Spaces:
Running
Running
Update models/qwen.py
Browse files- models/qwen.py +48 -33
models/qwen.py
CHANGED
@@ -3,31 +3,21 @@ import json
|
|
3 |
from PIL import Image
|
4 |
import torch
|
5 |
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
|
6 |
-
from qwen_vl_utils import process_vision_info
|
7 |
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor
|
|
|
8 |
|
9 |
-
#
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
#attn_implementation="flash_attention_2"
|
15 |
-
)
|
16 |
-
|
17 |
-
min_pixels = 256 * 28 * 28
|
18 |
-
max_pixels = 1080 * 28 * 28
|
19 |
-
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct", min_pixels=min_pixels, max_pixels=max_pixels)
|
20 |
|
21 |
-
# Initialize Pix2Struct OCR model
|
22 |
-
ocr_processor = Pix2StructProcessor.from_pretrained("google/pix2struct-textcaps-base")
|
23 |
-
ocr_model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base")
|
24 |
|
25 |
-
# Load prompt
|
26 |
def load_prompt():
|
27 |
-
with open("prompts/prompt.txt", "r") as f:
|
28 |
return f.read()
|
29 |
|
30 |
-
|
31 |
def try_extract_json(text):
|
32 |
try:
|
33 |
return json.loads(text)
|
@@ -50,20 +40,25 @@ def try_extract_json(text):
|
|
50 |
except json.JSONDecodeError:
|
51 |
return text
|
52 |
|
53 |
-
|
54 |
def extract_all_text_pix2struct(image: Image.Image):
|
55 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
predictions = ocr_model.generate(**inputs, max_new_tokens=512)
|
57 |
-
|
58 |
-
|
59 |
|
60 |
-
# Assign event/gateway names from OCR text
|
61 |
def assign_event_gateway_names_from_ocr(json_data: dict, ocr_text: str):
|
62 |
if not ocr_text or not json_data:
|
63 |
return json_data
|
64 |
|
65 |
-
lines = [line.strip() for line in ocr_text.split('\n') if line.strip()]
|
66 |
-
|
67 |
def assign_best_guess(obj):
|
68 |
if not obj.get("name") or obj["name"].strip() == "":
|
69 |
obj["name"] = "(label unknown)"
|
@@ -76,9 +71,29 @@ def assign_event_gateway_names_from_ocr(json_data: dict, ocr_text: str):
|
|
76 |
|
77 |
return json_data
|
78 |
|
79 |
-
|
80 |
def run_model(image: Image.Image):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
prompt = load_prompt()
|
|
|
82 |
messages = [
|
83 |
{
|
84 |
"role": "user",
|
@@ -89,21 +104,21 @@ def run_model(image: Image.Image):
|
|
89 |
}
|
90 |
]
|
91 |
|
92 |
-
text =
|
93 |
image_inputs, video_inputs = process_vision_info(messages)
|
94 |
|
95 |
-
inputs =
|
96 |
text=[text],
|
97 |
images=image_inputs,
|
98 |
videos=video_inputs,
|
99 |
padding=True,
|
100 |
return_tensors="pt"
|
101 |
-
).to(
|
102 |
|
103 |
-
generated_ids =
|
104 |
generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
|
105 |
|
106 |
-
output_text =
|
107 |
generated_ids_trimmed,
|
108 |
skip_special_tokens=True,
|
109 |
clean_up_tokenization_spaces=False
|
@@ -111,11 +126,11 @@ def run_model(image: Image.Image):
|
|
111 |
|
112 |
parsed_json = try_extract_json(output_text)
|
113 |
|
114 |
-
#
|
115 |
ocr_text = extract_all_text_pix2struct(image)
|
116 |
parsed_json = assign_event_gateway_names_from_ocr(parsed_json, ocr_text)
|
117 |
|
118 |
return {
|
119 |
"json": parsed_json,
|
120 |
"raw": output_text
|
121 |
-
}
|
|
|
3 |
from PIL import Image
|
4 |
import torch
|
5 |
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
|
|
|
6 |
from transformers import Pix2StructForConditionalGeneration, Pix2StructProcessor
|
7 |
+
from qwen_vl_utils import process_vision_info
|
8 |
|
9 |
+
# Globals (lazy-loaded at runtime)
|
10 |
+
qwen_model = None
|
11 |
+
qwen_processor = None
|
12 |
+
ocr_model = None
|
13 |
+
ocr_processor = None
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
|
|
|
|
|
|
15 |
|
|
|
16 |
def load_prompt():
|
17 |
+
with open("prompts/prompt.txt", "r", encoding="utf-8") as f:
|
18 |
return f.read()
|
19 |
|
20 |
+
|
21 |
def try_extract_json(text):
|
22 |
try:
|
23 |
return json.loads(text)
|
|
|
40 |
except json.JSONDecodeError:
|
41 |
return text
|
42 |
|
43 |
+
|
44 |
def extract_all_text_pix2struct(image: Image.Image):
|
45 |
+
global ocr_model, ocr_processor
|
46 |
+
|
47 |
+
if ocr_model is None or ocr_processor is None:
|
48 |
+
ocr_processor = Pix2StructProcessor.from_pretrained("google/pix2struct-textcaps-base")
|
49 |
+
ocr_model = Pix2StructForConditionalGeneration.from_pretrained("google/pix2struct-textcaps-base")
|
50 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
51 |
+
ocr_model = ocr_model.to(device)
|
52 |
+
|
53 |
+
inputs = ocr_processor(images=image, return_tensors="pt").to(ocr_model.device)
|
54 |
predictions = ocr_model.generate(**inputs, max_new_tokens=512)
|
55 |
+
return ocr_processor.decode(predictions[0], skip_special_tokens=True).strip()
|
56 |
+
|
57 |
|
|
|
58 |
def assign_event_gateway_names_from_ocr(json_data: dict, ocr_text: str):
|
59 |
if not ocr_text or not json_data:
|
60 |
return json_data
|
61 |
|
|
|
|
|
62 |
def assign_best_guess(obj):
|
63 |
if not obj.get("name") or obj["name"].strip() == "":
|
64 |
obj["name"] = "(label unknown)"
|
|
|
71 |
|
72 |
return json_data
|
73 |
|
74 |
+
|
75 |
def run_model(image: Image.Image):
|
76 |
+
global qwen_model, qwen_processor
|
77 |
+
|
78 |
+
if qwen_model is None or qwen_processor is None:
|
79 |
+
qwen_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
80 |
+
"Qwen/Qwen2.5-VL-7B-Instruct",
|
81 |
+
torch_dtype=torch.bfloat16,
|
82 |
+
device_map="auto"
|
83 |
+
# You can enable flash attention here if needed:
|
84 |
+
# attn_implementation="flash_attention_2"
|
85 |
+
)
|
86 |
+
|
87 |
+
min_pixels = 256 * 28 * 28
|
88 |
+
max_pixels = 1080 * 28 * 28
|
89 |
+
qwen_processor = AutoProcessor.from_pretrained(
|
90 |
+
"Qwen/Qwen2.5-VL-7B-Instruct",
|
91 |
+
min_pixels=min_pixels,
|
92 |
+
max_pixels=max_pixels
|
93 |
+
)
|
94 |
+
|
95 |
prompt = load_prompt()
|
96 |
+
|
97 |
messages = [
|
98 |
{
|
99 |
"role": "user",
|
|
|
104 |
}
|
105 |
]
|
106 |
|
107 |
+
text = qwen_processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
108 |
image_inputs, video_inputs = process_vision_info(messages)
|
109 |
|
110 |
+
inputs = qwen_processor(
|
111 |
text=[text],
|
112 |
images=image_inputs,
|
113 |
videos=video_inputs,
|
114 |
padding=True,
|
115 |
return_tensors="pt"
|
116 |
+
).to(qwen_model.device)
|
117 |
|
118 |
+
generated_ids = qwen_model.generate(**inputs, max_new_tokens=5000)
|
119 |
generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
|
120 |
|
121 |
+
output_text = qwen_processor.batch_decode(
|
122 |
generated_ids_trimmed,
|
123 |
skip_special_tokens=True,
|
124 |
clean_up_tokenization_spaces=False
|
|
|
126 |
|
127 |
parsed_json = try_extract_json(output_text)
|
128 |
|
129 |
+
# OCR post-processing
|
130 |
ocr_text = extract_all_text_pix2struct(image)
|
131 |
parsed_json = assign_event_gateway_names_from_ocr(parsed_json, ocr_text)
|
132 |
|
133 |
return {
|
134 |
"json": parsed_json,
|
135 |
"raw": output_text
|
136 |
+
}
|