constraining individual weights

[Eduardinho17]

Hi,

I'm working on a data analysis and I'm giving 2 components to my data for NMF. Is it possible to constrain only 1 of the component's weight? Also, what we know is that in the second component, the weight should decay exponentially. Is it possible, for instance, to only fix a parameter of the exponential function?

[maffettone]

The first thing should be possible, but the second is not currently possible. It does sound like an interesting feature worth developing though. I'll have a think on that. One approach for now would be to initialize the second set of weights to an exponentially decaying guess, and try and nudge the optimization in that direction.

X = load_my_dataset() # torch tensor shape (m_example, n_features)
weights_1 = load_expected_weights() # torch tensor shape (m_examples, k_components)
weights_2 = load_best_guess_exponential_weights() 

model = cmf.nmf.models.NMF(X.shape,
                           2,
                           initial_weights=[weights_1, weights_2],
                           fix_weights=[True, False])
loss = model.fit(X) # Learning curve of loss over timesteps
learned_weights, learned_components = model.W.detach().numpy(), model.H.detach().numpy()