Spaces:
Running
Running
Update models/gpt4o.py
Browse files- models/gpt4o.py +111 -111
models/gpt4o.py
CHANGED
@@ -1,111 +1,111 @@
|
|
1 |
-
# gpt4o_pix2struct_ocr.py
|
2 |
-
|
3 |
-
import os
|
4 |
-
import json
|
5 |
-
import base64
|
6 |
-
from PIL import Image
|
7 |
-
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor
|
8 |
-
import numpy as np
|
9 |
-
|
10 |
-
import openai
|
11 |
-
|
12 |
-
model = "gpt-4o"
|
13 |
-
|
14 |
-
# Load Pix2Struct model + processor (vision-language OCR)
|
15 |
-
processor = Pix2StructProcessor.from_pretrained("google/pix2struct-textcaps-base")
|
16 |
-
pix2struct_model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base")
|
17 |
-
|
18 |
-
|
19 |
-
def load_prompt(prompt_file="
|
20 |
-
with open(prompt_file, "r", encoding="utf-8") as f:
|
21 |
-
return f.read().strip()
|
22 |
-
|
23 |
-
|
24 |
-
def try_extract_json(text):
|
25 |
-
try:
|
26 |
-
return json.loads(text)
|
27 |
-
except json.JSONDecodeError:
|
28 |
-
start = text.find('{')
|
29 |
-
if start == -1:
|
30 |
-
return None
|
31 |
-
brace_count = 0
|
32 |
-
json_candidate = ''
|
33 |
-
for i in range(start, len(text)):
|
34 |
-
if text[i] == '{':
|
35 |
-
brace_count += 1
|
36 |
-
elif text[i] == '}':
|
37 |
-
brace_count -= 1
|
38 |
-
json_candidate += text[i]
|
39 |
-
if brace_count == 0 and json_candidate.strip():
|
40 |
-
break
|
41 |
-
try:
|
42 |
-
return json.loads(json_candidate)
|
43 |
-
except json.JSONDecodeError:
|
44 |
-
return None
|
45 |
-
|
46 |
-
|
47 |
-
def encode_image_base64(image: Image.Image):
|
48 |
-
from io import BytesIO
|
49 |
-
buffer = BytesIO()
|
50 |
-
image.save(buffer, format="JPEG")
|
51 |
-
return base64.b64encode(buffer.getvalue()).decode("utf-8")
|
52 |
-
|
53 |
-
|
54 |
-
def extract_all_text_pix2struct(image: Image.Image):
|
55 |
-
inputs = processor(images=image, return_tensors="pt")
|
56 |
-
predictions = pix2struct_model.generate(**inputs, max_new_tokens=512)
|
57 |
-
output_text = processor.decode(predictions[0], skip_special_tokens=True)
|
58 |
-
return output_text.strip()
|
59 |
-
|
60 |
-
|
61 |
-
# Optional: assign best-matching label from full extracted text using proximity (simplified version)
|
62 |
-
def assign_event_gateway_names_from_ocr(image: Image.Image, json_data, ocr_text):
|
63 |
-
if not ocr_text:
|
64 |
-
return json_data
|
65 |
-
|
66 |
-
# You could use NLP matching or regex in complex cases
|
67 |
-
words = ocr_text.split()
|
68 |
-
|
69 |
-
def guess_name_fallback(obj):
|
70 |
-
if not obj.get("name") or obj["name"].strip() == "":
|
71 |
-
obj["name"] = "(label unknown)" # fallback if matching logic isn't yet implemented
|
72 |
-
|
73 |
-
for evt in json_data.get("events", []):
|
74 |
-
guess_name_fallback(evt)
|
75 |
-
|
76 |
-
for gw in json_data.get("gateways", []):
|
77 |
-
guess_name_fallback(gw)
|
78 |
-
|
79 |
-
return json_data
|
80 |
-
|
81 |
-
|
82 |
-
def run_model(image: Image.Image, api_key: str = None):
|
83 |
-
prompt_text = load_prompt()
|
84 |
-
encoded_image = encode_image_base64(image)
|
85 |
-
|
86 |
-
if not api_key:
|
87 |
-
return {"json": None, "raw": "⚠️ API key is missing. Please provide your OpenAI API key."}
|
88 |
-
|
89 |
-
client = openai.OpenAI(api_key=api_key)
|
90 |
-
response = client.chat.completions.create(
|
91 |
-
model=model,
|
92 |
-
messages=[
|
93 |
-
{
|
94 |
-
"role": "user",
|
95 |
-
"content": [
|
96 |
-
{"type": "text", "text": prompt_text},
|
97 |
-
{"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{encoded_image}"}}
|
98 |
-
]
|
99 |
-
}
|
100 |
-
],
|
101 |
-
max_tokens=5000
|
102 |
-
)
|
103 |
-
|
104 |
-
output_text = response.choices[0].message.content.strip()
|
105 |
-
parsed_json = try_extract_json(output_text)
|
106 |
-
|
107 |
-
# Vision-language OCR assist step (Pix2Struct)
|
108 |
-
full_ocr_text = extract_all_text_pix2struct(image)
|
109 |
-
parsed_json = assign_event_gateway_names_from_ocr(image, parsed_json, full_ocr_text)
|
110 |
-
|
111 |
-
return {"json": parsed_json, "raw": output_text}
|
|
|
1 |
+
# gpt4o_pix2struct_ocr.py
|
2 |
+
|
3 |
+
import os
|
4 |
+
import json
|
5 |
+
import base64
|
6 |
+
from PIL import Image
|
7 |
+
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor
|
8 |
+
import numpy as np
|
9 |
+
|
10 |
+
import openai
|
11 |
+
|
12 |
+
model = "gpt-4o"
|
13 |
+
|
14 |
+
# Load Pix2Struct model + processor (vision-language OCR)
|
15 |
+
processor = Pix2StructProcessor.from_pretrained("google/pix2struct-textcaps-base")
|
16 |
+
pix2struct_model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base")
|
17 |
+
|
18 |
+
|
19 |
+
def load_prompt(prompt_file="prompts/prompt.txt"):
|
20 |
+
with open(prompt_file, "r", encoding="utf-8") as f:
|
21 |
+
return f.read().strip()
|
22 |
+
|
23 |
+
|
24 |
+
def try_extract_json(text):
|
25 |
+
try:
|
26 |
+
return json.loads(text)
|
27 |
+
except json.JSONDecodeError:
|
28 |
+
start = text.find('{')
|
29 |
+
if start == -1:
|
30 |
+
return None
|
31 |
+
brace_count = 0
|
32 |
+
json_candidate = ''
|
33 |
+
for i in range(start, len(text)):
|
34 |
+
if text[i] == '{':
|
35 |
+
brace_count += 1
|
36 |
+
elif text[i] == '}':
|
37 |
+
brace_count -= 1
|
38 |
+
json_candidate += text[i]
|
39 |
+
if brace_count == 0 and json_candidate.strip():
|
40 |
+
break
|
41 |
+
try:
|
42 |
+
return json.loads(json_candidate)
|
43 |
+
except json.JSONDecodeError:
|
44 |
+
return None
|
45 |
+
|
46 |
+
|
47 |
+
def encode_image_base64(image: Image.Image):
|
48 |
+
from io import BytesIO
|
49 |
+
buffer = BytesIO()
|
50 |
+
image.save(buffer, format="JPEG")
|
51 |
+
return base64.b64encode(buffer.getvalue()).decode("utf-8")
|
52 |
+
|
53 |
+
|
54 |
+
def extract_all_text_pix2struct(image: Image.Image):
|
55 |
+
inputs = processor(images=image, return_tensors="pt")
|
56 |
+
predictions = pix2struct_model.generate(**inputs, max_new_tokens=512)
|
57 |
+
output_text = processor.decode(predictions[0], skip_special_tokens=True)
|
58 |
+
return output_text.strip()
|
59 |
+
|
60 |
+
|
61 |
+
# Optional: assign best-matching label from full extracted text using proximity (simplified version)
|
62 |
+
def assign_event_gateway_names_from_ocr(image: Image.Image, json_data, ocr_text):
|
63 |
+
if not ocr_text:
|
64 |
+
return json_data
|
65 |
+
|
66 |
+
# You could use NLP matching or regex in complex cases
|
67 |
+
words = ocr_text.split()
|
68 |
+
|
69 |
+
def guess_name_fallback(obj):
|
70 |
+
if not obj.get("name") or obj["name"].strip() == "":
|
71 |
+
obj["name"] = "(label unknown)" # fallback if matching logic isn't yet implemented
|
72 |
+
|
73 |
+
for evt in json_data.get("events", []):
|
74 |
+
guess_name_fallback(evt)
|
75 |
+
|
76 |
+
for gw in json_data.get("gateways", []):
|
77 |
+
guess_name_fallback(gw)
|
78 |
+
|
79 |
+
return json_data
|
80 |
+
|
81 |
+
|
82 |
+
def run_model(image: Image.Image, api_key: str = None):
|
83 |
+
prompt_text = load_prompt()
|
84 |
+
encoded_image = encode_image_base64(image)
|
85 |
+
|
86 |
+
if not api_key:
|
87 |
+
return {"json": None, "raw": "⚠️ API key is missing. Please provide your OpenAI API key."}
|
88 |
+
|
89 |
+
client = openai.OpenAI(api_key=api_key)
|
90 |
+
response = client.chat.completions.create(
|
91 |
+
model=model,
|
92 |
+
messages=[
|
93 |
+
{
|
94 |
+
"role": "user",
|
95 |
+
"content": [
|
96 |
+
{"type": "text", "text": prompt_text},
|
97 |
+
{"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{encoded_image}"}}
|
98 |
+
]
|
99 |
+
}
|
100 |
+
],
|
101 |
+
max_tokens=5000
|
102 |
+
)
|
103 |
+
|
104 |
+
output_text = response.choices[0].message.content.strip()
|
105 |
+
parsed_json = try_extract_json(output_text)
|
106 |
+
|
107 |
+
# Vision-language OCR assist step (Pix2Struct)
|
108 |
+
full_ocr_text = extract_all_text_pix2struct(image)
|
109 |
+
parsed_json = assign_event_gateway_names_from_ocr(image, parsed_json, full_ocr_text)
|
110 |
+
|
111 |
+
return {"json": parsed_json, "raw": output_text}
|