Create custom_st.py
#2
by
Samoed
- opened
- README.md +46 -39
- custom_st.py +110 -0
README.md
CHANGED
@@ -44,57 +44,64 @@ pip install -r requirements.txt
|
|
44 |
|
45 |
Then you can enter the directory to run the following command.
|
46 |
```python
|
47 |
-
from
|
48 |
-
from src.arguments import ModelArguments
|
49 |
-
from src.utils import load_processor
|
50 |
import torch
|
51 |
-
from transformers import HfArgumentParser, AutoProcessor
|
52 |
from PIL import Image
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
model.eval()
|
62 |
-
|
63 |
# Image + Text -> Text
|
64 |
inputs = processor(text='<|image|><|begin_of_text|> Represent the given image with the following question: What is in the image', images=[Image.open(
|
65 |
-
'figures/example.jpg')], return_tensors="pt")
|
66 |
-
|
67 |
-
|
68 |
string = 'A cat and a dog'
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
print(string, '=', model.compute_similarity(qry_output, tgt_output))
|
73 |
## A cat and a dog = tensor([[0.3965]], device='cuda:0', dtype=torch.bfloat16)
|
|
|
74 |
string = 'A cat and a tiger'
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
print(string, '=', model.compute_similarity(qry_output, tgt_output))
|
79 |
## A cat and a tiger = tensor([[0.3105]], device='cuda:0', dtype=torch.bfloat16)
|
|
|
80 |
# Text -> Image
|
81 |
-
inputs = processor(text='Find me an everyday image that matches the given caption: A cat and a dog.', return_tensors="pt")
|
82 |
-
|
83 |
-
|
84 |
string = '<|image|><|begin_of_text|> Represent the given image.'
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
print(string, '=', model.compute_similarity(qry_output, tgt_output))
|
89 |
## <|image|><|begin_of_text|> Represent the given image. = tensor([[0.4219]], device='cuda:0', dtype=torch.bfloat16)
|
90 |
-
|
91 |
-
inputs =
|
92 |
-
qry_output = model(
|
93 |
string = '<|image|><|begin_of_text|> Represent the given image.'
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
print(string, '=', model.compute_similarity(qry_output, tgt_output))
|
98 |
## <|image|><|begin_of_text|> Represent the given image. = tensor([[0.3887]], device='cuda:0', dtype=torch.bfloat16)
|
99 |
```
|
100 |
|
@@ -106,4 +113,4 @@ print(string, '=', model.compute_similarity(qry_output, tgt_output))
|
|
106 |
journal={arXiv preprint arXiv:2502.08468},
|
107 |
year={2025}
|
108 |
}
|
109 |
-
```
|
|
|
44 |
|
45 |
Then you can enter the directory to run the following command.
|
46 |
```python
|
47 |
+
from transformers import MllamaForConditionalGeneration, AutoProcessor
|
|
|
|
|
48 |
import torch
|
|
|
49 |
from PIL import Image
|
50 |
+
|
51 |
+
# Pooling and Normalization
|
52 |
+
def last_pooling(last_hidden_state, attention_mask, normalize=True):
|
53 |
+
sequence_lengths = attention_mask.sum(dim=1) - 1
|
54 |
+
batch_size = last_hidden_state.shape[0]
|
55 |
+
reps = last_hidden_state[torch.arange(batch_size, device=last_hidden_state.device), sequence_lengths]
|
56 |
+
if normalize:
|
57 |
+
reps = torch.nn.functional.normalize(reps, p=2, dim=-1)
|
58 |
+
return reps
|
59 |
+
|
60 |
+
def compute_similarity(q_reps, p_reps):
|
61 |
+
return torch.matmul(q_reps, p_reps.transpose(0, 1))
|
62 |
+
|
63 |
+
model_name = "intfloat/mmE5-mllama-11b-instruct"
|
64 |
+
|
65 |
+
# Load Processor and Model
|
66 |
+
processor = AutoProcessor.from_pretrained(model_name)
|
67 |
+
model = MllamaForConditionalGeneration.from_pretrained(
|
68 |
+
model_name, torch_dtype=torch.bfloat16
|
69 |
+
).to("cuda")
|
70 |
model.eval()
|
71 |
+
|
72 |
# Image + Text -> Text
|
73 |
inputs = processor(text='<|image|><|begin_of_text|> Represent the given image with the following question: What is in the image', images=[Image.open(
|
74 |
+
'figures/example.jpg')], return_tensors="pt").to("cuda")
|
75 |
+
qry_output = last_pooling(model(**inputs, return_dict=True, output_hidden_states=True).hidden_states[-1], inputs['attention_mask'])
|
76 |
+
|
77 |
string = 'A cat and a dog'
|
78 |
+
text_inputs = processor(text=string, return_tensors="pt").to("cuda")
|
79 |
+
tgt_output = last_pooling(model(**text_inputs, return_dict=True, output_hidden_states=True).hidden_states[-1], text_inputs['attention_mask'])
|
80 |
+
print(string, '=', compute_similarity(qry_output, tgt_output))
|
|
|
81 |
## A cat and a dog = tensor([[0.3965]], device='cuda:0', dtype=torch.bfloat16)
|
82 |
+
|
83 |
string = 'A cat and a tiger'
|
84 |
+
text_inputs = processor(text=string, return_tensors="pt").to("cuda")
|
85 |
+
tgt_output = last_pooling(model(**text_inputs, return_dict=True, output_hidden_states=True).hidden_states[-1], text_inputs['attention_mask'])
|
86 |
+
print(string, '=', compute_similarity(qry_output, tgt_output))
|
|
|
87 |
## A cat and a tiger = tensor([[0.3105]], device='cuda:0', dtype=torch.bfloat16)
|
88 |
+
|
89 |
# Text -> Image
|
90 |
+
inputs = processor(text='Find me an everyday image that matches the given caption: A cat and a dog.', return_tensors="pt").to("cuda")
|
91 |
+
qry_output = last_pooling(model(**inputs, return_dict=True, output_hidden_states=True).hidden_states[-1], inputs['attention_mask'])
|
92 |
+
|
93 |
string = '<|image|><|begin_of_text|> Represent the given image.'
|
94 |
+
tgt_inputs = processor(text=string, images=[Image.open('figures/example.jpg')], return_tensors="pt").to("cuda")
|
95 |
+
tgt_output = last_pooling(model(**tgt_inputs, return_dict=True, output_hidden_states=True).hidden_states[-1], tgt_inputs['attention_mask'])
|
96 |
+
print(string, '=', compute_similarity(qry_output, tgt_output))
|
|
|
97 |
## <|image|><|begin_of_text|> Represent the given image. = tensor([[0.4219]], device='cuda:0', dtype=torch.bfloat16)
|
98 |
+
|
99 |
+
inputs = processor(text='Find me an everyday image that matches the given caption: A cat and a tiger.', return_tensors="pt").to("cuda")
|
100 |
+
qry_output = last_pooling(model(**inputs, return_dict=True, output_hidden_states=True).hidden_states[-1], inputs['attention_mask'])
|
101 |
string = '<|image|><|begin_of_text|> Represent the given image.'
|
102 |
+
tgt_inputs = processor(text=string, images=[Image.open('figures/example.jpg')], return_tensors="pt").to("cuda")
|
103 |
+
tgt_output = last_pooling(model(**tgt_inputs, return_dict=True, output_hidden_states=True).hidden_states[-1], tgt_inputs['attention_mask'])
|
104 |
+
print(string, '=', compute_similarity(qry_output, tgt_output))
|
|
|
105 |
## <|image|><|begin_of_text|> Represent the given image. = tensor([[0.3887]], device='cuda:0', dtype=torch.bfloat16)
|
106 |
```
|
107 |
|
|
|
113 |
journal={arXiv preprint arXiv:2502.08468},
|
114 |
year={2025}
|
115 |
}
|
116 |
+
```
|
custom_st.py
ADDED
@@ -0,0 +1,110 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from io import BytesIO
|
2 |
+
from typing import Any, Dict, Optional, List
|
3 |
+
import torch
|
4 |
+
from PIL import Image
|
5 |
+
from transformers import AutoProcessor, MllamaForConditionalGeneration
|
6 |
+
from sentence_transformers.models import Transformer as BaseTransformer
|
7 |
+
|
8 |
+
|
9 |
+
class MultiModalTransformer(BaseTransformer):
|
10 |
+
def __init__(
|
11 |
+
self,
|
12 |
+
model_name_or_path: str,
|
13 |
+
cache_dir: Optional[str] = None,
|
14 |
+
tokenizer_args: Optional[Dict[str, Any]] = None,
|
15 |
+
**kwargs,
|
16 |
+
):
|
17 |
+
super().__init__(model_name_or_path, **kwargs)
|
18 |
+
if tokenizer_args is None:
|
19 |
+
tokenizer_args = {}
|
20 |
+
|
21 |
+
# Initialize processor
|
22 |
+
self.processor = AutoProcessor.from_pretrained(
|
23 |
+
model_name_or_path, cache_dir=cache_dir, **tokenizer_args
|
24 |
+
)
|
25 |
+
|
26 |
+
def _load_model(
|
27 |
+
self,
|
28 |
+
model_name_or_path: str,
|
29 |
+
config,
|
30 |
+
cache_dir: str,
|
31 |
+
backend: str,
|
32 |
+
is_peft_model: bool,
|
33 |
+
**model_args,
|
34 |
+
) -> None:
|
35 |
+
self.auto_model = MllamaForConditionalGeneration.from_pretrained(
|
36 |
+
model_name_or_path, torch_dtype=torch.bfloat16, cache_dir=cache_dir, **model_args
|
37 |
+
)
|
38 |
+
|
39 |
+
def forward(
|
40 |
+
self, features: Dict[str, torch.Tensor], **kwargs
|
41 |
+
) -> Dict[str, torch.Tensor]:
|
42 |
+
# Process inputs through the model
|
43 |
+
outputs = self.auto_model(
|
44 |
+
**features,
|
45 |
+
return_dict=True,
|
46 |
+
output_hidden_states=True,
|
47 |
+
**kwargs
|
48 |
+
)
|
49 |
+
|
50 |
+
# Apply last pooling and normalization
|
51 |
+
last_hidden_state = outputs.hidden_states[-1]
|
52 |
+
attention_mask = features["attention_mask"]
|
53 |
+
sentence_embedding = self._last_pooling(last_hidden_state, attention_mask)
|
54 |
+
|
55 |
+
features.update({"sentence_embedding": sentence_embedding})
|
56 |
+
return features
|
57 |
+
|
58 |
+
def _last_pooling(self, last_hidden_state: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
|
59 |
+
"""Apply last token pooling and L2 normalization"""
|
60 |
+
sequence_lengths = attention_mask.sum(dim=1) - 1
|
61 |
+
batch_size = last_hidden_state.shape[0]
|
62 |
+
reps = last_hidden_state[torch.arange(batch_size, device=last_hidden_state.device), sequence_lengths]
|
63 |
+
return torch.nn.functional.normalize(reps, p=2, dim=-1)
|
64 |
+
|
65 |
+
def tokenize(self, texts: List[List[Dict]] | List[str]) -> Dict[str, torch.Tensor]:
|
66 |
+
def process_text_item(item):
|
67 |
+
if isinstance(item, str):
|
68 |
+
return item, []
|
69 |
+
|
70 |
+
text, images = "", []
|
71 |
+
for sub_item in item:
|
72 |
+
if sub_item["type"] == "text":
|
73 |
+
text += sub_item["content"]
|
74 |
+
elif sub_item["type"] in ["image_bytes", "image_path"]:
|
75 |
+
text += "<|image|>"
|
76 |
+
if sub_item["type"] == "image_bytes":
|
77 |
+
img = Image.open(BytesIO(sub_item["content"])).convert("RGB")
|
78 |
+
else:
|
79 |
+
img = Image.open(sub_item["content"]).convert("RGB")
|
80 |
+
images.append(img)
|
81 |
+
else:
|
82 |
+
raise ValueError(f"Unknown data type {sub_item['type']}")
|
83 |
+
return text, images
|
84 |
+
|
85 |
+
all_texts, all_images = [], []
|
86 |
+
for item in texts:
|
87 |
+
text, images = process_text_item(item)
|
88 |
+
all_texts.append(text)
|
89 |
+
all_images.extend(images)
|
90 |
+
|
91 |
+
# Process inputs through the processor
|
92 |
+
if all_images:
|
93 |
+
inputs = self.processor(
|
94 |
+
text=all_texts,
|
95 |
+
images=all_images,
|
96 |
+
padding="longest",
|
97 |
+
truncation=True,
|
98 |
+
max_length=self.max_seq_length,
|
99 |
+
return_tensors="pt"
|
100 |
+
)
|
101 |
+
else:
|
102 |
+
inputs = self.processor(
|
103 |
+
text=all_texts,
|
104 |
+
padding="longest",
|
105 |
+
truncation=True,
|
106 |
+
max_length=self.max_seq_length,
|
107 |
+
return_tensors="pt"
|
108 |
+
)
|
109 |
+
|
110 |
+
return inputs
|