more efficient to select out the frozen codes and then do projection …

…in `indices_to_codes` for SimVQ
lucidrains · Nov 11, 2024 · 8f5b428 · 8f5b428
1 parent a0e8f2c
commit 8f5b428
Show file tree

Hide file tree

Showing 2 changed files with 5 additions and 4 deletions.
diff --git a/pyproject.toml b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "vector-quantize-pytorch"
-version = "1.20.3"
+version = "1.20.4"
 description = "Vector Quantization - Pytorch"
 authors = [
     { name = "Phil Wang", email = "[email protected]" }

diff --git a/vector_quantize_pytorch/sim_vq.py b/vector_quantize_pytorch/sim_vq.py
@@ -56,7 +56,7 @@ def __init__(
         if not exists(codebook_transform):
             codebook_transform = nn.Linear(dim, dim, bias = False)
 
-        self.codebook_to_codes = codebook_transform
+        self.code_transform = codebook_transform
 
         self.register_buffer('frozen_codebook', codebook)
 
@@ -72,15 +72,16 @@ def __init__(
 
     @property
     def codebook(self):
-        return self.codebook_to_codes(self.frozen_codebook)
+        return self.code_transform(self.frozen_codebook)
 
     def indices_to_codes(
         self,
         indices
     ):
         implicit_codebook = self.codebook
 
-        quantized = get_at('[c] d, b ... -> b ... d', implicit_codebook, indices)
+        frozen_codes = get_at('[c] d, b ... -> b ... d', self.frozen_codebook, indices)
+        quantized = self.code_transform(frozen_codes)
 
         if self.accept_image_fmap:
             quantized = rearrange(quantized, 'b ... d -> b d ...')