open raven
I'm trying to implement Yolo Architecture for Object localization task but it doesn't seem to learn anything i tried different learning rates , model sizes , batchsize and even different grid sizes but nothing seems to work after some time loss plateaus out it is not even overfitting to training data
here is the link of notebook in Kaggle : https://www.kaggle.com/code/parthvala/object-detection-testing
can someone please check it i don't know what is wrong with it
flint quartz
see it was decreasing
@open raven the thing is you are training whole network (assuming multi-million parameters) from scratch
its difficult/computational-expensive/time-consuming to train a whole NN from scratch/random-weights-initialization
to tackle this we use transfer learning, ie. you can use some pretrained YOLO model and finetune it on your dataset
Again im not sure if this is the only reason, there might be some other problems too, to be confirm calculate total trainable parameters in your model
open raven
Yeah the loss is decreasing but after some time plateaus out no matter what I do , if we change learning rate or batch size it may reach to 0.4 or 0.3 loss after so much training but even at that loss value it doesn't predict right labels for training data
And regarding time I trained for 3 or 4 hours but it was not even fitting training data
flint quartz
I go through your notebook and parameters arent an issue (120K params)
but also you have used custom loss function, torch dataset object, you have changed labling from yolo to coco then back to yolo....alot of places for error
One recommendation
I noticed you are only detecting only one object hence there is no need classification just do regression to obtain [x_center, y_center, w, h]
this will simplify the overall process