\[\newcommand{\ud}{\mathop{}\negthinspace\mathrm{d}} \newcommand{\pfrac}[2][x]{\frac{\partial #2}{\partial #1}}\]

Hyperparameter Tuning

In this part, we go through the basic steps for tuning the hyperparameters of Matilda. It is an example with specified dataset. The commands below would need to be run in settled conda environment, which could be referred to the Installation part.

Before Tuning the hyperparameter, the data should be appropriately prepared and loaded. The process could be referred to the corresponding parts in Implementing Matilda.

Main Functions for Matilda

The following functions are the core of Matilda. They have covered all of the function for Matilda. Their hyperparameters are controlling the distinct process or functions for Matilda. The detailed information of their hyperparameter could be viewed in the following two parts in this page.

Other classes or functions that help construct the main functions for matilda could be viewed in the sections of Others information. Generally, they are not directly/seperately used or called for other purposes.

Tunable hyperparameters

Tunable hyperparameter	Default value	Input type	Basic meaning
batch_size	64	int	Batch size for learning
epochs	30	int	Training epoches
fs_method	“IntegratedGradient”	string	Training epoches
hidden_rna	185	int	The number of neurons for RNA layer
hidden_adt	30	int	The number of neurons for ADT layer
hidden_atac	185	int	The number of neurons for ATAC layer
query	False	bool	If the input data is query of reference
lr	0.02	float	Learning rate for optimisation
seed	1	int	Global Random seed
simulation_ct	1	int	an index for which cell type to simulate, could be the real type label. “-1” means to simulate all types. Only be activated when simulation = True.
simulation_num	100	int	The number of cells to simulate for the specified cell type. Only be activated when simulation = True.
z_dim	100	int	Dimension of latent space

Functional hyperparameters

These hyperparameters are typically the switches of corresponding funtion/process. They are all Bool type. You could set as True for switching on all the functions at the same time in one completed process.

Functional hyperparameter	Default value	Controlling Function/process
augmentation	False	Data Augmentation
classification	False	Classification
dim_reduce	False	Dimension reduction
fs	False	Feature selection
simulation	False	Data simulation

Tuning strategies and methods

You could use the validation set for performance evaluation during tuning. If the size of dataset is larger, the classic Grid Search might take a long time to find out the optimized solution since it needs to be set manually. Bayesian optimisation could be a relatively efficient method without manual operation. If you have enough prior information for datasets or models’ expectation, it would be better. The metrics or benchmarks might need to be customised for multimodal sequencing data analysis for different process, such as the correlation visualisation for feature selection process or overall accuracy for classification…