SnapCCESS 0.2.2

Ensemble deep learning of embeddings for clustering multimodal single-cell omics data

SnapCCESS

A python package to generate ensemble deep learning of embeddings for clustering multimodal single-cell omics data

Installation

Stable version

pip install snapccess  --index-url https://pypi.org/simple

https://pypi.org/project/snapccess/

Development version

pip install snapccess  --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple

https://test.pypi.org/project/snapccess/

The functions in this package are described below.

snapshotVAE

Description

To create the VAE model

Usage

model = snapshotVAE(num_features=[nfeatures_rna,nfeatures_pro], num_hidden_features=[hidden_rna2,hidden_pro2], z_dim=z_dim)

Arguments

• num_features: a list of number of features of each modality
• num_hidden_features: the number of hidden features we will used in training the model, in our paper, we use hidden_rna=185, and hidden_pro=30
• z_dim: dimension of the latent space, in our paper, we use z_dim=100

Output

A VAE model

---

train_model

Description

Training a VAE model with Snapshot learning rate or constant learning rate

Usage

model,histroy,embedding = train_model(model, train_dl, valid_dl, lr=lr, epochs=epochs,epochs_per_cycle=epochs_per_cycle, save_path=\"\",snapshot=True,embedding_number=1)

Arguments

• model: a vae model
• train_dl: training dataset
• valid_dl: validation dataset
• lr: initial learning rate
• epochs: total number of train cycles for snapshot ensemble vae
• epochs_per_cycle: the number of epochs per cycle
• save_path: the output file path of embeddings, by default leave it blank will not save any embeddings into the save_path, but the train_model will return the embeddings
• snapshot: a boolean value to indicate the model whether to use the snapshot ensemble method or the traditional VAE method (with constant learning rate)
• embeddings_number: a value to indicate the index of embeddings in the output filename when apply the traditional VAE

Output

This function will return the model, the loss of training and validation dataset (history) and a list of the latent space embeddings (for Snapshot ensemble method) or a single embedding for traditional VAE method.

---

get_encodings

Description

To get the embeddings from model after training.

Usage

embedding = get_encodings(model,valid_dl)

Arguments

• model: a VAE model
• valid_dl: the dataset that used as input to training the VAE model

Output

Embedding of the valid_dl dataset in the VAE model, to convert it to a matrix, try pd.DataFrame(embedding.cpu().numpy())

---

nvidia_info

Description

To monitor the memory usage of GPU

Usage

memory = nvidia_info(pid)['memory']

Arguments

• pid: the pid of running script

Output

This function will return the memory usage of the pid process.