SkyrimFaceFinder

Simple Neural Network Identifier for Skyrim Vanilla Pregens.

Business Understanding

Skyrim recently had a re-release, and has a slider-based character build system. Sometimes you want to make a character that looks like someone, i.e. you want to make yourself, a celebrity, a specific piece of art, etc. The goal of this project is to use tensorflow to create a Neural Network that will find the closest pregen character to a given photograph/picture.

Constraints In order to limit this project (and allow a quick 'win' for me), I will make the following limitations. If you want to borrow this code to go beyond that, please do.

• Human Only - This will be so that the author can find the closest pregen to their own face.
• Specific pose - This allows a regular neural net (with creepy eigenfaces) and avoids the need for a CNN, which I'd doing plenty of over in my Dog Classifier, which at the time of this writing is still a WIP.
• Reduced to three to save on memory/runtime.

Data Understanding

Since my laptop cannot run graphics-intensive games, and my only game system is an xbox 360, I will stick to the standard pregens. This is good, since it limits the data. I found a decent source of picture here. This of course means I will need to play around with these images.

These give me literally 1 image for each pregen. I will need to manipulate these to make multiple versions of each of these pictures, likely by scalling up or down by a pixel

Data Preparation

Again, stealing a bit from my Dog Classifier, I can build a little python code that will manipulate these images. Mostly just offsetting them by 1-2 pixels in each direction, which also resizes them to make them 100 x 100.

Modeling

The network above should do the job.

Evaluation

Showing how well it did on my face. I included the female faces, and since it's a simple 1-layer network, it seems to focus on beards (or the lack thereof). Since I'm clean-shaven, it labels me female often. I asked on facebook (always a great source of information) for some volunteer faces. See The Project Itself.

Full disclosure: I'm using my test set more like a validation set, but it's a one-off project for fun that I spent maybe 3 hours on, so. Also, good to see accuracies hit 100%... of course it's just the same picture offset slightly, but still.

Examples:

• No Filter:

  

• Male Filter:

  

• Female Filter:

  

Never expected Redguard (figured Imperial, given my heritage) but I totally see it in the facial structure.

Deployment

This was a one-off project, so I have no plans to deploy it. Feel free to clone it and run it yourself. Instructions below for people not proficient in python/github

If you want to run it yourself, all you need is:

• a github account (if you don't have one but are looking at this anyway)
• Anaconda python version 3
• Run python 3.5 (open command Prompt: "conda install python=3.5.0")
• Tensorflow In command prompt: 'pip install tensorflow'
• 'git clone' this repo (click 'clone or download', copy the https url, then type 'git clone (url)' in the folder you want this folder in)
• Place a 100 x 100 picture where your face is roughly in the same location as the rest in the 'Input' folder
• 'jupyter notebook' into SkyrimFaceFinder.ipynb and Kernal->Restart and Run All It will give you your face and your Skyrim face.