch anything😅. I let yolo do that for me.
Does size L meet the requirements? I'm even doubtful about size M
Thanks. I will restart some experiments.
Size L is pretty quick to train.
Thats a very good question.
Your solutions for this challenge must be able to function in a resource-limited setting i.e. it should run on a low-resource smartphone. As such, we are imposing the following restrictions on resources: T4 GPU, maximum 9h training, maximum 3h inference Model frameworks must be appropriate for use on edge devices (e.g. ONNX, TensorFlow Lite)
I think yolo11l and m fits in well for this description. It can be used on onnx and isn't that heavy around 50mb. If you break the restrictions that's when your model doesn't fit the description.
ok !! Have you applied class balancing?
That's the next thing I'm going to do. Honestly I thought the classes are well balanced so I didn't do any sort of class balancing. But now I'm suffering to improve the score, so yeah why not 😅
Did it work out for you?
no, I haven’t balanced them yet either.
@nymfree are you still suffering to break the score?
still struggling. I see that you guys broke through.
Yes we really struggled to break the score. Did you try KFolds? Use a large number of folds and see whether it works.
tried kfolds and tuning some hyperparameters. can improve local score but not LB. Maybe my splits are not quite right.
What's your CV? Maybe there might be data leakage. Some models perform better than other models though. I know there is some kind of gap when I also post my submissions.
My splits were not quite right. I was doing something dumb. Re-running my experiments. Thanks @CodeJoe, I was about to give up on this competition.
Yolo can get you 0.800