SocialGANetwork

This repo contains a project to generate social media content by leveraging Generative Adversarial Networks

Social Media Pages:

Instagram: https://www.instagram.com/socialganetwork/

Approach

Original Plan

Initially the plan was to have a full end to end process automated through use of instagrams API. Naively I did not expect their API to be so locked down and have had to abandon many elements of my originalplans, at least in the short term.

I had originally envisaged the finished project having a process such as below:

1. Programatically choose a hashtag
   • take biggest trending hashtag
   • allow community engagement - voting on hashtag by commenting on images on the account - take most requested
2. Scrape the top ~10k from that hashtag
3. Train Model
4. Generate new images from G - 9 to create an instagram "grid" for each run
5. Automatically upload with hashtag, details of the methodology and some metrics of the image and run

Current state

A number of the above elements have had to be de-scoped or de-priortised - at present:

• A hashtag is determined by the user (me)

• The batch of images is downloaded using a 3rd party app 4K Stogram

• Upload is performed manually using a mobile browser emulator (uploading from desktop is prohibited)

• Each run so far is a copy of the same notebook with input parameters modified

Issues

• Garbage In/Garbage Out

  • I believe a major blocker for generating better images is the heterogeneity of instagram content with a particular hashtag. Hashtags are often "spammed" and attached to images which they are not directly relevant to.
  • Examples:
    • The top images for the hashtag sunset contains many images that we would not recognise as "sunsets".
    • Discovering the site ingramer highlighted how banal instagram's user's use of hashtags often is.
  • Ways around this would be
  • Attempt to find choose suitable, homogenous hashtags, manually chosen to be quite niche
  • alternative sources of images (such as google images)
  • use some other technique to determine which images are unlikely to belond in your set of training images before training the GAN.
  • I guess part of the objective of "predicting" instagram content should naturally include the influence of a hashtag's "abuse" and this will lead to more abstract impressions of images from a topic/hashtag

Future Developments:

• Incorporate laplacian pyramid approach
• Filter images before training
• Train/output higher resolution images
• AI Upscaling of images
• Twitter / NLP Equivalent

References:

A large proportion of the current code is taken from the following site: DCGAN Tutorial — PyTorch Tutorials 1.6.0 documentation