Near the beginning of winter break, I had an idea to contribute an implementation to ONNX, Microsoft’s Open Source neural network accelerator. I had just been dropped from the Aerospace internship (due to accepting a position for Facebook’s summer software engineering internship (more on that later I swear)), and AI ideas were still fresh in my mind, so it was a good opportunity to keep my programming skills sharp.

I’m proud to announce that I’ve done it! You can check out my work here. The full README.md is a bit too wordy to include (and honestly is redundant to list again here), but I’ll include the description section so you have an idea of what I was working on.

Description

Fully Convolutional Networks (FCNs) are a neural network model for real-time class-wise image segmentation. As the name implies, every weight layer in the network is convolutional. The final layer has the same height/width as the input image, making FCNs a useful tool for doing dense pixel-wise predictions without a significant amount of postprocessing. Being fully convolutional also provides great flexibility in the resolutions this model can handle.

The specific model I contributed can detect 20 different classes, corresponding to the COCO 2012 class analouges of the PASCAL VOC classes. The models were sourced from PyTorch models pre-trained on the COCO train2017 dataset using this class subset.

Oh, also, something I should mention: today is a double post! Go check out the previous post for more code stuff I did recently.