Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -11,107 +11,45 @@ import traceback
|
|
11 |
import warnings
|
12 |
import sys
|
13 |
|
14 |
-
# Suppress
|
15 |
warnings.filterwarnings("ignore", category=FutureWarning)
|
16 |
warnings.filterwarnings("ignore", message=".*_supports_sdpa.*")
|
17 |
|
18 |
-
#
|
19 |
-
def
|
20 |
-
"""
|
21 |
-
import importlib.util
|
22 |
-
import types
|
23 |
-
|
24 |
-
# Create a custom import hook
|
25 |
-
class Florence2ImportHook:
|
26 |
-
def find_spec(self, fullname, path, target=None):
|
27 |
-
if "florence2" in fullname.lower() or "modeling_florence2" in fullname:
|
28 |
-
return importlib.util.spec_from_loader(fullname, Florence2Loader())
|
29 |
-
return None
|
30 |
-
|
31 |
-
class Florence2Loader:
|
32 |
-
def create_module(self, spec):
|
33 |
-
return None
|
34 |
-
|
35 |
-
def exec_module(self, module):
|
36 |
-
# Load the original module
|
37 |
-
import importlib.machinery
|
38 |
-
import importlib.util
|
39 |
-
|
40 |
-
# Find the actual florence2 module
|
41 |
-
for path in sys.path:
|
42 |
-
florence_path = os.path.join(path, "modeling_florence2.py")
|
43 |
-
if os.path.exists(florence_path):
|
44 |
-
spec = importlib.util.spec_from_file_location("modeling_florence2", florence_path)
|
45 |
-
if spec and spec.loader:
|
46 |
-
spec.loader.exec_module(module)
|
47 |
-
|
48 |
-
# Patch the module after loading
|
49 |
-
if hasattr(module, 'Florence2ForConditionalGeneration'):
|
50 |
-
original_init = module.Florence2ForConditionalGeneration.__init__
|
51 |
-
|
52 |
-
def patched_init(self, config):
|
53 |
-
# Add the missing attribute before calling super().__init__
|
54 |
-
self._supports_sdpa = False
|
55 |
-
original_init(self, config)
|
56 |
-
|
57 |
-
module.Florence2ForConditionalGeneration.__init__ = patched_init
|
58 |
-
module.Florence2ForConditionalGeneration._supports_sdpa = False
|
59 |
-
break
|
60 |
-
|
61 |
-
# Install the import hook
|
62 |
-
hook = Florence2ImportHook()
|
63 |
-
sys.meta_path.insert(0, hook)
|
64 |
-
|
65 |
-
# Apply the fix before any model imports
|
66 |
-
try:
|
67 |
-
fix_florence2_import()
|
68 |
-
except Exception as e:
|
69 |
-
print(f"Warning: Could not apply import hook: {e}")
|
70 |
-
|
71 |
-
# Alternative fix: Monkey-patch transformers before importing utils
|
72 |
-
def monkey_patch_transformers():
|
73 |
-
"""Monkey patch transformers to handle _supports_sdpa"""
|
74 |
try:
|
75 |
import transformers.modeling_utils as modeling_utils
|
76 |
|
|
|
77 |
original_check = modeling_utils.PreTrainedModel._check_and_adjust_attn_implementation
|
78 |
|
79 |
def patched_check(self, *args, **kwargs):
|
80 |
-
#
|
81 |
if not hasattr(self, '_supports_sdpa'):
|
82 |
-
self
|
|
|
83 |
try:
|
84 |
return original_check(self, *args, **kwargs)
|
85 |
except AttributeError as e:
|
86 |
if '_supports_sdpa' in str(e):
|
87 |
-
# Return
|
88 |
return "eager"
|
89 |
raise
|
90 |
|
91 |
modeling_utils.PreTrainedModel._check_and_adjust_attn_implementation = patched_check
|
92 |
-
|
93 |
-
# Also patch the getter
|
94 |
-
original_getattr = modeling_utils.PreTrainedModel.__getattribute__
|
95 |
-
|
96 |
-
def patched_getattr(self, name):
|
97 |
-
if name == '_supports_sdpa' and not hasattr(self, '_supports_sdpa'):
|
98 |
-
return False
|
99 |
-
return original_getattr(self, name)
|
100 |
-
|
101 |
-
modeling_utils.PreTrainedModel.__getattribute__ = patched_getattr
|
102 |
-
|
103 |
-
print("Successfully patched transformers for Florence2 compatibility")
|
104 |
|
105 |
except Exception as e:
|
106 |
print(f"Warning: Could not patch transformers: {e}")
|
107 |
|
108 |
-
# Apply the
|
109 |
-
|
110 |
|
111 |
-
# Now import the utils
|
112 |
-
from util.utils import check_ocr_box, get_yolo_model, get_som_labeled_img
|
113 |
|
114 |
-
# Download repository
|
115 |
repo_id = "microsoft/OmniParser-v2.0"
|
116 |
local_dir = "weights"
|
117 |
|
@@ -121,75 +59,105 @@ if not os.path.exists(local_dir):
|
|
121 |
else:
|
122 |
print(f"Weights already exist at: {local_dir}")
|
123 |
|
124 |
-
# Custom function to load caption model
|
125 |
def load_caption_model_safe(model_name="florence2", model_name_or_path="weights/icon_caption"):
|
126 |
-
"""Safely load caption model
|
127 |
|
128 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
129 |
|
|
|
130 |
try:
|
131 |
-
# Method 1: Try the original function with patching
|
132 |
-
from util.utils import get_caption_model_processor
|
133 |
return get_caption_model_processor(model_name, model_name_or_path)
|
134 |
-
except
|
135 |
-
|
136 |
-
print(f"SDPA error detected, trying alternative loading method...")
|
137 |
-
else:
|
138 |
-
raise
|
139 |
|
140 |
-
# Method 2: Load
|
141 |
try:
|
142 |
from transformers import AutoProcessor, AutoModelForCausalLM
|
143 |
|
144 |
-
print(f"Loading caption model from {model_name_or_path}
|
145 |
|
146 |
-
# Load processor
|
147 |
processor = AutoProcessor.from_pretrained(
|
148 |
model_name_or_path,
|
149 |
-
trust_remote_code=True
|
150 |
-
revision="main"
|
151 |
)
|
152 |
|
153 |
-
#
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
for config in configs_to_try:
|
162 |
-
try:
|
163 |
-
model = AutoModelForCausalLM.from_pretrained(
|
164 |
-
model_name_or_path,
|
165 |
-
trust_remote_code=True,
|
166 |
-
device_map="auto" if torch.cuda.is_available() else None,
|
167 |
-
**config
|
168 |
-
)
|
169 |
-
|
170 |
-
# Ensure the attribute exists
|
171 |
-
if not hasattr(model, '_supports_sdpa'):
|
172 |
-
model._supports_sdpa = False
|
173 |
-
|
174 |
-
print(f"Model loaded successfully with config: {config}")
|
175 |
-
break
|
176 |
-
|
177 |
-
except Exception as e:
|
178 |
-
print(f"Failed with config {config}: {e}")
|
179 |
-
continue
|
180 |
|
181 |
-
|
182 |
-
|
|
|
183 |
|
184 |
-
|
185 |
-
if device.type == 'cuda' and not next(model.parameters()).is_cuda:
|
186 |
model = model.to(device)
|
187 |
|
|
|
188 |
return {'model': model, 'processor': processor}
|
189 |
|
190 |
except Exception as e:
|
191 |
-
print(f"
|
192 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
193 |
|
194 |
# Load models
|
195 |
try:
|
@@ -205,9 +173,9 @@ except Exception as e:
|
|
205 |
print(f"Critical error loading models: {e}")
|
206 |
print(traceback.format_exc())
|
207 |
caption_model_processor = None
|
208 |
-
|
209 |
|
210 |
-
#
|
211 |
MARKDOWN = """
|
212 |
# OmniParser V2 Pro🔥
|
213 |
|
@@ -220,7 +188,6 @@ MARKDOWN = """
|
|
220 |
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
221 |
print(f"Using device: {DEVICE}")
|
222 |
|
223 |
-
# Custom CSS for UI enhancement
|
224 |
custom_css = """
|
225 |
body { background-color: #f0f2f5; }
|
226 |
.gradio-container { font-family: 'Segoe UI', sans-serif; max-width: 1400px; margin: auto; }
|
@@ -230,8 +197,6 @@ button:hover { transform: translateY(-2px); box-shadow: 0 4px 12px rgba(0,0,0,0.
|
|
230 |
.output-image { border: 2px solid #e1e4e8; border-radius: 8px; }
|
231 |
#input_image { border: 2px dashed #4a90e2; border-radius: 8px; }
|
232 |
#input_image:hover { border-color: #2c5aa0; }
|
233 |
-
.gr-box { border-radius: 8px; }
|
234 |
-
.gr-padded { padding: 16px; }
|
235 |
"""
|
236 |
|
237 |
@spaces.GPU
|
@@ -243,22 +208,19 @@ def process(
|
|
243 |
use_paddleocr,
|
244 |
imgsz
|
245 |
) -> tuple:
|
246 |
-
"""Process image with error handling
|
247 |
|
248 |
-
# Input validation
|
249 |
if image_input is None:
|
250 |
return None, "⚠️ Please upload an image for processing."
|
251 |
|
252 |
-
|
253 |
-
|
254 |
-
return None, "⚠️ Caption model not loaded. There was an error during initialization. Please check the logs."
|
255 |
|
256 |
try:
|
257 |
-
|
258 |
-
|
259 |
-
f"iou_threshold={iou_threshold}, use_paddleocr={use_paddleocr}, imgsz={imgsz}")
|
260 |
|
261 |
-
# Calculate overlay ratio
|
262 |
image_width = image_input.size[0]
|
263 |
box_overlay_ratio = max(0.5, min(2.0, image_width / 3200))
|
264 |
|
@@ -269,7 +231,7 @@ def process(
|
|
269 |
'thickness': max(int(3 * box_overlay_ratio), 1),
|
270 |
}
|
271 |
|
272 |
-
#
|
273 |
try:
|
274 |
ocr_bbox_rslt, is_goal_filtered = check_ocr_box(
|
275 |
image_input,
|
@@ -280,42 +242,37 @@ def process(
|
|
280 |
use_paddleocr=use_paddleocr
|
281 |
)
|
282 |
|
283 |
-
# Handle None result from OCR
|
284 |
if ocr_bbox_rslt is None:
|
285 |
-
print("OCR returned None, using empty results")
|
286 |
text, ocr_bbox = [], []
|
287 |
else:
|
288 |
text, ocr_bbox = ocr_bbox_rslt
|
289 |
|
290 |
-
|
291 |
-
if
|
292 |
-
|
293 |
-
if ocr_bbox is None:
|
294 |
-
ocr_bbox = []
|
295 |
-
|
296 |
print(f"OCR found {len(text)} text regions")
|
297 |
|
298 |
except Exception as e:
|
299 |
-
print(f"OCR error: {e}
|
300 |
text, ocr_bbox = [], []
|
301 |
|
302 |
-
#
|
303 |
try:
|
304 |
-
# Ensure
|
305 |
if isinstance(caption_model_processor, dict) and 'model' in caption_model_processor:
|
306 |
model = caption_model_processor['model']
|
307 |
if not hasattr(model, '_supports_sdpa'):
|
308 |
-
model
|
309 |
|
310 |
dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(
|
311 |
image_input,
|
312 |
yolo_model,
|
313 |
BOX_TRESHOLD=box_threshold,
|
314 |
output_coord_in_ratio=True,
|
315 |
-
ocr_bbox=ocr_bbox
|
316 |
draw_bbox_config=draw_bbox_config,
|
317 |
caption_model_processor=caption_model_processor,
|
318 |
-
ocr_text=text
|
319 |
iou_threshold=iou_threshold,
|
320 |
imgsz=imgsz
|
321 |
)
|
@@ -324,121 +281,100 @@ def process(
|
|
324 |
raise ValueError("Failed to generate labeled image")
|
325 |
|
326 |
except Exception as e:
|
327 |
-
print(f"
|
328 |
-
|
329 |
-
return image_input, f"⚠️ Error during element detection: {str(e)}"
|
330 |
|
331 |
-
# Decode
|
332 |
try:
|
333 |
image = Image.open(io.BytesIO(base64.b64decode(dino_labled_img)))
|
334 |
-
print('Successfully decoded processed image')
|
335 |
except Exception as e:
|
336 |
-
print(f"
|
337 |
-
return image_input, f"⚠️ Error decoding
|
338 |
|
339 |
-
# Format
|
340 |
if parsed_content_list and len(parsed_content_list) > 0:
|
341 |
parsed_text = "🎯 **Detected Elements:**\n\n"
|
342 |
for i, v in enumerate(parsed_content_list):
|
343 |
-
if v:
|
344 |
-
parsed_text += f"**
|
345 |
else:
|
346 |
-
parsed_text = "ℹ️ No UI elements detected. Try adjusting the
|
347 |
|
348 |
-
print(f'
|
349 |
return image, parsed_text
|
350 |
|
351 |
except Exception as e:
|
352 |
-
|
353 |
-
print(f"Error during processing: {e}")
|
354 |
print(traceback.format_exc())
|
355 |
-
return None,
|
356 |
|
357 |
-
# Build
|
358 |
-
with gr.Blocks(css=custom_css, theme=gr.themes.Soft()
|
359 |
gr.Markdown(MARKDOWN)
|
360 |
|
361 |
-
|
362 |
-
|
363 |
-
gr.Markdown("### ⚠️ Warning: Caption model failed to load. Some features may not work.")
|
364 |
|
365 |
with gr.Row():
|
366 |
-
# Left sidebar: Upload and settings
|
367 |
with gr.Column(scale=1):
|
368 |
-
with gr.Accordion("📤 Upload
|
369 |
image_input_component = gr.Image(
|
370 |
type='pil',
|
371 |
-
label='Upload Screenshot
|
372 |
elem_id="input_image"
|
373 |
)
|
374 |
|
375 |
gr.Markdown("### 🎛️ Detection Settings")
|
376 |
|
377 |
-
|
378 |
-
|
379 |
-
|
380 |
-
|
381 |
-
|
382 |
-
|
383 |
-
|
384 |
-
|
385 |
-
)
|
386 |
-
|
387 |
-
iou_threshold_component = gr.Slider(
|
388 |
-
label='🔲 IOU Threshold',
|
389 |
-
minimum=0.01,
|
390 |
-
maximum=1.0,
|
391 |
-
step=0.01,
|
392 |
-
value=0.1,
|
393 |
-
info="Controls overlap filtering"
|
394 |
-
)
|
395 |
-
|
396 |
-
use_paddleocr_component = gr.Checkbox(
|
397 |
-
label='🔤 Use PaddleOCR',
|
398 |
-
value=True,
|
399 |
-
info="✓ PaddleOCR | ✗ EasyOCR"
|
400 |
-
)
|
401 |
-
|
402 |
-
imgsz_component = gr.Slider(
|
403 |
-
label='📐 Detection Image Size',
|
404 |
-
minimum=640,
|
405 |
-
maximum=1920,
|
406 |
-
step=32,
|
407 |
-
value=640,
|
408 |
-
info="Higher = better accuracy but slower"
|
409 |
-
)
|
410 |
|
411 |
-
|
412 |
-
|
413 |
-
|
414 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
415 |
)
|
416 |
|
417 |
-
gr.
|
418 |
-
|
419 |
-
|
420 |
-
|
421 |
-
- **Complex UIs:** Lower box threshold to 0.03
|
422 |
-
- **Too many boxes:** Increase IOU threshold
|
423 |
-
""")
|
424 |
|
425 |
-
# Right main area: Results tabs
|
426 |
with gr.Column(scale=2):
|
427 |
with gr.Tabs():
|
428 |
-
with gr.Tab("🖼️
|
429 |
image_output_component = gr.Image(
|
430 |
type='pil',
|
431 |
-
label='
|
432 |
-
elem_classes=["output-image"]
|
433 |
)
|
434 |
|
435 |
-
with gr.Tab("📝
|
436 |
text_output_component = gr.Markdown(
|
437 |
-
value="*
|
438 |
-
elem_classes=["parsed-text"]
|
439 |
)
|
440 |
|
441 |
-
# Button click event
|
442 |
submit_button_component.click(
|
443 |
fn=process,
|
444 |
inputs=[
|
@@ -452,13 +388,9 @@ with gr.Blocks(css=custom_css, theme=gr.themes.Soft(), title="OmniParser V2 Pro"
|
|
452 |
show_progress=True
|
453 |
)
|
454 |
|
455 |
-
# Launch
|
456 |
if __name__ == "__main__":
|
457 |
try:
|
458 |
-
# Set environment variables
|
459 |
-
os.environ['TRANSFORMERS_OFFLINE'] = '0'
|
460 |
-
os.environ['HF_HUB_OFFLINE'] = '0'
|
461 |
-
|
462 |
demo.queue(max_size=10)
|
463 |
demo.launch(
|
464 |
share=False,
|
@@ -467,5 +399,4 @@ if __name__ == "__main__":
|
|
467 |
server_port=7860
|
468 |
)
|
469 |
except Exception as e:
|
470 |
-
print(f"
|
471 |
-
print(traceback.format_exc())
|
|
|
11 |
import warnings
|
12 |
import sys
|
13 |
|
14 |
+
# Suppress warnings
|
15 |
warnings.filterwarnings("ignore", category=FutureWarning)
|
16 |
warnings.filterwarnings("ignore", message=".*_supports_sdpa.*")
|
17 |
|
18 |
+
# Simple monkey patch for transformers - avoid recursion
|
19 |
+
def simple_patch_transformers():
|
20 |
+
"""Simple patch to fix _supports_sdpa issue"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
try:
|
22 |
import transformers.modeling_utils as modeling_utils
|
23 |
|
24 |
+
# Store original method
|
25 |
original_check = modeling_utils.PreTrainedModel._check_and_adjust_attn_implementation
|
26 |
|
27 |
def patched_check(self, *args, **kwargs):
|
28 |
+
# Simply set the attribute if it doesn't exist
|
29 |
if not hasattr(self, '_supports_sdpa'):
|
30 |
+
object.__setattr__(self, '_supports_sdpa', False)
|
31 |
+
|
32 |
try:
|
33 |
return original_check(self, *args, **kwargs)
|
34 |
except AttributeError as e:
|
35 |
if '_supports_sdpa' in str(e):
|
36 |
+
# Return default attention implementation
|
37 |
return "eager"
|
38 |
raise
|
39 |
|
40 |
modeling_utils.PreTrainedModel._check_and_adjust_attn_implementation = patched_check
|
41 |
+
print("Applied simple transformers patch")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
|
43 |
except Exception as e:
|
44 |
print(f"Warning: Could not patch transformers: {e}")
|
45 |
|
46 |
+
# Apply the patch BEFORE importing utils
|
47 |
+
simple_patch_transformers()
|
48 |
|
49 |
+
# Now import the utils
|
50 |
+
from util.utils import check_ocr_box, get_yolo_model, get_caption_model_processor, get_som_labeled_img
|
51 |
|
52 |
+
# Download repository
|
53 |
repo_id = "microsoft/OmniParser-v2.0"
|
54 |
local_dir = "weights"
|
55 |
|
|
|
59 |
else:
|
60 |
print(f"Weights already exist at: {local_dir}")
|
61 |
|
62 |
+
# Custom function to load caption model
|
63 |
def load_caption_model_safe(model_name="florence2", model_name_or_path="weights/icon_caption"):
|
64 |
+
"""Safely load caption model"""
|
65 |
|
66 |
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
67 |
|
68 |
+
# Method 1: Try original function
|
69 |
try:
|
|
|
|
|
70 |
return get_caption_model_processor(model_name, model_name_or_path)
|
71 |
+
except Exception as e:
|
72 |
+
print(f"Original loading failed: {e}, trying alternative...")
|
|
|
|
|
|
|
73 |
|
74 |
+
# Method 2: Load with specific configs
|
75 |
try:
|
76 |
from transformers import AutoProcessor, AutoModelForCausalLM
|
77 |
|
78 |
+
print(f"Loading caption model from {model_name_or_path}...")
|
79 |
|
|
|
80 |
processor = AutoProcessor.from_pretrained(
|
81 |
model_name_or_path,
|
82 |
+
trust_remote_code=True
|
|
|
83 |
)
|
84 |
|
85 |
+
# Load model with safer config
|
86 |
+
model = AutoModelForCausalLM.from_pretrained(
|
87 |
+
model_name_or_path,
|
88 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
89 |
+
trust_remote_code=True,
|
90 |
+
attn_implementation="eager", # Use eager attention
|
91 |
+
low_cpu_mem_usage=True
|
92 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
|
94 |
+
# Ensure attribute exists (using object.__setattr__ to avoid recursion)
|
95 |
+
if not hasattr(model, '_supports_sdpa'):
|
96 |
+
object.__setattr__(model, '_supports_sdpa', False)
|
97 |
|
98 |
+
if device.type == 'cuda':
|
|
|
99 |
model = model.to(device)
|
100 |
|
101 |
+
print("Model loaded successfully with alternative method")
|
102 |
return {'model': model, 'processor': processor}
|
103 |
|
104 |
except Exception as e:
|
105 |
+
print(f"Alternative loading also failed: {e}")
|
106 |
+
|
107 |
+
# Method 3: Manual loading as last resort
|
108 |
+
try:
|
109 |
+
print("Attempting manual model loading...")
|
110 |
+
|
111 |
+
# Import required modules
|
112 |
+
from transformers import AutoProcessor, AutoConfig
|
113 |
+
import importlib.util
|
114 |
+
|
115 |
+
# Load processor
|
116 |
+
processor = AutoProcessor.from_pretrained(
|
117 |
+
model_name_or_path,
|
118 |
+
trust_remote_code=True
|
119 |
+
)
|
120 |
+
|
121 |
+
# Load config
|
122 |
+
config = AutoConfig.from_pretrained(
|
123 |
+
model_name_or_path,
|
124 |
+
trust_remote_code=True
|
125 |
+
)
|
126 |
+
|
127 |
+
# Manually import and instantiate model
|
128 |
+
model_file = os.path.join(model_name_or_path, "modeling_florence2.py")
|
129 |
+
if os.path.exists(model_file):
|
130 |
+
spec = importlib.util.spec_from_file_location("modeling_florence2_custom", model_file)
|
131 |
+
module = importlib.util.module_from_spec(spec)
|
132 |
+
spec.loader.exec_module(module)
|
133 |
+
|
134 |
+
# Get model class
|
135 |
+
if hasattr(module, 'Florence2ForConditionalGeneration'):
|
136 |
+
model_class = module.Florence2ForConditionalGeneration
|
137 |
+
|
138 |
+
# Create model instance
|
139 |
+
model = model_class(config)
|
140 |
+
|
141 |
+
# Set the attribute before loading weights
|
142 |
+
object.__setattr__(model, '_supports_sdpa', False)
|
143 |
+
|
144 |
+
# Load weights
|
145 |
+
weight_file = os.path.join(model_name_or_path, "model.safetensors")
|
146 |
+
if os.path.exists(weight_file):
|
147 |
+
from safetensors.torch import load_file
|
148 |
+
state_dict = load_file(weight_file)
|
149 |
+
model.load_state_dict(state_dict, strict=False)
|
150 |
+
|
151 |
+
if device.type == 'cuda':
|
152 |
+
model = model.to(device)
|
153 |
+
model = model.half() # Use half precision
|
154 |
+
|
155 |
+
print("Model loaded successfully with manual method")
|
156 |
+
return {'model': model, 'processor': processor}
|
157 |
+
|
158 |
+
except Exception as e:
|
159 |
+
print(f"Manual loading failed: {e}")
|
160 |
+
raise RuntimeError(f"Could not load model with any method: {e}")
|
161 |
|
162 |
# Load models
|
163 |
try:
|
|
|
173 |
print(f"Critical error loading models: {e}")
|
174 |
print(traceback.format_exc())
|
175 |
caption_model_processor = None
|
176 |
+
yolo_model = None
|
177 |
|
178 |
+
# UI Configuration
|
179 |
MARKDOWN = """
|
180 |
# OmniParser V2 Pro🔥
|
181 |
|
|
|
188 |
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
189 |
print(f"Using device: {DEVICE}")
|
190 |
|
|
|
191 |
custom_css = """
|
192 |
body { background-color: #f0f2f5; }
|
193 |
.gradio-container { font-family: 'Segoe UI', sans-serif; max-width: 1400px; margin: auto; }
|
|
|
197 |
.output-image { border: 2px solid #e1e4e8; border-radius: 8px; }
|
198 |
#input_image { border: 2px dashed #4a90e2; border-radius: 8px; }
|
199 |
#input_image:hover { border-color: #2c5aa0; }
|
|
|
|
|
200 |
"""
|
201 |
|
202 |
@spaces.GPU
|
|
|
208 |
use_paddleocr,
|
209 |
imgsz
|
210 |
) -> tuple:
|
211 |
+
"""Process image with error handling"""
|
212 |
|
|
|
213 |
if image_input is None:
|
214 |
return None, "⚠️ Please upload an image for processing."
|
215 |
|
216 |
+
if caption_model_processor is None or yolo_model is None:
|
217 |
+
return None, "⚠️ Models not loaded properly. Please restart the application."
|
|
|
218 |
|
219 |
try:
|
220 |
+
print(f"Processing: box_threshold={box_threshold}, iou_threshold={iou_threshold}, "
|
221 |
+
f"use_paddleocr={use_paddleocr}, imgsz={imgsz}")
|
|
|
222 |
|
223 |
+
# Calculate overlay ratio
|
224 |
image_width = image_input.size[0]
|
225 |
box_overlay_ratio = max(0.5, min(2.0, image_width / 3200))
|
226 |
|
|
|
231 |
'thickness': max(int(3 * box_overlay_ratio), 1),
|
232 |
}
|
233 |
|
234 |
+
# OCR processing
|
235 |
try:
|
236 |
ocr_bbox_rslt, is_goal_filtered = check_ocr_box(
|
237 |
image_input,
|
|
|
242 |
use_paddleocr=use_paddleocr
|
243 |
)
|
244 |
|
|
|
245 |
if ocr_bbox_rslt is None:
|
|
|
246 |
text, ocr_bbox = [], []
|
247 |
else:
|
248 |
text, ocr_bbox = ocr_bbox_rslt
|
249 |
|
250 |
+
text = text if text is not None else []
|
251 |
+
ocr_bbox = ocr_bbox if ocr_bbox is not None else []
|
252 |
+
|
|
|
|
|
|
|
253 |
print(f"OCR found {len(text)} text regions")
|
254 |
|
255 |
except Exception as e:
|
256 |
+
print(f"OCR error: {e}")
|
257 |
text, ocr_bbox = [], []
|
258 |
|
259 |
+
# Object detection and captioning
|
260 |
try:
|
261 |
+
# Ensure model has _supports_sdpa attribute
|
262 |
if isinstance(caption_model_processor, dict) and 'model' in caption_model_processor:
|
263 |
model = caption_model_processor['model']
|
264 |
if not hasattr(model, '_supports_sdpa'):
|
265 |
+
object.__setattr__(model, '_supports_sdpa', False)
|
266 |
|
267 |
dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(
|
268 |
image_input,
|
269 |
yolo_model,
|
270 |
BOX_TRESHOLD=box_threshold,
|
271 |
output_coord_in_ratio=True,
|
272 |
+
ocr_bbox=ocr_bbox,
|
273 |
draw_bbox_config=draw_bbox_config,
|
274 |
caption_model_processor=caption_model_processor,
|
275 |
+
ocr_text=text,
|
276 |
iou_threshold=iou_threshold,
|
277 |
imgsz=imgsz
|
278 |
)
|
|
|
281 |
raise ValueError("Failed to generate labeled image")
|
282 |
|
283 |
except Exception as e:
|
284 |
+
print(f"Detection error: {e}")
|
285 |
+
return image_input, f"⚠️ Error during detection: {str(e)}"
|
|
|
286 |
|
287 |
+
# Decode image
|
288 |
try:
|
289 |
image = Image.open(io.BytesIO(base64.b64decode(dino_labled_img)))
|
|
|
290 |
except Exception as e:
|
291 |
+
print(f"Image decode error: {e}")
|
292 |
+
return image_input, f"⚠️ Error decoding image: {str(e)}"
|
293 |
|
294 |
+
# Format results
|
295 |
if parsed_content_list and len(parsed_content_list) > 0:
|
296 |
parsed_text = "🎯 **Detected Elements:**\n\n"
|
297 |
for i, v in enumerate(parsed_content_list):
|
298 |
+
if v:
|
299 |
+
parsed_text += f"**Element {i}:** {v}\n"
|
300 |
else:
|
301 |
+
parsed_text = "ℹ️ No UI elements detected. Try adjusting the thresholds."
|
302 |
|
303 |
+
print(f'Processing complete. Found {len(parsed_content_list)} elements.')
|
304 |
return image, parsed_text
|
305 |
|
306 |
except Exception as e:
|
307 |
+
print(f"Processing error: {e}")
|
|
|
308 |
print(traceback.format_exc())
|
309 |
+
return None, f"⚠️ Error: {str(e)}"
|
310 |
|
311 |
+
# Build UI
|
312 |
+
with gr.Blocks(css=custom_css, theme=gr.themes.Soft()) as demo:
|
313 |
gr.Markdown(MARKDOWN)
|
314 |
|
315 |
+
if caption_model_processor is None or yolo_model is None:
|
316 |
+
gr.Markdown("### ⚠️ Warning: Models failed to load. Please check logs.")
|
|
|
317 |
|
318 |
with gr.Row():
|
|
|
319 |
with gr.Column(scale=1):
|
320 |
+
with gr.Accordion("📤 Upload & Settings", open=True):
|
321 |
image_input_component = gr.Image(
|
322 |
type='pil',
|
323 |
+
label='Upload Screenshot',
|
324 |
elem_id="input_image"
|
325 |
)
|
326 |
|
327 |
gr.Markdown("### 🎛️ Detection Settings")
|
328 |
|
329 |
+
box_threshold_component = gr.Slider(
|
330 |
+
label='Box Threshold',
|
331 |
+
minimum=0.01,
|
332 |
+
maximum=1.0,
|
333 |
+
step=0.01,
|
334 |
+
value=0.05,
|
335 |
+
info="Lower = more detections"
|
336 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
337 |
|
338 |
+
iou_threshold_component = gr.Slider(
|
339 |
+
label='IOU Threshold',
|
340 |
+
minimum=0.01,
|
341 |
+
maximum=1.0,
|
342 |
+
step=0.01,
|
343 |
+
value=0.1,
|
344 |
+
info="Overlap filtering"
|
345 |
+
)
|
346 |
+
|
347 |
+
use_paddleocr_component = gr.Checkbox(
|
348 |
+
label='Use PaddleOCR',
|
349 |
+
value=True
|
350 |
+
)
|
351 |
+
|
352 |
+
imgsz_component = gr.Slider(
|
353 |
+
label='Image Size',
|
354 |
+
minimum=640,
|
355 |
+
maximum=1920,
|
356 |
+
step=32,
|
357 |
+
value=640
|
358 |
)
|
359 |
|
360 |
+
submit_button_component = gr.Button(
|
361 |
+
value='🚀 Process',
|
362 |
+
variant='primary'
|
363 |
+
)
|
|
|
|
|
|
|
364 |
|
|
|
365 |
with gr.Column(scale=2):
|
366 |
with gr.Tabs():
|
367 |
+
with gr.Tab("🖼️ Result"):
|
368 |
image_output_component = gr.Image(
|
369 |
type='pil',
|
370 |
+
label='Annotated Image'
|
|
|
371 |
)
|
372 |
|
373 |
+
with gr.Tab("📝 Elements"):
|
374 |
text_output_component = gr.Markdown(
|
375 |
+
value="*Results will appear here...*"
|
|
|
376 |
)
|
377 |
|
|
|
378 |
submit_button_component.click(
|
379 |
fn=process,
|
380 |
inputs=[
|
|
|
388 |
show_progress=True
|
389 |
)
|
390 |
|
391 |
+
# Launch
|
392 |
if __name__ == "__main__":
|
393 |
try:
|
|
|
|
|
|
|
|
|
394 |
demo.queue(max_size=10)
|
395 |
demo.launch(
|
396 |
share=False,
|
|
|
399 |
server_port=7860
|
400 |
)
|
401 |
except Exception as e:
|
402 |
+
print(f"Launch failed: {e}")
|
|