API document¶

This section provides detailed API documentation for all public functions and classes in ResPAN.

Preprocessing¶

def data_preprocessing(adata):
   """Function used to preprocess our data with batch effect
   """

adata: The datasets with batch effect in AnnData form.

ResPAN¶

class Mish(nn.Module):
 """A class implementation for the state-of-the-art activation function, Mish.
 """

class discriminator(nn.Module):
 """The discrimimator structure of our AWGAN.
     Layer: N(2000)->1024->512->256->128->1
 """

class generator(nn.Module):
   """The generator structure of our AWGAN.
   Layer: 2000->1024->512->256->512->1024->2000
   with skip connection
   """

def run_respan(adata, batch_key='batch', order=None, epoch=300, batch=1024, lambda_1=10.,
            reduction='pca', subsample=3000, k1=None, k2=None, filtering=False,
            n_critic=10, seed=999, b1=0.9, b2=0.999, lr=0.0001, opt='AdamW'):
"""The main entry of our model
"""

adata: Data with batch effect.
batch_key: The index name of batch information in the given dataset.
order: The batch sequence for training or none.
epoch: The number of iteraiton steps.
batch: Input batch.
lambda_1: A hyperparameter used to control the weights of gradient penalty.
reduction: Method we used for dimension reduction, including None, ‘cca’, ‘pca’ and ‘kpca’.
subsample: The size of our data used to generate training dataset.
k1: The number of nearest neighbors we used across different batches.
k2: The number of nearest neighbors we used inner one batch.
filtering: A bool value used to determine whether we used top features selection or not.
n_critic: A step value used to determine the training times of generator.
seed: Random seed we used in our training process.
b1,b2: Hyperparameters used in AdamW optimizer.
lr: Learning rate.
opt: The name of our optimizer.