fix-matryoshka-normalization (#24)
Browse files- fix: matryoshka normalization (a4de150f6b126c6bb858fd0b999cf4862b75d327)
- custom_st.py +4 -5
- modeling_jina_embeddings_v4.py +1 -0
custom_st.py
CHANGED
@@ -45,7 +45,6 @@ class Transformer(nn.Module):
|
|
45 |
self.model = AutoModel.from_pretrained(
|
46 |
model_name_or_path, config=self.config, cache_dir=cache_dir, **model_kwargs
|
47 |
)
|
48 |
-
|
49 |
self.processor = AutoProcessor.from_pretrained(
|
50 |
model_name_or_path,
|
51 |
cache_dir=cache_dir,
|
@@ -133,14 +132,13 @@ class Transformer(nn.Module):
|
|
133 |
if k.startswith("text_") and k != "text_indices"
|
134 |
}
|
135 |
text_indices = features.get("text_indices", [])
|
136 |
-
|
137 |
-
with torch.autocast(device_type=device):
|
138 |
text_embeddings = self.model(
|
139 |
**text_batch, task_label=task
|
140 |
).single_vec_emb
|
141 |
if self.config.truncate_dim:
|
142 |
text_embeddings = text_embeddings[:, : self.config.truncate_dim]
|
143 |
-
|
144 |
for i, embedding in enumerate(text_embeddings):
|
145 |
all_embeddings.append((text_indices[i], embedding))
|
146 |
|
@@ -152,12 +150,13 @@ class Transformer(nn.Module):
|
|
152 |
}
|
153 |
image_indices = features.get("image_indices", [])
|
154 |
|
155 |
-
with torch.autocast(device_type=device):
|
156 |
img_embeddings = self.model(
|
157 |
**image_batch, task_label=task
|
158 |
).single_vec_emb
|
159 |
if self.config.truncate_dim:
|
160 |
img_embeddings = img_embeddings[:, : self.config.truncate_dim]
|
|
|
161 |
|
162 |
for i, embedding in enumerate(img_embeddings):
|
163 |
all_embeddings.append((image_indices[i], embedding))
|
|
|
45 |
self.model = AutoModel.from_pretrained(
|
46 |
model_name_or_path, config=self.config, cache_dir=cache_dir, **model_kwargs
|
47 |
)
|
|
|
48 |
self.processor = AutoProcessor.from_pretrained(
|
49 |
model_name_or_path,
|
50 |
cache_dir=cache_dir,
|
|
|
132 |
if k.startswith("text_") and k != "text_indices"
|
133 |
}
|
134 |
text_indices = features.get("text_indices", [])
|
135 |
+
with torch.autocast(device_type=device, dtype=torch.bfloat16):
|
|
|
136 |
text_embeddings = self.model(
|
137 |
**text_batch, task_label=task
|
138 |
).single_vec_emb
|
139 |
if self.config.truncate_dim:
|
140 |
text_embeddings = text_embeddings[:, : self.config.truncate_dim]
|
141 |
+
text_embeddings = torch.nn.functional.normalize(text_embeddings, p=2, dim=-1)
|
142 |
for i, embedding in enumerate(text_embeddings):
|
143 |
all_embeddings.append((text_indices[i], embedding))
|
144 |
|
|
|
150 |
}
|
151 |
image_indices = features.get("image_indices", [])
|
152 |
|
153 |
+
with torch.autocast(device_type=device, dtype=torch.bfloat16):
|
154 |
img_embeddings = self.model(
|
155 |
**image_batch, task_label=task
|
156 |
).single_vec_emb
|
157 |
if self.config.truncate_dim:
|
158 |
img_embeddings = img_embeddings[:, : self.config.truncate_dim]
|
159 |
+
img_embeddings = torch.nn.functional.normalize(img_embeddings, p=2, dim=-1)
|
160 |
|
161 |
for i, embedding in enumerate(img_embeddings):
|
162 |
all_embeddings.append((image_indices[i], embedding))
|
modeling_jina_embeddings_v4.py
CHANGED
@@ -350,6 +350,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
|
|
350 |
embeddings = embeddings.single_vec_emb
|
351 |
if truncate_dim is not None:
|
352 |
embeddings = embeddings[:, :truncate_dim]
|
|
|
353 |
else:
|
354 |
embeddings = embeddings.multi_vec_emb
|
355 |
if return_multivector and not return_numpy:
|
|
|
350 |
embeddings = embeddings.single_vec_emb
|
351 |
if truncate_dim is not None:
|
352 |
embeddings = embeddings[:, :truncate_dim]
|
353 |
+
embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=-1)
|
354 |
else:
|
355 |
embeddings = embeddings.multi_vec_emb
|
356 |
if return_multivector and not return_numpy:
|