seaandsailor

Model checkpointing using meta-graphs in TensorFlow

jfsantos — Thu, 01 Dec 2016 12:00:00 -0500

Deep learning models usually take a while to train, even on GPUs, so being able to checkpoint intermediate stages in training is really important. Frameworks like Keras usually offer functionality to store learned variables like weights and biases, but for resuming training for a checkpoint, one might also need to …

A really brief introduction to audio signal processing in Julia

jfsantos — Mon, 03 Nov 2014 23:00:00 -0500

Introduction

If you are into scientific computing, you probably have already heard about Julia, the magical language that aims to be (almost) as fast as C and as easy as MATLAB and Python to write. I have been playing with Julia for more than a year now, and I really …

Configuring MATLAB to work with XCode 6 on Yosemite

jfsantos — Tue, 28 Oct 2014 17:00:00 -0400

After I upgraded my Mac to Yosemite (10.10), I had a lot of trouble to get MATLAB 2013a working properly. First, I started getting a NullPointerException from Java when trying to start it. Turns out this was a problem with some Java classes used for drawing MATLAB's GUI and …

Using stochastic neurons for conditional computation

jfsantos — Sun, 27 Apr 2014 16:00:00 -0400

As stated in my previous post, I had some issues when trying to train my network while using sparse-coded speech frames as input/output. The network was getting stuck at the same point after processing a single batch, and processing additional batches did anything training-wise (performance, weights, etc., were all …

Using an auditory-inspired representation for speech

jfsantos — Mon, 31 Mar 2014 14:00:00 -0400

I previously described an approach to representing speech signals by decomposing them to an arbitrary dictionary (using a sparse coding algorithm such as Orthogonal Matching Pursuit). In that post, I showed that learning a representation from the data by using a dictionary learning method could be useful. However, there were …

Experiments with a 2 layer MLP incorporating phone information

jfsantos — Wed, 19 Mar 2014 21:00:00 -0400

During the spring break I decided to run some experiments with MLPs as generative models, using both acoustic samples and phone codes as input. The experiment's objective is two-fold: I wanted to investigate if using information from surrounding frames improves the models (when compared to what our colleagues have found …

Dictionary Learning and Sparse Coding for Speech Signals

jfsantos — Tue, 25 Feb 2014 17:00:00 -0500

Sparse signal approximations are the basis for a variety of signal processing techniques. Such approximations are usually employed with the objective of having a signal representation that is more meaningful, malleable, and robust to noise than the ones obtained by standard transform methods [Sturm2009]. The so-called dictionary based methods (DBM …

Speech signal representations

jfsantos — Sat, 01 Feb 2014 18:00:00 -0500

One of the objectives of our project is to learn a useful representation from speech signals that can be used to synthesize new (arbitrary) sentences. There are many different ways of representing speech signals; those representations are usually tailored to specific applications. In speech recognition, for example, we want to …

Personal research journal: deep learning for speech synthesis

jfsantos — Sat, 25 Jan 2014 15:00:00 -0500

This is the introduction to a series of reports on my experiments on deep learning methods for speech synthesis. These experiments are part of my coursework for Dr. Yoshua Bengio's Representation Learning course at Université de Montréal. All the related code is going to be posted at a GitHub repository …

My Emacs configuration for MATLAB, LaTeX, and Python

jfsantos — Sun, 28 Jul 2013 12:00:00 -0400

Emacs is a main part of my daily life. I write code and papers using it, so you could say that anything that my research will leave (hopefully) for the next generations was probably typed using it. Not anybody loves Emacs, and I will not try to convince you to …

WinPython: a portable scientific Python distribution for Windows

jfsantos — Tue, 25 Jun 2013 21:00:00 -0400

Recently I developed a piece of software for administering speech intelligibility listening tests and discovered a small issue: I am not an administrator of the computer that I am going to use for the tests, so I cannot install any software. Since the computer is running Windows, there is no …

Using Vagrant to make reproducible development and test environments

jfsantos — Sun, 19 May 2013 16:00:00 -0400

Working with research-grade software is hard sometimes. We all have heard before (more than once, probably) the "it worked perfectly on my computer" story. The problem is that setting up a development environment takes a lot of time, and usually you'll end up forgetting which libraries and pieces of software …