| layout | title | date | tags | categories | |||
|---|---|---|---|---|---|---|---|
post |
reactionary |
2017-06-08 13:13:30 -0700 |
|
|
Create a machine-defined relationship between observed contant and human facial response using Youtube 'reaction videos'. Use this model to create new facial reactions for content that I create and compile it into a single looping video. There is an inherent performativity to the 'reaction video' format which exposes the social utility of the reaction which is to express a relative and categorizable ethical framework. This points to a tendency to integrate the exterior image into our understanding of our own relative righteousness.
Pix2Pix cGAN using Torch and Lua
Neural Enhance CNN using Tensorflow and Python
After several failed attempts to adapt Deep Multi-Scale Video Prediction to my needs. I turned to Pix2Pix.
Trained Pix2Pix on dataset, made new videos and used those to generate new reaction faces, these new videos were enlarged with Neural Enhance
2400 pngs taken from 40 different youtube reaction videos. 3 out of 4 of each videos frames were dropped and one second at 30 fps was taken out of the resulting reaction and source video.
Trained six different times with slight changes to the dataset between each training. Models were trained on a Google Cloud computer with 32 vCPUS and 128 GB of RAM. Training took roughly 3 days for each model.
I made a series of videos and then generated reactions to them using the model. The final results were composed into a single-channel looping video piece along with a stream-of-consciousness text that I wrote that addresses several of the themes.
Deep multi-scale video prediction beyond mean square error Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network Image-to-Image Translation with Conditional Adversarial Networks