2
$\begingroup$

With yolov4, I am training an 80k images dataset that is used to classify different species of fish. Currently, I am using the following pre-trained weights: yolov4.conv.137 .

Now I was wondering if this is a backbone or weights trained with the COCO dataset? Would I benefit better from using these pre-trained weight files or just training from a blank slate?

$\endgroup$
0

1 Answer 1

0
$\begingroup$

My general intution says yes, your model might get benefit from pre-trained weights.

But, in transfer learning, two can things happen -

  1. Positive transfer - The pretrained model on big dataset performs very well when trained on a new dataset
  2. Negative transfer - The pretrained model on big dataset has a performance decline when trained on a new dataset

I will suggest a small experiment -

make a small subset of your fish dataset of ~ 10% ~ 8K images having same class proportion as the original and train your model for both cases, with transfer learning and without transfer learning. This small experiment will quickly determine which method is the most successful in your case (Fish dataset and YOLOv4).

More detailed discussion on the same is here - https://stats.stackexchange.com/questions/450801/the-negative-transfer-problem-in-machine-learning

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.