Closed CarloLucibello closed 1 year ago
I am working on fixing the DenseNets, and porting their weights. That, along with some docs updates, should land in under a week and should be the last thing we handle in v0.8.
Since we can load the old version of DenseNet and the change should be non-breaking we can consider that non-blocking for the v0.8 release and be part of a patch release shortly after.
I think the change might be breaking 😬 When I compare our implementation with the one in the paper, we are doing something different. For example, compare https://github.com/pytorch/vision/blob/d2f7486ccaef461913cdb51990ff353addf6f064/torchvision/models/densenet.py#L96-L124 with https://github.com/FluxML/Metalhead.jl/blob/010d4bc72989e4392d2ce0b89e72b8d640927dd8/src/convnets/densenet.jl#L49-L53 and you realise that we are not combining the features at one higher level as it should be according to the paper.
ok, let's wait for the fix then
The issues for the milestone have been cleared. Anything else before tagging?