A custom GPT model which was trained upon svg files.
Specifically the flat emoji variants from Microsoft's FluentUI repo.
These svn files only consist of "stand-alone" path elements which should make it simpler to train upon and sample from.
training and dataset
Both Tokenizer and Model were trained using aitextgen
The python file which was used for training, the .txt file dataset and a few generated samples can be found here
post processing and extracting .svg files from generated samples
# Extract from generated output and into a seperate .svg file all sequences which starts with <svg and ends with:
# A. </svg>
# B. If the sequence does not end with </svg> then find the last > in the sequence and append </svg> to it
generated samples
The generated samples below were also created with this script
