Open YangHai-1218 opened 4 years ago
question +1
I'd like to ask the same question.
zylo117 said that the weights he provides are transferred and finetuned, not trained from scratch, (https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch/issues/140#issuecomment-616344730) while he also said, somewhere in Issues, that he has experiences training a model from scratch by this code.
So I'd like to know whether zylo117 obtained mAP results by models he trained from scratch (possibly with ImageNet-pretrained EfficientNet backbone), which are comparable to mAP results in readme.md.
Thanks.
After looking into so many issues and replies, I really doubt that no one(including zylo117 himself) has ever reproduced his result reported on COCO dataset. It seems that zylo117 just transfer weights from tensorflow and finetune it a little bit. And that should be the reason that why he implements conv and pooling operator exactly the same with tensorflow because he needs to load weights from tf.
Hello, thanks for your work! But I was just wondering how did you get the results? Did you use the pretrained weights from https://github.com/google/automl/tree/master/efficientdet and convert them to pytorch version to load the weights and then use the train.py to fintune the weights? Or you just trained the models from scratch?