Error in scanvi_ref multiclass accuracy calculation (#2938)

added -1 to the squeeze in order to deal with the case of dim==1 in a
multiclass batch where there is 1 sample for the usage of scvi-> scanvi
transfer learning

close #2934

---------

Co-authored-by: Can Ergen <canergen.ac@gmail.com>

CHANGELOG.md

@@ -98,6 +98,7 @@ to [Semantic Versioning]. Full commit history is available in the
 
 #### Fixed
 
+- Fix logging of accuracy for cases with 1 sample per class in scANVI {pr}`2938`.
 - Disable adversarial classifier if training with a single batch.
     Previously this raised a None error {pr}`2914`.
 - {meth}`~scvi.model.SCVI.get_normalized_expression` fixed for Poisson distribution and

src/scvi/train/_trainingplans.py

@@ -741,7 +741,7 @@ def compute_and_log_metrics(
             return
 
         classification_loss = loss_output.classification_loss
-        true_labels = loss_output.true_labels.squeeze()
+        true_labels = loss_output.true_labels.squeeze(-1)
         logits = loss_output.logits
         predicted_labels = torch.argmax(logits, dim=-1)