Spaces:
Running
Running
Commit
·
fd0aa67
1
Parent(s):
7302c8f
add: FigureAnnotatorFromPageImage.extract_structured_output
Browse files
medrag_multi_modal/assistant/figure_annotation.py
CHANGED
@@ -4,20 +4,31 @@ from typing import Union
|
|
4 |
import cv2
|
5 |
import weave
|
6 |
from PIL import Image
|
|
|
7 |
from rich.progress import track
|
8 |
|
9 |
from ..utils import get_wandb_artifact, read_jsonl_file
|
10 |
from .llm_client import LLMClient
|
11 |
|
12 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
class FigureAnnotatorFromPageImage(weave.Model):
|
14 |
-
|
|
|
15 |
|
16 |
@weave.op()
|
17 |
def annotate_figures(
|
18 |
self, page_image: Image.Image
|
19 |
) -> dict[str, Union[Image.Image, str]]:
|
20 |
-
annotation = self.
|
21 |
system_prompt="""
|
22 |
You are an expert in the domain of scientific textbooks, especially medical texts.
|
23 |
You are presented with a page from a scientific textbook from the domain of biology, specifically anatomy.
|
@@ -43,16 +54,27 @@ Here are some clues you need to follow:
|
|
43 |
)
|
44 |
return {"page_image": page_image, "annotations": annotation}
|
45 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
@weave.op()
|
47 |
def predict(self, image_artifact_address: str):
|
48 |
artifact_dir = get_wandb_artifact(image_artifact_address, "dataset")
|
49 |
metadata = read_jsonl_file(os.path.join(artifact_dir, "metadata.jsonl"))
|
50 |
annotations = []
|
51 |
for item in track(metadata, description="Annotating images:"):
|
52 |
-
|
53 |
-
|
54 |
-
)
|
55 |
page_image = cv2.cvtColor(page_image, cv2.COLOR_BGR2RGB)
|
56 |
page_image = Image.fromarray(page_image)
|
57 |
-
|
|
|
|
|
|
|
|
|
58 |
return annotations
|
|
|
4 |
import cv2
|
5 |
import weave
|
6 |
from PIL import Image
|
7 |
+
from pydantic import BaseModel
|
8 |
from rich.progress import track
|
9 |
|
10 |
from ..utils import get_wandb_artifact, read_jsonl_file
|
11 |
from .llm_client import LLMClient
|
12 |
|
13 |
|
14 |
+
class FigureAnnotation(BaseModel):
|
15 |
+
figure_id: str
|
16 |
+
figure_description: str
|
17 |
+
|
18 |
+
|
19 |
+
class FigureAnnotations(BaseModel):
|
20 |
+
annotations: list[FigureAnnotation]
|
21 |
+
|
22 |
+
|
23 |
class FigureAnnotatorFromPageImage(weave.Model):
|
24 |
+
figure_extraction_llm_client: LLMClient
|
25 |
+
structured_output_llm_client: LLMClient
|
26 |
|
27 |
@weave.op()
|
28 |
def annotate_figures(
|
29 |
self, page_image: Image.Image
|
30 |
) -> dict[str, Union[Image.Image, str]]:
|
31 |
+
annotation = self.figure_extraction_llm_client.predict(
|
32 |
system_prompt="""
|
33 |
You are an expert in the domain of scientific textbooks, especially medical texts.
|
34 |
You are presented with a page from a scientific textbook from the domain of biology, specifically anatomy.
|
|
|
54 |
)
|
55 |
return {"page_image": page_image, "annotations": annotation}
|
56 |
|
57 |
+
@weave.op
|
58 |
+
def extract_structured_output(self, annotations: str) -> FigureAnnotations:
|
59 |
+
return self.structured_output_llm_client.predict(
|
60 |
+
system_prompt="You are suppossed to extract a list of figure annotations consisting of figure IDs and corresponding figure descriptions.",
|
61 |
+
user_prompt=[annotations],
|
62 |
+
schema=FigureAnnotations,
|
63 |
+
)
|
64 |
+
|
65 |
@weave.op()
|
66 |
def predict(self, image_artifact_address: str):
|
67 |
artifact_dir = get_wandb_artifact(image_artifact_address, "dataset")
|
68 |
metadata = read_jsonl_file(os.path.join(artifact_dir, "metadata.jsonl"))
|
69 |
annotations = []
|
70 |
for item in track(metadata, description="Annotating images:"):
|
71 |
+
page_image_file = os.path.join(artifact_dir, f"page{item['page_idx']}.png")
|
72 |
+
page_image = cv2.imread(page_image_file)
|
|
|
73 |
page_image = cv2.cvtColor(page_image, cv2.COLOR_BGR2RGB)
|
74 |
page_image = Image.fromarray(page_image)
|
75 |
+
figure_extracted_annotations = self.annotate_figures(page_image=page_image)
|
76 |
+
figure_extracted_annotations["annotations"] = self.extract_structured_output(
|
77 |
+
figure_extracted_annotations["annotations"]
|
78 |
+
).model_dump()
|
79 |
+
annotations.append(figure_extracted_annotations)
|
80 |
return annotations
|
medrag_multi_modal/assistant/llm_client.py
CHANGED
@@ -12,6 +12,7 @@ from ..utils import base64_encode_image
|
|
12 |
class ClientType(str, Enum):
|
13 |
GEMINI = "gemini"
|
14 |
MISTRAL = "mistral"
|
|
|
15 |
|
16 |
|
17 |
GOOGLE_MODELS = [
|
@@ -54,6 +55,8 @@ MISTRAL_MODELS = [
|
|
54 |
"open-mixtral-8x22b",
|
55 |
]
|
56 |
|
|
|
|
|
57 |
|
58 |
class LLMClient(weave.Model):
|
59 |
model_name: str
|
@@ -65,6 +68,8 @@ class LLMClient(weave.Model):
|
|
65 |
client_type = ClientType.GEMINI
|
66 |
elif model_name in MISTRAL_MODELS:
|
67 |
client_type = ClientType.MISTRAL
|
|
|
|
|
68 |
else:
|
69 |
raise ValueError(f"Invalid model name: {model_name}")
|
70 |
super().__init__(model_name=model_name, client_type=client_type)
|
@@ -139,6 +144,51 @@ class LLMClient(weave.Model):
|
|
139 |
)
|
140 |
return response.choices[0].message.content
|
141 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
@weave.op()
|
143 |
def predict(
|
144 |
self,
|
@@ -150,5 +200,7 @@ class LLMClient(weave.Model):
|
|
150 |
return self.execute_gemini_sdk(user_prompt, system_prompt, schema)
|
151 |
elif self.client_type == ClientType.MISTRAL:
|
152 |
return self.execute_mistral_sdk(user_prompt, system_prompt, schema)
|
|
|
|
|
153 |
else:
|
154 |
raise ValueError(f"Invalid client type: {self.client_type}")
|
|
|
12 |
class ClientType(str, Enum):
|
13 |
GEMINI = "gemini"
|
14 |
MISTRAL = "mistral"
|
15 |
+
OPENAI = "openai"
|
16 |
|
17 |
|
18 |
GOOGLE_MODELS = [
|
|
|
55 |
"open-mixtral-8x22b",
|
56 |
]
|
57 |
|
58 |
+
OPENAI_MODELS = ["gpt-4o", "gpt-4o-2024-08-06", "gpt-4o-mini", "gpt-4o-mini-2024-07-18"]
|
59 |
+
|
60 |
|
61 |
class LLMClient(weave.Model):
|
62 |
model_name: str
|
|
|
68 |
client_type = ClientType.GEMINI
|
69 |
elif model_name in MISTRAL_MODELS:
|
70 |
client_type = ClientType.MISTRAL
|
71 |
+
elif model_name in OPENAI_MODELS:
|
72 |
+
client_type = ClientType.OPENAI
|
73 |
else:
|
74 |
raise ValueError(f"Invalid model name: {model_name}")
|
75 |
super().__init__(model_name=model_name, client_type=client_type)
|
|
|
144 |
)
|
145 |
return response.choices[0].message.content
|
146 |
|
147 |
+
@weave.op()
|
148 |
+
def execute_openai_sdk(
|
149 |
+
self,
|
150 |
+
user_prompt: Union[str, list[str]],
|
151 |
+
system_prompt: Optional[Union[str, list[str]]] = None,
|
152 |
+
schema: Optional[Any] = None,
|
153 |
+
) -> Union[str, Any]:
|
154 |
+
from openai import OpenAI
|
155 |
+
|
156 |
+
system_prompt = (
|
157 |
+
[system_prompt] if isinstance(system_prompt, str) else system_prompt
|
158 |
+
)
|
159 |
+
user_prompt = [user_prompt] if isinstance(user_prompt, str) else user_prompt
|
160 |
+
|
161 |
+
system_messages = [
|
162 |
+
{"role": "system", "content": prompt} for prompt in system_prompt
|
163 |
+
]
|
164 |
+
user_messages = []
|
165 |
+
for prompt in user_prompt:
|
166 |
+
if isinstance(prompt, Image.Image):
|
167 |
+
user_messages.append(
|
168 |
+
{
|
169 |
+
"type": "image_url",
|
170 |
+
"image_url": {
|
171 |
+
"url": base64_encode_image(prompt, "image/png"),
|
172 |
+
},
|
173 |
+
},
|
174 |
+
)
|
175 |
+
else:
|
176 |
+
user_messages.append({"type": "text", "text": prompt})
|
177 |
+
messages = system_messages + [{"role": "user", "content": user_messages}]
|
178 |
+
|
179 |
+
client = OpenAI(api_key=os.environ.get("OPENAI_API_KEY"))
|
180 |
+
|
181 |
+
if schema is None:
|
182 |
+
completion = client.chat.completions.create(
|
183 |
+
model=self.model_name, messages=messages
|
184 |
+
)
|
185 |
+
return completion.choices[0].message.content
|
186 |
+
|
187 |
+
completion = weave.op()(client.beta.chat.completions.parse)(
|
188 |
+
model=self.model_name, messages=messages, response_format=schema
|
189 |
+
)
|
190 |
+
return completion.choices[0].message.parsed
|
191 |
+
|
192 |
@weave.op()
|
193 |
def predict(
|
194 |
self,
|
|
|
200 |
return self.execute_gemini_sdk(user_prompt, system_prompt, schema)
|
201 |
elif self.client_type == ClientType.MISTRAL:
|
202 |
return self.execute_mistral_sdk(user_prompt, system_prompt, schema)
|
203 |
+
elif self.client_type == ClientType.OPENAI:
|
204 |
+
return self.execute_openai_sdk(user_prompt, system_prompt, schema)
|
205 |
else:
|
206 |
raise ValueError(f"Invalid client type: {self.client_type}")
|
pyproject.toml
CHANGED
@@ -43,6 +43,7 @@ dependencies = [
|
|
43 |
"instructor>=1.6.3",
|
44 |
"jsonlines>=4.0.0",
|
45 |
"opencv-python>=4.10.0.84",
|
|
|
46 |
]
|
47 |
|
48 |
[project.optional-dependencies]
|
@@ -71,6 +72,7 @@ core = [
|
|
71 |
"instructor>=1.6.3",
|
72 |
"jsonlines>=4.0.0",
|
73 |
"opencv-python>=4.10.0.84",
|
|
|
74 |
]
|
75 |
|
76 |
dev = ["pytest>=8.3.3", "isort>=5.13.2", "black>=24.10.0", "ruff>=0.6.9"]
|
|
|
43 |
"instructor>=1.6.3",
|
44 |
"jsonlines>=4.0.0",
|
45 |
"opencv-python>=4.10.0.84",
|
46 |
+
"openai>=1.52.2",
|
47 |
]
|
48 |
|
49 |
[project.optional-dependencies]
|
|
|
72 |
"instructor>=1.6.3",
|
73 |
"jsonlines>=4.0.0",
|
74 |
"opencv-python>=4.10.0.84",
|
75 |
+
"openai>=1.52.2",
|
76 |
]
|
77 |
|
78 |
dev = ["pytest>=8.3.3", "isort>=5.13.2", "black>=24.10.0", "ruff>=0.6.9"]
|