Deeply learning 3D-aware node embeddings for RNA secondary structure graphs.
Learn unsupervised node embeddings using method from https://arxiv.org/abs/1905.12265
To train model on RNA graphs in ./directory:
First preprocess graphs (compute angles and one-hot features from the rna_classes features) by
python data_processing/process_graphs.py -i [directory] -o [preprocessed_dir]
Then learn embeddings using context prediction by running
python pretrain_context.py --train_dir [preprocessed_dir]
Default values for context prediction hyperparams are K=1, r1 = 1, r2=2. (K,r1,r2) can be changed by adding arguments
python train.py --train_dir [preprocessed_dir] --K ... --r1 ... --r2 ...
To compute embeddings for graphs in ./directory and save them to a new dir, run
python embeddings.py -i [directory] -o [write_directory]
An example of how to use a saved model to warm-start embeddings with the pretrained embeddings is given in tasks/train_mg.py