A pytorch implementation of paper Scene Graph Disentanglement and Composition for Generalizable Complex Image Generation
conda env create -n disco python=3.8
conda activate disco
pip install -r requirement.txt
Download VisualGenome dataset here or wherever convenient for you. Then run script ./DisCo/prepare_data/filter_visual_genome.py and ./DisCo/prepare_data/construct_textual_graph.py to prepare data.
accelerate launch train_disco.py --use_ema --resolution=512 --batch_size=8 --gradient_accumulation_steps=2 --gradient_checkpointing --max_train_steps=50000 --learning_rate=1e-05 --lr_scheduler="linear" --checkpointing_steps 5000
Clone the DisCo repository from https://github.com/jmeyer24/DisCo.git into the scratch (?) folder Adapt it, prepare the data and then train the whole thing
## Get the repository (slightly modified)
git clone https://github.com/jmeyer24/DisCo.git
## get the datasets and models
bash DisCo/preparations.sh# Train the model with standard setup
accelerate launch ./DisCo/train_disco.py --use_ema --resolution=512 --batch_size=8 --gradient_accumulation_steps=2 --gradient_checkpointing --max_train_steps=50000 --learning_rate=1e-05 --lr_scheduler="linear" --checkpointing_steps 5000