Finetuning about efficientdet-pytorch

[rwightman]

You can use the training script here for custom datasets so long as you have your custom dataset in a coco compatible annotation format (train / val json files).

To finetune, during model creation, adding a one liner that relplaces the HeadNet would be the easiest solution.

model.class_net = HeadNet(config, num_outputs=your_num_classes, norm_kwargs=dict(eps=.001, momentum=.01))

You could probably get a bit more granular and just replace the final layer of the head net
model.class_net.predict = ... but that requires some more code.

In both cases you need to hack together a weight init that only inits the modified head weights properly, especially the predict weights. It's best to do this by modifying the create factory, or creating your own... at line 28 after pretrained weights are loaded, not where the FIXME is.

https://github.com/rwightman/efficientdet-pytorch/blob/master/effdet/factory.py#L28

[rwightman]

Also, default training hparams won't necessarily work great for finetuning. I'd start with --opt adam --sched plateau and lr between 1e-3 to 1e-4.

[Naivepro1990]

Thanks,

It will be good idea to include the finetuning as an option in the code, by default it should be false.

And instead of hardcoding the prediction classes, it should be a parameter like in the original code.

[rwightman]

Yup, it's definitely on the todo list, likely next thing to add here with result of my resnet backbone experiements, but have a number of other project threads to wrap up first :)

…

On Mon, Jul 6, 2020 at 5:48 PM Naivepro1990 ***@***.***> wrote: Thanks, It will be good idea to include the finetuning as an option in the code, by default it should be false. And instead of hardcoding the prediction classes should be a parameter — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#49 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABLQICAVX524RSL7VXWHW3LR2JWGHANCNFSM4ORGIMGQ> .

[Naivepro1990]

@rwightman It seems the above solution doesn;t quite work for finetuning as the location you specified HeadNet is not defined. I did tried to import but still it gives error.