Anyone tried RetinaNet callbacks?

(Austin) #1

I know that this is just a dev notebook that hasn’t been covered or thoroughly tested, but curious if anyone else tried out this model and already solved the issues I’ve noticed with two callbacks.

The two callbacks I’m mainly interested in are to_fp16 and EarlyStoppingCallback.

Adding learn.to_fp16() causes an error in fitting, I think possibly stemming from the RetinaNetFocalLoss class. End of error is:

AttributeError: ‘int’ object has no attribute ‘dtype’

Could provide full traceback if anyone is interested in looking at it.

Also, I’ve tried to add a callback_fns of partial(EarlyStoppingCallback, monitor='valid_loss', min_delta=1e-3, patience=3), but when fitting I get the warning:

UserWarning: <class ‘fastai.callbacks.tracker.EarlyStoppingCallback’> conditioned on metric valid_loss which is not available. Available metrics are: train_loss, valid_loss, time

If anyone else has already started playing with the RetinaNet model in the pascal.ipynb notebook and has figured out either of these things, I’d be really interested in hearing!


Side note, if anyone has trained this on pascal 2007 please share the best mAP you got! :slight_smile:

0 Likes

(mohamed amin houidi) #2

the best mAP i got was around 17%. i tried many things but it still remains at 17%.

1 Like