vector-quantize-pytorch 1.22.2__tar.gz → 1.22.4__tar.gz

{vector_quantize_pytorch-1.22.2 → vector_quantize_pytorch-1.22.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vector-quantize-pytorch
-Version: 1.22.2
+Version: 1.22.4
 Summary: Vector Quantization - Pytorch
 Project-URL: Homepage, https://pypi.org/project/vector-quantize-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/vector-quantizer-pytorch
@@ -324,7 +324,7 @@ This repository should also automatically synchronizing the codebooks in a multi
 
 <img src="./images/simvq.png" width="400px"></img>
 
-A <a href="https://arxiv.org/abs/2411.02038">new ICLR 2025 paper</a> proposes a scheme where the codebook is frozen, and the codes are implicitly generated through a linear projection. The authors claim this setup leads to less codebook collapse as well as easier convergence. I have found this to perform even better when paired with <a href="https://arxiv.org/abs/2410.06424">rotation trick</a> from Fifty et al., and expanding the linear projection to a small one layer MLP. You can experiment with it as so
+A <a href="https://arxiv.org/abs/2411.02038">new paper</a> proposes a scheme where the codebook is frozen, and the codes are implicitly generated through a linear projection. The authors claim this setup leads to less codebook collapse as well as easier convergence. I have found this to perform even better when paired with <a href="https://arxiv.org/abs/2410.06424">rotation trick</a> from Fifty et al., and expanding the linear projection to a small one layer MLP. You can experiment with it as so
 
 Update: hearing mixed results
 

{vector_quantize_pytorch-1.22.2 → vector_quantize_pytorch-1.22.4}/README.md

@@ -280,7 +280,7 @@ This repository should also automatically synchronizing the codebooks in a multi
 
 <img src="./images/simvq.png" width="400px"></img>
 
-A <a href="https://arxiv.org/abs/2411.02038">new ICLR 2025 paper</a> proposes a scheme where the codebook is frozen, and the codes are implicitly generated through a linear projection. The authors claim this setup leads to less codebook collapse as well as easier convergence. I have found this to perform even better when paired with <a href="https://arxiv.org/abs/2410.06424">rotation trick</a> from Fifty et al., and expanding the linear projection to a small one layer MLP. You can experiment with it as so
+A <a href="https://arxiv.org/abs/2411.02038">new paper</a> proposes a scheme where the codebook is frozen, and the codes are implicitly generated through a linear projection. The authors claim this setup leads to less codebook collapse as well as easier convergence. I have found this to perform even better when paired with <a href="https://arxiv.org/abs/2410.06424">rotation trick</a> from Fifty et al., and expanding the linear projection to a small one layer MLP. You can experiment with it as so
 
 Update: hearing mixed results
 

{vector_quantize_pytorch-1.22.2 → vector_quantize_pytorch-1.22.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "vector-quantize-pytorch"
-version = "1.22.2"
+version = "1.22.4"
 description = "Vector Quantization - Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{vector_quantize_pytorch-1.22.2 → vector_quantize_pytorch-1.22.4}/vector_quantize_pytorch/residual_vq.py

@@ -85,6 +85,8 @@ class MLP(Module):
     ):
         one_headed = codes.ndim == 2
 
+        condition, _ = pack([condition], 'b * d') # handle condition with ndim 2 - (batch, dim)
+
         if one_headed:
             codes = rearrange(codes, 'c d -> 1 c d')
 

{vector_quantize_pytorch-1.22.2 → vector_quantize_pytorch-1.22.4}/vector_quantize_pytorch/vector_quantize_pytorch.py

@@ -559,9 +559,18 @@ class EuclideanCodebook(Module):
 
         else:
             if exists(codebook_transform_fn):
-                quantize = einx.get_at('h b n [c] d, h b n -> h b n d', transformed_embed, embed_ind)
+                # quantize = einx.get_at('h b n [c] d, h b n -> h b n d', transformed_embed, embed_ind)
+
+                repeated_embed_ind = repeat(embed_ind, 'h b n -> h b n 1 d', d = transformed_embed.shape[-1])
+                quantize = transformed_embed.gather(-2, repeated_embed_ind)
+                quantize = rearrange(quantize, 'h b n 1 d -> h b n d')
+
             else:
-                quantize = einx.get_at('h [c] d, h b n -> h b n d', embed, embed_ind)
+                # quantize = einx.get_at('h [c] d, h b n -> h b n d', embed, embed_ind)
+
+                repeated_embed = repeat(embed, 'h c d -> h b c d', b = embed_ind.shape[1])
+                repeated_embed_ind = repeat(embed_ind, 'h b n -> h b n d', d = embed.shape[-1])
+                quantize = repeated_embed.gather(-2, repeated_embed_ind)
 
         if self.training and self.ema_update and not freeze_codebook:
 
@@ -767,9 +776,18 @@ class CosineSimCodebook(Module):
 
         else:
             if exists(codebook_transform_fn):
-                quantize = einx.get_at('h b n [c] d, h b n -> h b n d', transformed_embed, embed_ind)
+                # quantize = einx.get_at('h b n [c] d, h b n -> h b n d', transformed_embed, embed_ind)
+
+                repeated_embed_ind = repeat(embed_ind, 'h b n -> h b n 1 d', d = transformed_embed.shape[-1])
+                quantize = transformed_embed.gather(-2, repeated_embed_ind)
+                quantize = rearrange(quantize, 'h b n 1 d -> h b n d')
+
             else:
-                quantize = einx.get_at('h [c] d, h b n -> h b n d', embed, embed_ind)
+                # quantize = einx.get_at('h [c] d, h b n -> h b n d', embed, embed_ind)
+
+                repeated_embed = repeat(embed, 'h c d -> h b c d', b = embed_ind.shape[1])
+                repeated_embed_ind = repeat(embed_ind, 'h b n -> h b n d', d = embed.shape[-1])
+                quantize = repeated_embed.gather(-2, repeated_embed_ind)
 
         if self.training and self.ema_update and not freeze_codebook:
             if exists(mask):