PyData Barcelona 2017

Deep learning workshop presented at PyData...

I presented a workshop on convolutional neural networks for audio processing at the PyData conference in Barcelona, along with my colleague Olga Slizovskaia. We introduced basic data processing routines with numpy and scipy, and then hyper-parameter visualization using TensorBoard. The introductory slides for the presentation can be found here. These routines are at the core of the source separation repository DeepConvSep and allowed me to train my model with hundreds of Gb of data.

During the three days of workshops and presentations, I learned a lot. I will try to summarize the things which I have seen.

I liked the tricks to speedup data reading and processing, given by Guillerm Borell. Numexpr is faster than numpy when performing different computations on big arrays. More speedup can be achieved with numba, which compiles python code into native machine instructions, either on the CPU or GPU. Think of it like cython on steroids. Don’t believe me? Try the python timing and profiling routines. Then, I was happy to find out that I am not the only one storing data as binary. I actually wrote my own routines for that, but if you look for something more fancy and also compatible with pandas, check h5py. Speaking of that, the following workshop by Francesc Alted was also very cool. Very similar to h5py, you can find feather which is a very fast format to store data. If you want something better for big data, check out fastparquet.

During the second and the third day, I have attended a few talks. Yufeng Guo is a developer advocate at Google and he gave a talk about wide and deep architectures with TensorFlow. This talk is largely based on a workshop they did last year. I am not working on natural language processing but this talk and the talk by Pascal van Kooten on creating boots were interesting.

Then, I attented the keynote by Gema Parreño who won the NASA asteroid challenge. I really liked the slides on the classification of neural networks, neural network zoo and Tensor playground which does some basic simulations for various architectures and problems.