amiralavi / scipr Goto Github PK

Add tensorboard writers. Tensorboard is included in one of our dependencies, pytorch, so it makes sense to interact with it through pytorch.

Add arguments to the SCIPR.fit function:

• tensorboard=True; enable tensorboard logging
• tensorboard_dir=None; if specified, write tensorboard session to this directory. If not specified, warn that it's not specified and let user know what auto-generated path is being used.

On SCIPR.fit, add tensorboard tags for:

• number of matches
• avg distance between matches

Add support to feed an adata object into the fit and predict functions, something like:

def fit(self, adata, batch_key, source, target)

Where batch_key is the name of the column in obsm which contains the batch annotations, and source and target are the batch annotations of the source and target batches.

def transform(self, adata, batch_key, batch, inplace=False)

batch is the batch annotation for the batch to transform. If inplace==True then replace the cells (observations) in the adata.X which have the batch annotation with the new transformed values. Otherwise, return the tuple (transformed numpy.ndarray, row indexer into adata of the transformed cells)

• scipr version: 0.2.1
• Python version: 3.9.7
• Operating System: Mac OS Monterey

Description

Describe what you were trying to get done.
I 'd like to adapt and use scipr with flow and mass cytomtery data sets.

Tell us what happened, what went wrong, and what you expected to happen.
Using simulated data, MNN matching, and an "affine" transform, the transformed object ends up being a single point repeated N times. If I use "rigid", I get points on a curve.

What I Did

As input, I simulated a toy 2D data set containing 500 points in 4 clusters (data set "A" in the code below). Then I affine-transformed "A" and called that "B".

Paste the command(s) you ran and the output.

import scipr
from scipr.matching import Closest, Hungarian, Greedy, MNN
from scipr.transform import Affine, Rigid

closest = Closest()
MNN = MNN()
rigid = Rigid()
affine = Affine()

model = scipr.SCIPR(match_algo=MNN,
transform_algo=affine,
input_normalization='l2',
n_iter = 100)
model.fit(A, B)
aligned = model.transform(A)

I get back a single coordinate:
aligned[:5,]
array([[0.71251583, 0.68878496],
[0.71251583, 0.68878496],
[0.71251583, 0.68878496],
[0.71251583, 0.68878496],
[0.71251583, 0.68878496]])

Part of the issue is the matching. If I replace the "match_alg"o with one that just passes a predefined set of matching points, then I get back a transformed object that looks reasonably like the reference, but ONLY if the "transform_algo" is "rigid". For "affine" I get back a straight-line. Any ideas why?

THANK YOU.

Hamid

If there was a crash, please include the traceback here.

Add a boolean flag, "verbosity", set to 0 by default to have quite functions. If things work fine there should be nothing written to stdout. If user sets verbosity > 0, then emit some logging info.