GSOC2019-Introduction
Hello Everyone!
This year I have been fortunate enough to be selected for GSOC 2019. I will be working with the chainer organisation and my project is titled ‘Building Recurrent Neural Networks for ChainerX’.
I first started contributing to the ChainerX codebase in March. I started working on building various routines such as sigmoid, exponentiation. The members of the library are very helpful and supportive. They helped me understand the problem statement in depth and their requirements. This in-turn helped me write a good proposal.
Get Aniket Didolkar’s stories in your inbox
Join Medium for free to get updates from this writer.
The first week I spent mainly going through the codebase of chainer and understanding how recurrent neural networks were implemented in chainer. I got in touch with my mentor and we discussed in brief about the project. My mentor then gave me a specific routine to work on - n_step_lstm, which is basically an n-layered lstm cell. Have a look here for further details on lstms.
I started with the CPU implementation of this routine. I had to figure out a lot of stuff on my own. This was very different than the routines which I had built before. This was way more challenging. For me the most exciting part was to have the opportuinity to dive deep into the architecture of the LSTM which is one of the most prominent deep learning models used today. It took me about 10 days to finish the CPU implementation. I got to learn a lot from this.
After this, I had to start with the GPU implementation which involved usage of CUDNN, which is a GPU accelerated deep neural network library built by NVIDIA. I had some basic experience of working with CUDA in the past. I started by going through the documentation of CUDNN which was very daunting at first. I will cover my in-depth experience with CUDNN in the next blog post.
Stay Tuned!
