fix: remove padding vectors

#1
Files changed (1) hide show
  1. similarity.py +2 -0
similarity.py CHANGED
@@ -217,6 +217,8 @@ class JinaV4SimilarityMapper:
217
  truncate_dim=self.num_vectors
218
  )
219
  image_embeddings = image_embeddings[0] # [num_patches/num_vectors, embed_dim]
 
 
220
 
221
  # <|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Describe the image.<|im_end|>\n
222
  vision_start_position_from_start = 3 + 1
 
217
  truncate_dim=self.num_vectors
218
  )
219
  image_embeddings = image_embeddings[0] # [num_patches/num_vectors, embed_dim]
220
+ non_zero_mask = (image_embeddings.abs().sum(dim=1) > 0)
221
+ image_embeddings = image_embeddings[non_zero_mask]
222
 
223
  # <|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Describe the image.<|im_end|>\n
224
  vision_start_position_from_start = 3 + 1