Configuring

Below is a sample configuration file for the train.py script.

# Logging and Paths
wandb_proj: NoSigmoidTest # Weights and Biases project to log results to.
wandb_save: True # Whether or not to log to Weights and Biases.
log_file: ./logs/scratch_testing.log # Location of log file
model_save_dir: ./best_models # Location to save best models
data_cache_dir: /data/cb/samsl/Adapting_PLM_DTI/dataset # Location of downloaded data (use `conplex_dti download`)

# Misc
device: 0 # CUDA device to use for training
replicate: 0 # Random seed
verbosity: 3 # Verbosity level for logging

# Task and Dataset
task: davis # Benchmark task - one of "davis", "bindingdb", "biosnap", "biosnap_prot", "biosnap_mol", "dti_dg"
contrastive_split: within # Train/test split for contrastive learning is between target classes or within target classes

# Model and Featurizers
drug_featurizer: MorganFeaturizer # Featurizer for small molecule drug SMILES (see `conplex_dti.featurizer` documentation)
target_featurizer: ProtBertFeaturizer # Featurizer for protein sequences (see `conplex_dti.featurizer` documentation)
model_architecture: SimpleCoembeddingNoSigmoid # Model architecture (see `conplex_dti.models` documentation)
latent_dimension: 1024 # Dimension of shared co-embedding space
latent_distance: "Cosine" # Distance metric to use in learned co-embedding space

# Training
epochs: 50 # Number of epochs to train for

## Batching
batch_size: 32 # Size of batch for binary data set
contrastive_batch_size: 256 # Size of batch for contrastive data set
shuffle: True # Whether to shuffle training data before batching
num_workers: 0 # Number of workers for PyTorch DataLoader
every_n_val: 1 # How often to run validation during training (epochs)

## Learning Rate
lr: 1e-4 # Learning rate for binary training
lr_t0: 10 # With annealing, reset learning rate to initial value after this many epochs for binary traniing

## Contrastive
contrastive: True # Whether to use contrastive learning
clr: 1e-5 # Learning rate for contrastive training
clr_t0: 10 # With annealing, reset learning rate to initial value after this many epochs for contrastive training

## Margin
margin_fn: 'tanh_decay' # Margin annealing function to use for contrastive triplet distance loss
margin_max: 0.25 # Maximum margin value
margin_t0: 10 # With annealing, reset margin to initial value after this many epochs for contrastive training