rmcavoy

@dvlshah That is actually incorrect. See my newer issue on convolutions where I found that the paper is including the backbone (minus the classifier layer) and bifpn+ heads in its...

From examining the original paper and this code, I believe D_class needs to be set by setting the optional parameter stacked_convs as an argument to the retinahead. To be blunt,...

Unfortunately I don't know of any faithful implementations currently. Signatrix also doesn't faithfully implement EfficientDet as they aren't using the all stride2 version of EfficientNet that the original EfficientDet paper...

I am guessing that if we get a good PyTorch implementation it will come from Facebook making one to put in Detection2.

If you were to print out all the feature levels/blocks in the EfficientNet, there are 7 of them and we are extracting the 3rd, 5th and 7th EfficientNet feature level/block....

Its not a lie per se. It is just exceedingly misleading to suggest you extracted feature 3-7 without adding in the extra equations that explain how you manufactured two of...

I went through and confirmed that if depthwise separable convolutions are used everywhere along with fixing the code so it is properly setting Dbifpn, Dclass/box, Wbifpn, Wclass/box etc to the...

Edit: Turns out Figure 3 is just misleading.

@sevakon Yes, from my tests of the number of parameters, every part of the detection network uses depthwise separable convolutions instead of normal convolutions. This makes sense with the original...

Also note that you need to make sure there is a batch norm after every depthwise separable layer to match the original paper.