segment-anything: fix class name typo

Note: weights are impacted

scripts/prepare_test_weights.py

@@ -581,7 +581,7 @@ def convert_sam():
         "convert_segment_anything.py",
         "tests/weights/sam_vit_h_4b8939.pth",
         "tests/weights/segment-anything-h.safetensors",
-        expected_hash="3b73b2fd",
+        expected_hash="b62ad5ed",
     )
 
 

src/refiners/foundationals/segment_anything/mask_decoder.py

@@ -5,7 +5,7 @@ import refiners.fluxion.layers as fl
 from refiners.fluxion.context import Contexts
 from refiners.foundationals.segment_anything.transformer import (
     SparseCrossDenseAttention,
-    TwoWayTranformerLayer,
+    TwoWayTransformerLayer,
 )
 
 
@@ -210,7 +210,7 @@ class MaskDecoder(fl.Chain):
             EmbeddingsAggregator(num_output_mask=num_output_mask),
             Transformer(
                 *(
-                    TwoWayTranformerLayer(
+                    TwoWayTransformerLayer(
                         embedding_dim=embedding_dim,
                         num_heads=8,
                         feed_forward_dim=feed_forward_dim,

src/refiners/foundationals/segment_anything/transformer.py

@@ -116,7 +116,7 @@ class DenseCrossSparseAttention(fl.Chain):
         )
 
 
-class TwoWayTranformerLayer(fl.Chain):
+class TwoWayTransformerLayer(fl.Chain):
     def __init__(
         self,
         embedding_dim: int,

tests/foundationals/segment_anything/test_sam.py

@@ -21,7 +21,7 @@ from refiners.fluxion.model_converter import ModelConverter
 from refiners.fluxion.utils import image_to_tensor, load_tensors, no_grad
 from refiners.foundationals.segment_anything.image_encoder import FusedSelfAttention
 from refiners.foundationals.segment_anything.model import SegmentAnythingH
-from refiners.foundationals.segment_anything.transformer import TwoWayTranformerLayer
+from refiners.foundationals.segment_anything.transformer import TwoWayTransformerLayer
 
 # See predictor_example.ipynb official notebook
 PROMPTS: list[SAMPrompt] = [
@@ -188,7 +188,7 @@ def test_two_way_transformer(facebook_sam_h: FacebookSAM) -> None:
     dense_positional_embedding = torch.randn(1, 64 * 64, 256, device=facebook_sam_h.device)
     sparse_embedding = torch.randn(1, 3, 256, device=facebook_sam_h.device)
 
-    refiners_layer = TwoWayTranformerLayer(
+    refiners_layer = TwoWayTransformerLayer(
         embedding_dim=256, feed_forward_dim=2048, num_heads=8, device=facebook_sam_h.device
     )
     facebook_layer = facebook_sam_h.mask_decoder.transformer.layers[1]  # type: ignore