Plotting

After training, you can plot the training history. This can help understand whether the model has converged or is overfitting:

model.plot_history()

Visualizing the representations (embeddings):

1. PCA of plain representations and the GMM (means and samples):

   model.plot_latent_space()
   

2. UMAP of representations using scanpy:

   data.obsm['latent'] = model.get_representation()
   data.obs['cluster'] = model.clustering().astype(str)
   
   sc.pp.neighbors(data, use_rep='latent')
   sc.tl.umap(data, min_dist=1.0)
   sc.pl.umap(data, color='observable')
   sc.pl.umap(data, color='cluster')
   

3. Covariate representations (2D):

   cov_rep = model.get_covariate_representation()
   
   import seaborn as sns
   sns.scatterplot(x=cov_rep[:, 0], y=cov_rep[:, 1], hue=data.obs[data.obs["train_val_test"]=="train"]["Site"].values)
   

See this notebook for the example plots:

human bonemarrow example notebook