rmcavoy

Feat_channels in retinahead should be W_bifpn

5

> Box/class prediction network – we fix their width to be always the same as BiFPN (i.e., Wpred = Wbifpn), but The feat_channels variable used in the retinahead should be...

Convolutions don't match paper

13

According to the original paper (see quotes below), the convolutions in this implementation appear to be incorrect since among other reasons normal convolutions are used. With regard to the bifpn,...

Incorrect feature map extraction

7

One of the large differences between this and the original implementation is this implementation follows the diagram in the original paper and extracts feature levels 3, 4, 5, 6 and...

The official Tensorflow implementation is now available.

9

https://github.com/google/automl/blob/master/efficientdet/efficientdet_arch.py

The below lines with torch.ones and torch.zeros in the focus loss are allocating arrays on one device and then transfering them to gpu using .cuda. This is inefficient especially in...