Fix bug of importing __version__

Major updates:

• Add support for mean_vals, min_vals, and max_vals for mean and range constraints (depending on distributions).
• Fixing issues related to batch normalization (dataset repeated/shuffled for training and inference).
• Examples tested on both single-cell genomics and proteomics data.
• Rich docstrings added to all updated methods.

• Refactor initialization.
• Improve performance on Gaussian likelihood:
  • mean_vals for centering; by default, it is set to be the observed means of all Gaussian features
• Improve range constraints:
  • min_vals and max_vals; the default values depends on the likelihood.

Minor updates:

• Minor update of training augment
• Add num_repeat option for training small datasets
• Set default values: 'dist_block' = ['Gaussian', ...], learning_rate=3-4.

• Allow dim_input_arr and dim_block to be None.
• Adjust batch size when larger than the sample size.

What's Changed

Major updates

• Fix bug about masking #3 when computing latent embeddings
• Add MMD loss for batch correction in the latent space
• Add options for continuing training
• Add options for skip connections to improve imputation quality
• Fix bug about activation function (sigmoid for Bernoulli, scaled for NB, which was used in the initial version)

Minor updates

• v0.1.0 slight improvement #5
• v0.2.0 add max_vals argument, which can be provided for each block
• Improve efficiency by decorating tf.function
• Improve documentation comments

Test

Time consumption on integrating DOGMA-seq, CITE-seq, and ASAP-seq datasets (30987 cells and 42598 features):

• On CPU with 12 cores and 128GB of RAM, 80s per epoch, ~11h for 500 epochs
• On NVIDIA GeForce RTX 4090, 10s per epoch, ~1.4h for 500 epochs
• On NVIDIA A100 (Colab pro+), 30s per epoch, ~4h for 500 epochs

Version as supplements to the paper