Swipr - Swipr Script Service

Wednesday, December 12, 2018

Now that we have a trained model, we need to consider how to interact with it. The full tech of the script can be found here on GitHub but we’ll walk through it here too. The overview of this step is that we have a python script acting as a local server where the client side of the script receives input and returns output to to external consumers, and the server side loads the PyTorch model and runs the computation. …

Swipr - Fast.ai and CNNs

Tuesday, December 11, 2018

Finally we can consider creating the image classification model. The jupyter notebook that was used for this process can be found here on GitHub. DataLoader The first thing the creation of a PyTorch model requires is the creation of a DataLoader, which is a neat structure for handling the data we want to go over and what their labels should be. It loads x’s (inputs) and maps them to y’s (outputs). …

Swipr - Data Cleaning

Tuesday, December 11, 2018

Back in part 3, we said: Of course, pictures of beaches, food, and dogs are all common Instagram subjects, even on profiles of stereotypically self centered selfie-obsessed young twenties girls. We will come back to this point in detail later on, but for now, let's assume that each profile is pure and ideal for its category. Now it is time to come back to that. A bunch of Instagram photos, no matter from whose profile, are of course not pure and ideal for their category, and pictures of beaches, foods, and boyfriends abound. …

Swipr - Data Collection (Part 2 of 2)

Tuesday, December 11, 2018

Downloading from Instagram There are a number of solutions for scraping Instagram, and it’s mostly a pick-your-poison kind of affair. we picked the first well-supported google result for “Instagram scaper” and came up with Instalooter. It seemed to come with good documentation and sufficient automation facilities. Rate Limits We are told that prior to April 2018, downloading from Instagram was a much more lackadaisical affair. There was allegedly a generous rate limit for a given Instagram login token, upwards of 5,000 page requests per hour. This mythical time must have been the golden era for Instagram scraping. …

Swipr - Data Collection (Part 1 of 2)

Tuesday, December 11, 2018

Data Collection In previous posts we laid out what we expect our end result to look like, but underlying all of that is of course the ML model making our auto-tinder more than just a simple always-swipe-right bot. Like all deep learning projects, before we can consider training our model or even what our model architecture will be, we need to consider the problem of data. What, exactly do we need to collect? …

Swipr - Scope

Tuesday, December 11, 2018

From the outset we wanted to have, at the most basic level, the ability to use a CPU-only machine to judge whether or not we should swipe right on a given photograph. Additionally, it does us no good to just have a model that does this, it needs to be useful in some form beyond merely existing, meaning it has to exist as some kind of accessible application. Finally, it would be preferable to have this system be accessible from anywhere. Binding it to our local machine is useful insofar as this being an academic exercise is concerned, but it’d be a lot nicer to turn it on and off as we see fit, when we see fit. …

Swipr - Overview

Tuesday, November 27, 2018

The Goal Continuing the quest to learn modern Machine Learning, I thought it would be fun to create a Tinder auto-swiper that could potentially reject matches I don’t want and accept matches I do want. One of my friends had the brilliant idea of naming this tinderbot “Swiper”, like the fox from Dora the Explorer, famous for being repelled by the intrepid protagonists with the incantation of “Swiper, no swiping!” …

Machine Levine Mk. II

Saturday, November 3, 2018

Machine Levine Mk. II is the second of the Matt Bots, and the first of which can produce any kind of coherent output. A brief description of his capabilities are described at his homepage over at http://machinelevine.winetech.com/bots/ml2, but we’ll deal with some deeper ideas behind his creation and operation here. Data Collection Mk. II uses the same initial corpus that Mk. I uses. For more details on data collection, refer to Mk. I’s Data Collection section. …

Machine Levine Mk. I

Friday, November 2, 2018

This is a description of how the first of the Matt Bots came into existence. Few things in this document should be interpreted as the best or even a correct way to do anything - this bot is an academic exercise. That being said, here is how he was made. Data Collection Like the general post describes, the data used to train the Matt model was derived from Matt Levine’s at-the-time 571-count archive of Money Stuff articles. …

Machine Levine

Thursday, November 1, 2018

The Goal In order to better understand machine learning, I decided to see if I could get a neural network to write articles like Matt Levine. The original goal was to sort Spotify songs by male or female vocals, but I had been learning all of this stuff by following along with the fast.ai courses, and that specific use case was a bit afield of the coursework and forum discussions happening in the MOOC. Fast.ai, great as is it, has a focus on image data and NLP applications. As a compromise, I decided to learn by creating a bot that writes just like Matt Levine of Bloomberg. …