Spaces:
Running
Running
Update models/aya_vision.py
Browse files- models/aya_vision.py +37 -26
models/aya_vision.py
CHANGED
@@ -6,37 +6,32 @@ import torch
|
|
6 |
from transformers import AutoProcessor, AutoModelForImageTextToText
|
7 |
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor
|
8 |
|
9 |
-
# Set Hugging Face Token
|
10 |
hf_token = os.getenv("HF_TOKEN")
|
11 |
|
12 |
-
#
|
13 |
-
|
14 |
-
|
15 |
-
model = AutoModelForImageTextToText.from_pretrained(
|
16 |
-
model_id, device_map="auto", torch_dtype=torch.float16
|
17 |
-
)
|
18 |
|
19 |
-
|
20 |
-
ocr_processor =
|
21 |
-
ocr_model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base")
|
22 |
|
23 |
# Load prompt
|
24 |
def load_prompt():
|
25 |
with open("prompts/prompt.txt", "r", encoding="utf-8") as f:
|
26 |
return f.read()
|
27 |
|
28 |
-
|
|
|
29 |
def try_extract_json(text):
|
30 |
if not text or not text.strip():
|
31 |
return None
|
32 |
try:
|
33 |
return json.loads(text)
|
34 |
except json.JSONDecodeError:
|
35 |
-
# Try extracting JSON substring by brace balancing
|
36 |
start = text.find('{')
|
37 |
if start == -1:
|
38 |
return None
|
39 |
-
|
40 |
brace_count = 0
|
41 |
json_candidate = ''
|
42 |
for i in range(start, len(text)):
|
@@ -48,26 +43,33 @@ def try_extract_json(text):
|
|
48 |
json_candidate += char
|
49 |
if brace_count == 0:
|
50 |
break
|
51 |
-
|
52 |
try:
|
53 |
return json.loads(json_candidate)
|
54 |
except json.JSONDecodeError:
|
55 |
return None
|
56 |
|
57 |
-
|
|
|
58 |
def extract_all_text_pix2struct(image: Image.Image):
|
59 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
predictions = ocr_model.generate(**inputs, max_new_tokens=512)
|
61 |
output_text = ocr_processor.decode(predictions[0], skip_special_tokens=True)
|
62 |
return output_text.strip()
|
63 |
|
64 |
-
|
|
|
65 |
def assign_event_gateway_names_from_ocr(json_data: dict, ocr_text: str):
|
66 |
if not ocr_text or not json_data:
|
67 |
return json_data
|
68 |
|
69 |
-
lines = [line.strip() for line in ocr_text.split('\n') if line.strip()]
|
70 |
-
|
71 |
def assign_best_guess(obj):
|
72 |
if not obj.get("name") or obj["name"].strip() == "":
|
73 |
obj["name"] = "(label unknown)"
|
@@ -80,8 +82,18 @@ def assign_event_gateway_names_from_ocr(json_data: dict, ocr_text: str):
|
|
80 |
|
81 |
return json_data
|
82 |
|
83 |
-
|
|
|
84 |
def run_model(image: Image.Image):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
prompt = load_prompt()
|
86 |
|
87 |
messages = [
|
@@ -94,34 +106,33 @@ def run_model(image: Image.Image):
|
|
94 |
}
|
95 |
]
|
96 |
|
97 |
-
inputs =
|
98 |
messages,
|
99 |
padding=True,
|
100 |
add_generation_prompt=True,
|
101 |
tokenize=True,
|
102 |
return_dict=True,
|
103 |
return_tensors="pt"
|
104 |
-
).to(
|
105 |
|
106 |
-
gen_tokens =
|
107 |
**inputs,
|
108 |
max_new_tokens=5000,
|
109 |
do_sample=True,
|
110 |
temperature=0.3,
|
111 |
)
|
112 |
|
113 |
-
output_text =
|
114 |
gen_tokens[0][inputs.input_ids.shape[1]:],
|
115 |
skip_special_tokens=True
|
116 |
)
|
117 |
|
118 |
parsed_json = try_extract_json(output_text)
|
119 |
|
120 |
-
#
|
121 |
ocr_text = extract_all_text_pix2struct(image)
|
122 |
parsed_json = assign_event_gateway_names_from_ocr(parsed_json, ocr_text)
|
123 |
|
124 |
-
# Return both parsed and raw
|
125 |
return {
|
126 |
"json": parsed_json,
|
127 |
"raw": output_text
|
|
|
6 |
from transformers import AutoProcessor, AutoModelForImageTextToText
|
7 |
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor
|
8 |
|
9 |
+
# Set Hugging Face Token from env
|
10 |
hf_token = os.getenv("HF_TOKEN")
|
11 |
|
12 |
+
# Lazy-load model objects
|
13 |
+
aya_model = None
|
14 |
+
aya_processor = None
|
|
|
|
|
|
|
15 |
|
16 |
+
ocr_model = None
|
17 |
+
ocr_processor = None
|
|
|
18 |
|
19 |
# Load prompt
|
20 |
def load_prompt():
|
21 |
with open("prompts/prompt.txt", "r", encoding="utf-8") as f:
|
22 |
return f.read()
|
23 |
|
24 |
+
|
25 |
+
# Try extracting JSON from text
|
26 |
def try_extract_json(text):
|
27 |
if not text or not text.strip():
|
28 |
return None
|
29 |
try:
|
30 |
return json.loads(text)
|
31 |
except json.JSONDecodeError:
|
|
|
32 |
start = text.find('{')
|
33 |
if start == -1:
|
34 |
return None
|
|
|
35 |
brace_count = 0
|
36 |
json_candidate = ''
|
37 |
for i in range(start, len(text)):
|
|
|
43 |
json_candidate += char
|
44 |
if brace_count == 0:
|
45 |
break
|
|
|
46 |
try:
|
47 |
return json.loads(json_candidate)
|
48 |
except json.JSONDecodeError:
|
49 |
return None
|
50 |
|
51 |
+
|
52 |
+
# OCR text from Pix2Struct
|
53 |
def extract_all_text_pix2struct(image: Image.Image):
|
54 |
+
global ocr_processor, ocr_model
|
55 |
+
|
56 |
+
if ocr_processor is None or ocr_model is None:
|
57 |
+
ocr_processor = Pix2StructProcessor.from_pretrained("google/pix2struct-textcaps-base")
|
58 |
+
ocr_model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base")
|
59 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
60 |
+
ocr_model = ocr_model.to(device)
|
61 |
+
|
62 |
+
inputs = ocr_processor(images=image, return_tensors="pt").to(ocr_model.device)
|
63 |
predictions = ocr_model.generate(**inputs, max_new_tokens=512)
|
64 |
output_text = ocr_processor.decode(predictions[0], skip_special_tokens=True)
|
65 |
return output_text.strip()
|
66 |
|
67 |
+
|
68 |
+
# Add fallback names if missing
|
69 |
def assign_event_gateway_names_from_ocr(json_data: dict, ocr_text: str):
|
70 |
if not ocr_text or not json_data:
|
71 |
return json_data
|
72 |
|
|
|
|
|
73 |
def assign_best_guess(obj):
|
74 |
if not obj.get("name") or obj["name"].strip() == "":
|
75 |
obj["name"] = "(label unknown)"
|
|
|
82 |
|
83 |
return json_data
|
84 |
|
85 |
+
|
86 |
+
# Main inference function
|
87 |
def run_model(image: Image.Image):
|
88 |
+
global aya_model, aya_processor
|
89 |
+
|
90 |
+
if aya_model is None or aya_processor is None:
|
91 |
+
model_id = "CohereForAI/aya-vision-8b"
|
92 |
+
aya_processor = AutoProcessor.from_pretrained(model_id)
|
93 |
+
aya_model = AutoModelForImageTextToText.from_pretrained(
|
94 |
+
model_id, device_map="auto", torch_dtype=torch.float16
|
95 |
+
)
|
96 |
+
|
97 |
prompt = load_prompt()
|
98 |
|
99 |
messages = [
|
|
|
106 |
}
|
107 |
]
|
108 |
|
109 |
+
inputs = aya_processor.apply_chat_template(
|
110 |
messages,
|
111 |
padding=True,
|
112 |
add_generation_prompt=True,
|
113 |
tokenize=True,
|
114 |
return_dict=True,
|
115 |
return_tensors="pt"
|
116 |
+
).to(aya_model.device)
|
117 |
|
118 |
+
gen_tokens = aya_model.generate(
|
119 |
**inputs,
|
120 |
max_new_tokens=5000,
|
121 |
do_sample=True,
|
122 |
temperature=0.3,
|
123 |
)
|
124 |
|
125 |
+
output_text = aya_processor.tokenizer.decode(
|
126 |
gen_tokens[0][inputs.input_ids.shape[1]:],
|
127 |
skip_special_tokens=True
|
128 |
)
|
129 |
|
130 |
parsed_json = try_extract_json(output_text)
|
131 |
|
132 |
+
# OCR enhancement
|
133 |
ocr_text = extract_all_text_pix2struct(image)
|
134 |
parsed_json = assign_event_gateway_names_from_ocr(parsed_json, ocr_text)
|
135 |
|
|
|
136 |
return {
|
137 |
"json": parsed_json,
|
138 |
"raw": output_text
|