Freeze backbone at epoch 0, unfreeze at epoch 1

[hktxt]

Hey, thanks for your implementation. I just don't understand this: # Freeze backbone at epoch 0, unfreeze at epoch 1 in train loop.

By the way, no close() were called, while using open(). I think switch it to with open() is more elegant. And, add a param for training from scratch, even though we can do this just comment some lines.

[glenn-jocher]

@hktxt this optionally freezes the darknet53 backbone for the first epoch so only non-backbone layers are trained.

with open() commands don't require close() commands.

[Pari-singh]

Hi @glenn-jocher ,

1. I want to understand, is that freezing of the final layer and then unfreezing are in case of (no) transfer learning, ie., if I do a custom data train and do not use the tag --transfer, given my data is new, it will freeze the last layers
2. I want to freeze the lowest layers i.e., the first 2, how would you suggest me to modify the train.py?

It would be really great to hear your comments! Thanks

[glenn-jocher]

@Pari-singh --transfer freezes layers, which I do not recommend. Simply train normally for best results from whatever starting point you wish using python3 train.py --weights start_from_here.pt

[Pari-singh]

@glenn-jocher which eventually means training it more on entire model, correct? I was wondering if there was a feature for gradual unfreezing till the layer we want

[glenn-jocher]

@Pari-singh no there's no feature like this. If you want the fastest results, simply start from pretrained yolov3-spp. python3 train.py --weights yolov3-spp.weights

If you want the best results, train on randomly initialized weights: python3 train.py --weights ''

[Pari-singh]

Yes, thanks for the insight @glenn-jocher

[glenn-jocher]

You're welcome, @Pari-singh! If you need further assistance, feel free to ask. Good luck with your training!