First week into GSoC with mlpack
Hey everyone, I intend to post a blog every week documenting what my plan was, what I did, and the things I learned or faced.
So with that let’s get started because I want to keep the blogs short but informational.
I proposed to implement two things which are the ResNet model builder which could build all the ResNet variants and MobileNet V1. At this point it’s not exactly the first week and I am into my second week so yeah a bit of a lie with the title but it looks catchy, doesn’t it :p
Till this point I have implemented All the building layers I need for the resnet models except a bottleneck block which is required for ResNet models above 34 excluding it, I also faced a segmentation error that took two days to debug because at this point I have written a good bit of code and to find such a thing to debug it was a pain, also I used gdb for the first time and it is easy to use, anyway back to the segmentation error so reading through the code and comparing it the existing models in mlpack I found out that there is a function named GetModel which should return the reference to the model which you can further use to train and do things but my function missed this little dude & so it wasn’t returning a reference anymore and when I called the predict function it ran into a segmentation error, but yeah returning a reference fixed that.
With that solved we uncovered that the padding layer in mlpack only accounted for a single filter, so to fix that we created a quick PR that fixed that issue so for now with all the things solved I continued my work on ResNet but for now, I am facing a dimensions mismatch while adding outputs in with the shortcut connections so need to debug what is causing that and hopefully with that I will be able to complete ResNet18 and ResNet34 which will allow me to work on the bottleNeck block for ResNet50, ResNet101, and ResNet152.
Thanks for reading this far.
References and resources if you want to follow up with my work
- Github
- Padding Layer PR
- And at last the ResNet PR
