Home

Awesome

Trick or TReAT : Thematic Reinforcement for Artistic Typography

This is the official repository accompanying the ICCC 2019 paper "Trick or TReAT : Thematic Reinforcement for Artistic Typography", by Purva Tendulkar, Kalpesh Krishna, Ramprasaath R. Selvaraju & Devi Parikh. The demo can be found here.

Full text available at: https://arxiv.org/abs/1903.07820

Steps for Training

  1. Create your primary directoy and enter it.

  2. Create folders save, args, split, to save the trained model, arguments, and train-val split respectively.

  3. git clone https://github.com/purvaten/treat.git

  4. Train model

python treat/train.py \
--alpha 0.25 \
--data cliparts,letters \
--cliparts_dir treat/clipart_imgs \
--letters_dir treat/letter_imgs \
--datalimit 40250 \
--model alexnet

where treat/clipart_imgs and treat/letter_imgs should contain the clipart and letter images respectively for training. They each currently contain a single prototype image only.

NOTE : The letter images must be named in a specific format : TYPE + LETTER + ID, where TYPE = upper or lower, LETTER = a, b, ... z, and ID is a number (optional). For example, lowera1.png, upperb4.png, etc.

Results

WordThemeTReAT
stormweather, disasterstorm
mousecomputermouse
fishoceanfish
churchJesus, Godchurch

NOTE : The generated TReAT is independent of the sequence of provided themes.