v0.1.2

Added a custom iterator support

v0.1.1

• Adaptive learning rate with some modification
• A bit more careful work with VRAM
• Minor improvements

v0.1.0

Some major updates for the original algorithm:

• A method for the projection of new data point using existing feature weights is added, model.project(Y). In this case, only factors will be trained, with loadings fixed.
• Now, learning rate is adaptive and is decreased during if the model's loglikelihood doesn't increase.

v0.0.3

In this version, we added a method to generate a counts reconstruction based on learned Poisson PCA: model.reconstruct_counts(). Also some minor modifications and improvements.

v0.0.2

Here we added major improvements:

1. Now we avoid calculation of dense full-size matrix via batching calculation of log-likelihood;
2. We implemented cyclic coordinate descent.

Learning rate is still fixed, in our tests this approach is good enough for most of the situations (line search is very computationally expensive)

v0.0.1

The first release of the package.

Here we're implementing quick optimisation for Poisson GLM-PCA in PyTorch. In contrast to R's fastglmpca package, with PyTorch computations might be done with GPU which significantly decreases computational time.