Code base corresponding to extended abstract accepted at the Machine Learning for Health (ML4H) at Neurips 2019 workshop.
Note: Data can be provided on request or obtained directly from ppmi-info.org
python train_ppmi.py --epoch 200 --batch_size 16 --experiment "?" --feature_size 441 --num_heads 4 --embedding_size 64 --block_shape 16 --meth --spect --num_datasets 2 --log_interval 10 --save --runs 3 --classification --learning_rate 0.00003 --dropout_keep_prob 1.0 --cuda --early_stop_epochs 20 --hidden_dim 256 --augment --aug_frac 0.1
parameter | description |
---|---|
epoch | number of training epochs |
batch_size | mini-batch size |
experiment | description of experiment |
feature_size | number of encoder output features |
num_head | number of heads for multi-head attention mechanism |
embedding_size | embedding dimension of encoder output |
block_shape | compressed embedding dimension within MHCA mechanism |
meth | use methylation data |
spect | use SPECT data |
num_datasets | how many unique datasets/modes are being used |
log_interval | how ofter to print progress to screen |
save | save the model |
runs | number of independent runs |
classification | true = classification, false = regression (currently not fully supported) |
learning_rate | learning rate |
dropout_keep_prob | dropout used within MHCA model |
cuda | execute model on GPU |
early_stop_epochs | how many epochs to wait until termenanting training |
hidden_dim | dimensionality of final hidden space |
augment | perform data augmentation |
aug_frac | fraction of data augmentation to apply |