unmixing 0.2.0

No project description provided

This repo is an implementation of stochastic linear unmixing using pytorch.
It is a fast and efficient, and can unmix even if there are null values.

$H$ is a 2d metrix, each row in it is a single end member.
Each column is a single feature of the end members.
No null values are allowed in $H$.
$O$ is a 2d metrix, each row a single observed instance.
The columns of $O$ are the same as the columns of $H$.
Null values are allowed in $O$.
The main class of this repo, LinearUnmixing, is calculating $R$ matrix
that minimizes the loss between $O$ and $H \cdot R$.
$R$ is constrained to be non-negative and each row in R sums to 1.
Reconstruction error can be used for filtering, since there is no
assessment for $R$, but we can asses the distance between $O$ and $H \cdot R$.

requirements:
numpy
torch

Installation:
pip install unmixing

Usage example:

n_endmembers = 5
n_features = 4
n_rows = 1_000

endmembers = torch.randint(1, 10, (n_endmembers, n_features), dtype=torch.float32)
non_negative = torch.randn((n_rows, n_endmembers)) ** 2
rates = non_negative / non_negative.sum(axis=1).reshape((-1, 1))
observed = torch.matmul(rates, endmembers)
lm = LinearUnmixing(endmembers, observed)
lm.unmix()
pred_rates = lm.get_rates()

rates_errors = (pred_rates - rates.numpy()).sum(axis=1).round(2)
print(f"first example - max rate error: {rates_errors.max()}")

reconstruction_error_per_observation = lm.get_errors()
print(f"first example - mean absolute reconstruction error: "
      f"{np.mean(np.abs(reconstruction_error_per_observation))}")

detailed_reconstruction_error = lm.get_errors_per_column()
print(f"first example - mean absolute reconstruction error per entry: "
      f"{np.mean(np.abs(detailed_reconstruction_error))}")