feat: Add random state feature. #150
Conversation
john-halloran
commented
Jun 6, 2025
john-halloran
commented
- feat: Added random_state feature for reproducibility.
| components=None, | ||
| random_state=None, | ||
| ): | ||
|
|
||
| self.MM = MM |
There was a problem hiding this comment.
Changed to n_components, which is what sklearn.decomposition.NMF uses.
| MM, | ||
| Y0=None, | ||
| X0=None, | ||
| A=None, |
There was a problem hiding this comment.
There are many different standards for what to name these matrices. Zero agreement between sources that use NMF. I'm inclined to eventually use what sklearn.decomposition.non_negative_factorization uses, which would mean MM->X, X->W, Y->H. But I'd like to leave this as is for the moment until there's a consensus about what would be the most clear or standard. If people will be finding this tool from the sNMF paper, there's also an argument for using the X, Y, and A names because that was used there.
| @@ -4,8 +4,20 @@ | |||
|
|
|||
|
|
|||
| class SNMFOptimizer: | |||
| def __init__(self, MM, Y0=None, X0=None, A=None, rho=1e12, eta=610, max_iter=500, tol=5e-7, components=None): | |||
There was a problem hiding this comment.
we need a docstring here and in the init. Please see scikit-package FAQ about how to write these. Also, look at Yucong's code or diffpy.utils?
There was a problem hiding this comment.
Added one here. The package init dates back to the old codebase, but as soon as that is updated it will get a docstring as well.
| @@ -15,23 +27,22 @@ def __init__(self, MM, Y0=None, X0=None, A=None, rho=1e12, eta=610, max_iter=500 | |||
| # Capture matrix dimensions | |||
| self.N, self.M = MM.shape | |||
| self.num_updates = 0 | |||
| self.rng = np.random.default_rng(random_state) | |||
There was a problem hiding this comment.
can we have a more descriptive variable name? Is this a range? What is the range?
| if self.A is None: | ||
| self.A = np.ones((self.K, self.M)) + np.random.randn(self.K, self.M) * 1e-3 # Small perturbation | ||
| self.A = np.ones((self.K, self.M)) + self.rng.normal(0, 1e-3, size=(self.K, self.M)) |
There was a problem hiding this comment.
K and M are probably good names if the matrix decomposition equation is in hte docstring, so they get defined there.
Thanks, will work on resolving these. To be clear, for things like the docstrings would you prefer I make new PRs, get those merged, then rebase this one, or just add to this existing PR? |
|
For now, I will assume anything given as feedback in this PR could be in scope to include. |
| @@ -17,6 +17,64 @@ def __init__( | |||
| components=None, | |||
| random_state=None, | |||
| ): | |||
| """Run sNMF based on an ndarray, parameters, and either a number | |||
There was a problem hiding this comment.
This is fantastic! Thanks for this. Please see here for our docstring standards, I am not sure if you looked at it:
https://scikit-package.github.io/scikit-package/programming-guides/billinge-group-standards.html#docstrings
For classes it is a bit tricky because what info do we put in the "Class" docstring and what info do we put in the "constructor" (i.e., the __init__()) docstring. After some googling we came up with the breakout that is shown in the DiffractionObjects class that is shown there. We would be after something similar here.
By way of example, I would probably do like this in this case
def SNMFOptimizer:
'''Configuration and methods to run the stretched NMF algorithm, sNMF
Instantiating the SNMFOptimizer class runs all the analysis
immediately. The results can then be accessed as instance attributes
of the class (X, Y, and A).
Please see <reference to paper> for more details about the algorithm.
Attributes
-----------
mm : ndarray
The array containing the data to be decomposed. Shape is (length_of_signal, number_of_conditions)
y0 : ndarray
The array containing initial guesses for the component weights
at each stretching condition. Shape is (number_of_components, number_of_conditions
...
'''
put future development plans into issues, not in the docstring. Just describe the current behavior. Try and keep it brief but highly informational.
To conform to PEP8 standards I lower-cased the variables. I know they correspond to matrices but we should decide which standard to break. The tie-breaker should probably be scikit-learn. Whatever they do, let's do that. Let's also add a small comment (not in the docstring) to remind ourselves in the future if it breaks PEP8 or it will annoy me every time we revisit it and I will try and change it back......
Conditions on instantiation will go in the constructor docstring.
That one describes the init method so should look more like a function docstring. It would look something like....
def __init__(mm....)
'''Initialize a SNMFOptimizer instance and run the optimization
Parameters
------------
mm : ndarray
The array containing the data to be decomposed. Shape is (length_of_signal, number_of_conditions)
y0 : ndarray Optional. Defaults to None
The array containing initial guesses for the component weights
at each stretching condition. Shape is (number_of_components, number_of_conditions
...
I think there was some text before about how Y0 was required. But if it is required it may be better to make it a required (positional) variable in the constructor and not have it optional. we can discuss design decisions too if you like.
