Implementation details for Efficient Det

[wuhy08]

Hi @mingxingtan ,
I am not sure if it is OK to ask here since EfficientDet model is not open source yet. I have been reading your new Efficient Det paper. Such a nice work! I have some questions about the implementation, though.

1. For Efficient Net, the max resolution resize is 32x (224->7), while in the EfficientDet paper, you mentioned that P_7 scales down to 128x. I wonder how do you deal with that? Do you add some additional levels on Efficient Net?

2. Did you add any lateral convolution layers from the feature maps from the backbone? Are output channels all the same in the BiFPN?

Thank you!