Yet-another-efficientdet-pytorch: Are your models in readme.md trained from scratch?

[tetsu-kikuchi]

Hello, @zylo117. Thank you for your work. I'd like to ask you a question.

Are the models you provide and list in readme.md, which achieve SOTA performance, trained from scratch by yourself ?
I mean, did you start training the models with only ImageNet-pretrained EfficientNet backbone, while other parts (such as BiFPN and headers) randomly initialized ?

I found you commented that the weights you provide are transferred and fine-tuned, not trained from scratch.
(https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch/issues/140#issuecomment-616344730)
I wonder what you mean by 'transferred'. Does this mean that you transferred already-trained EfficientDet models from some other repositories, and then fine-tuned them by your codes? In that case, from which repository did you transfer models ?

I am still a novice in this field, and apologize you in advance if I make some misunderstandings. Thanks.

[zylo117]

Transferred from the official effdet and then finetuned by a little bit.

[tetsu-kikuchi]

@zylo117
I got it. Thank you for your quick reply !

[tetsu-kikuchi]

@zylo117
I suppose it is good to write it in readme.md that you transferred from the official effdet and then finetuned.