Last year I was working on implementing Faster R-CNN from scratch using the original paper. Pleased to say that I got it working and spent some time this month porting it to TensorFlow as well and polishing things up a bit. I’ve documented some of my embarrassing setbacks and what I learned from them. Hopefully others will find this useful, even if working on a completely different model.
It works quite well, is easy to set up, and I’d like to think it is pretty clean and readable. If nothing else, it’s a fun comparison between PyTorch and TensorFlow.