Bugfix/no ref/ohe jitter

CHANGELOG.md

@@ -54,12 +54,12 @@
 * Fix `methods/scgpt_zeroshot` and `methods/scgpt_finetuned` failing to build: stop installing `flash-attn`. Both
     scripts pass `use_fast_transformer=False`, so it was never used. Behind it sat three more pins with no python 3.12
     wheels, now `numpy<2`, `torchtext==0.17.2` and `transformers==4.36.2`.
-
 * Split Scanorama into two methods/scores
     - Split Scanorama into embedding (integrate) and count-correction (correct) modes, instead of running both together. 
         This makes clear what the reported score(s) are describing, and also corrects the misleadingly low score that 
         the combined method receives. The scores for each component  are in line with their scores from v1, where the modes 
-        were separated.  
+        were separated.
+* Remove jitter from the `embed_cell_types` control method, distinguishing its behavior from `embed_cell_types_jittered`.
 
 # task_batch_integration 2.0.0
 

src/control_methods/utils.py

@@ -40,7 +40,7 @@ def _randomize_graph(adata, partition=None, neighbors_key="neighbors"):
     return adata
 
 
-def _perfect_embedding(partition, jitter=0.01):
+def _perfect_embedding(partition, jitter=None):
     """
     Taken and adapted from opsca-v1:
     https://github.com/openproblems-bio/openproblems/blob/acf5c95a7306b819c4a13972783433d0a48f769b/openproblems/tasks/_batch_integration/_common/methods/baseline.py#L37
@@ -51,6 +51,7 @@ def _perfect_embedding(partition, jitter=0.01):
     embedding = OneHotEncoder().fit_transform(
         LabelEncoder().fit_transform(partition)[:, None]
     )
+    embedding = embedding.toarray()
     if jitter is not None:
         embedding = embedding + np.random.uniform(-1 * jitter, jitter, embedding.shape)
     return np.asarray(embedding)