Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,13 +1,14 @@
|
|
1 |
-
import
|
2 |
-
import
|
3 |
-
|
4 |
from PIL import Image
|
|
|
|
|
5 |
import torch
|
6 |
import re
|
7 |
import logging
|
8 |
-
|
9 |
-
import
|
10 |
-
from io import BytesIO
|
11 |
|
12 |
# Configure logging
|
13 |
logging.basicConfig(level=logging.INFO)
|
@@ -20,7 +21,7 @@ tokenizer = None
|
|
20 |
model_name = "microsoft/GUI-Actor-2B-Qwen2-VL"
|
21 |
model_loaded = False
|
22 |
|
23 |
-
def load_model():
|
24 |
"""Load model with proper error handling and fallback strategies"""
|
25 |
global model, processor, tokenizer, model_loaded
|
26 |
|
@@ -39,8 +40,8 @@ def load_model():
|
|
39 |
|
40 |
model = Qwen2VLForConditionalGeneration.from_pretrained(
|
41 |
model_name,
|
42 |
-
torch_dtype=torch.
|
43 |
-
device_map=
|
44 |
trust_remote_code=True,
|
45 |
low_cpu_mem_usage=True
|
46 |
).eval()
|
@@ -52,17 +53,17 @@ def load_model():
|
|
52 |
logger.info("Trying AutoProcessor and AutoModel fallback...")
|
53 |
|
54 |
try:
|
55 |
-
from transformers import AutoProcessor,
|
56 |
|
57 |
processor = AutoProcessor.from_pretrained(
|
58 |
model_name,
|
59 |
trust_remote_code=True
|
60 |
)
|
61 |
|
62 |
-
model =
|
63 |
model_name,
|
64 |
-
torch_dtype=torch.
|
65 |
-
device_map=
|
66 |
trust_remote_code=True,
|
67 |
low_cpu_mem_usage=True
|
68 |
).eval()
|
@@ -74,7 +75,7 @@ def load_model():
|
|
74 |
logger.info("Trying generic transformers approach...")
|
75 |
|
76 |
# Last fallback - try loading as generic model
|
77 |
-
from transformers import AutoConfig,
|
78 |
import transformers
|
79 |
|
80 |
config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
|
@@ -96,8 +97,8 @@ def load_model():
|
|
96 |
model = ModelClass.from_pretrained(
|
97 |
model_name,
|
98 |
config=config,
|
99 |
-
torch_dtype=torch.
|
100 |
-
device_map=
|
101 |
trust_remote_code=True,
|
102 |
low_cpu_mem_usage=True
|
103 |
).eval()
|
@@ -116,8 +117,30 @@ def load_model():
|
|
116 |
model_loaded = False
|
117 |
return False
|
118 |
|
119 |
-
|
120 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
121 |
# Pattern untuk mencari koordinat dalam berbagai format
|
122 |
patterns = [
|
123 |
r'click\s*\(\s*(\d+(?:\.\d+)?)\s*,\s*(\d+(?:\.\d+)?)\s*\)', # click(x, y)
|
@@ -143,38 +166,11 @@ def extract_coordinates(text: str) -> List[Tuple[float, float]]:
|
|
143 |
# Default ke center jika tidak ditemukan
|
144 |
return [(0.5, 0.5)]
|
145 |
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
return "Model not loaded properly", 0.5, 0.5
|
151 |
-
|
152 |
try:
|
153 |
-
conversation = [
|
154 |
-
{
|
155 |
-
"role": "system",
|
156 |
-
"content": [
|
157 |
-
{
|
158 |
-
"type": "text",
|
159 |
-
"text": "You are a GUI agent. You are given a task and a screenshot of the screen. You need to perform a series of pyautogui actions to complete the task. Please provide the click coordinates.",
|
160 |
-
}
|
161 |
-
]
|
162 |
-
},
|
163 |
-
{
|
164 |
-
"role": "user",
|
165 |
-
"content": [
|
166 |
-
{
|
167 |
-
"type": "image",
|
168 |
-
"image": pil_image,
|
169 |
-
},
|
170 |
-
{
|
171 |
-
"type": "text",
|
172 |
-
"text": instruction,
|
173 |
-
},
|
174 |
-
],
|
175 |
-
},
|
176 |
-
]
|
177 |
-
|
178 |
# Apply chat template
|
179 |
text = processor.apply_chat_template(
|
180 |
conversation,
|
@@ -190,15 +186,11 @@ def inference(pil_image: Image.Image, instruction: str):
|
|
190 |
text=[text],
|
191 |
images=[image],
|
192 |
return_tensors="pt",
|
193 |
-
padding=True,
|
194 |
-
truncation=True,
|
195 |
-
max_length=512
|
196 |
)
|
197 |
|
198 |
-
# Move inputs to the same device as model
|
199 |
-
if torch.cuda.is_available():
|
200 |
-
inputs = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}
|
201 |
-
|
202 |
# Generate response with proper error handling
|
203 |
with torch.no_grad():
|
204 |
try:
|
@@ -226,97 +218,119 @@ def inference(pil_image: Image.Image, instruction: str):
|
|
226 |
|
227 |
# Extract coordinates
|
228 |
coordinates = extract_coordinates(response)
|
229 |
-
px, py = coordinates[0]
|
230 |
|
231 |
-
return
|
|
|
|
|
|
|
|
|
232 |
|
233 |
except Exception as e:
|
234 |
logger.error(f"Inference error: {e}")
|
235 |
-
return
|
|
|
|
|
|
|
|
|
236 |
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
-
|
245 |
-
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
|
|
|
|
|
251 |
|
252 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
253 |
|
254 |
-
|
255 |
-
|
256 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
257 |
|
258 |
-
#
|
259 |
-
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
|
264 |
-
|
265 |
-
|
266 |
-
|
267 |
-
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
with gr.Column():
|
278 |
-
response_output = gr.Textbox(
|
279 |
-
label="Model Response",
|
280 |
-
lines=5,
|
281 |
-
interactive=False
|
282 |
-
)
|
283 |
-
with gr.Row():
|
284 |
-
x_output = gr.Number(
|
285 |
-
label="X Coordinate (normalized)",
|
286 |
-
precision=4,
|
287 |
-
interactive=False
|
288 |
-
)
|
289 |
-
y_output = gr.Number(
|
290 |
-
label="Y Coordinate (normalized)",
|
291 |
-
precision=4,
|
292 |
-
interactive=False
|
293 |
-
)
|
294 |
-
|
295 |
-
# Status indicator
|
296 |
-
with gr.Row():
|
297 |
-
gr.Markdown(f"**Model Status:** {'✅ Loaded' if model_loaded else '❌ Not Loaded'}")
|
298 |
-
gr.Markdown(f"**Device:** {'GPU' if torch.cuda.is_available() else 'CPU'}")
|
299 |
-
|
300 |
-
# Examples
|
301 |
-
gr.Examples(
|
302 |
-
examples=[
|
303 |
-
["Click on the search button", None],
|
304 |
-
["Select the dropdown menu", None],
|
305 |
-
["Click on the submit form", None],
|
306 |
-
],
|
307 |
-
inputs=[instruction_input, image_input],
|
308 |
-
)
|
309 |
-
|
310 |
-
# Event handlers
|
311 |
-
submit_btn.click(
|
312 |
-
fn=process_image,
|
313 |
-
inputs=[image_input, instruction_input],
|
314 |
-
outputs=[response_output, x_output, y_output]
|
315 |
-
)
|
316 |
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
321 |
-
|
322 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from fastapi import FastAPI, HTTPException
|
2 |
+
from fastapi.responses import JSONResponse
|
3 |
+
from pydantic import BaseModel
|
4 |
from PIL import Image
|
5 |
+
from io import BytesIO
|
6 |
+
import base64
|
7 |
import torch
|
8 |
import re
|
9 |
import logging
|
10 |
+
import asyncio
|
11 |
+
from contextlib import asynccontextmanager
|
|
|
12 |
|
13 |
# Configure logging
|
14 |
logging.basicConfig(level=logging.INFO)
|
|
|
21 |
model_name = "microsoft/GUI-Actor-2B-Qwen2-VL"
|
22 |
model_loaded = False
|
23 |
|
24 |
+
async def load_model():
|
25 |
"""Load model with proper error handling and fallback strategies"""
|
26 |
global model, processor, tokenizer, model_loaded
|
27 |
|
|
|
40 |
|
41 |
model = Qwen2VLForConditionalGeneration.from_pretrained(
|
42 |
model_name,
|
43 |
+
torch_dtype=torch.float32,
|
44 |
+
device_map=None, # CPU only
|
45 |
trust_remote_code=True,
|
46 |
low_cpu_mem_usage=True
|
47 |
).eval()
|
|
|
53 |
logger.info("Trying AutoProcessor and AutoModel fallback...")
|
54 |
|
55 |
try:
|
56 |
+
from transformers import AutoProcessor, AutoModel
|
57 |
|
58 |
processor = AutoProcessor.from_pretrained(
|
59 |
model_name,
|
60 |
trust_remote_code=True
|
61 |
)
|
62 |
|
63 |
+
model = AutoModel.from_pretrained(
|
64 |
model_name,
|
65 |
+
torch_dtype=torch.float32,
|
66 |
+
device_map=None,
|
67 |
trust_remote_code=True,
|
68 |
low_cpu_mem_usage=True
|
69 |
).eval()
|
|
|
75 |
logger.info("Trying generic transformers approach...")
|
76 |
|
77 |
# Last fallback - try loading as generic model
|
78 |
+
from transformers import AutoConfig, AutoTokenizer
|
79 |
import transformers
|
80 |
|
81 |
config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
|
|
|
97 |
model = ModelClass.from_pretrained(
|
98 |
model_name,
|
99 |
config=config,
|
100 |
+
torch_dtype=torch.float32,
|
101 |
+
device_map=None,
|
102 |
trust_remote_code=True,
|
103 |
low_cpu_mem_usage=True
|
104 |
).eval()
|
|
|
117 |
model_loaded = False
|
118 |
return False
|
119 |
|
120 |
+
@asynccontextmanager
|
121 |
+
async def lifespan(app: FastAPI):
|
122 |
+
# Startup
|
123 |
+
logger.info("Starting up GUI-Actor API...")
|
124 |
+
await load_model()
|
125 |
+
yield
|
126 |
+
# Shutdown
|
127 |
+
logger.info("Shutting down GUI-Actor API...")
|
128 |
+
|
129 |
+
# Initialize FastAPI app with lifespan
|
130 |
+
app = FastAPI(
|
131 |
+
title="GUI-Actor API",
|
132 |
+
version="1.0.0",
|
133 |
+
lifespan=lifespan
|
134 |
+
)
|
135 |
+
|
136 |
+
class Base64Request(BaseModel):
|
137 |
+
image_base64: str
|
138 |
+
instruction: str
|
139 |
+
|
140 |
+
def extract_coordinates(text):
|
141 |
+
"""
|
142 |
+
Extract coordinates from model output text
|
143 |
+
"""
|
144 |
# Pattern untuk mencari koordinat dalam berbagai format
|
145 |
patterns = [
|
146 |
r'click\s*\(\s*(\d+(?:\.\d+)?)\s*,\s*(\d+(?:\.\d+)?)\s*\)', # click(x, y)
|
|
|
166 |
# Default ke center jika tidak ditemukan
|
167 |
return [(0.5, 0.5)]
|
168 |
|
169 |
+
def cpu_inference(conversation, model, tokenizer, processor):
|
170 |
+
"""
|
171 |
+
Inference function untuk CPU with better error handling
|
172 |
+
"""
|
|
|
|
|
173 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
174 |
# Apply chat template
|
175 |
text = processor.apply_chat_template(
|
176 |
conversation,
|
|
|
186 |
text=[text],
|
187 |
images=[image],
|
188 |
return_tensors="pt",
|
189 |
+
padding=True, # Enable padding
|
190 |
+
truncation=True, # Enable truncation for long texts
|
191 |
+
max_length=512 # Set reasonable max length
|
192 |
)
|
193 |
|
|
|
|
|
|
|
|
|
194 |
# Generate response with proper error handling
|
195 |
with torch.no_grad():
|
196 |
try:
|
|
|
218 |
|
219 |
# Extract coordinates
|
220 |
coordinates = extract_coordinates(response)
|
|
|
221 |
|
222 |
+
return {
|
223 |
+
"topk_points": coordinates,
|
224 |
+
"response": response,
|
225 |
+
"success": True
|
226 |
+
}
|
227 |
|
228 |
except Exception as e:
|
229 |
logger.error(f"Inference error: {e}")
|
230 |
+
return {
|
231 |
+
"topk_points": [(0.5, 0.5)],
|
232 |
+
"response": f"Error during inference: {str(e)}",
|
233 |
+
"success": False
|
234 |
+
}
|
235 |
|
236 |
+
@app.get("/")
|
237 |
+
async def root():
|
238 |
+
return {
|
239 |
+
"message": "GUI-Actor API is running",
|
240 |
+
"status": "healthy",
|
241 |
+
"model_loaded": model_loaded,
|
242 |
+
"model_name": model_name
|
243 |
+
}
|
244 |
+
|
245 |
+
@app.post("/click/base64")
|
246 |
+
async def predict_click_base64(data: Base64Request):
|
247 |
+
if not model_loaded:
|
248 |
+
raise HTTPException(
|
249 |
+
status_code=503,
|
250 |
+
detail="Model not loaded properly"
|
251 |
+
)
|
252 |
|
253 |
+
try:
|
254 |
+
# Decode base64 to image
|
255 |
+
try:
|
256 |
+
# Handle data URL format
|
257 |
+
if "," in data.image_base64:
|
258 |
+
image_data = base64.b64decode(data.image_base64.split(",")[-1])
|
259 |
+
else:
|
260 |
+
image_data = base64.b64decode(data.image_base64)
|
261 |
+
except Exception as e:
|
262 |
+
raise HTTPException(status_code=400, detail=f"Invalid base64 image: {e}")
|
263 |
+
|
264 |
+
try:
|
265 |
+
pil_image = Image.open(BytesIO(image_data)).convert("RGB")
|
266 |
+
except Exception as e:
|
267 |
+
raise HTTPException(status_code=400, detail=f"Invalid image format: {e}")
|
268 |
|
269 |
+
conversation = [
|
270 |
+
{
|
271 |
+
"role": "system",
|
272 |
+
"content": [
|
273 |
+
{
|
274 |
+
"type": "text",
|
275 |
+
"text": "You are a GUI agent. You are given a task and a screenshot of the screen. You need to perform a series of pyautogui actions to complete the task. Please provide the click coordinates.",
|
276 |
+
}
|
277 |
+
]
|
278 |
+
},
|
279 |
+
{
|
280 |
+
"role": "user",
|
281 |
+
"content": [
|
282 |
+
{
|
283 |
+
"type": "image",
|
284 |
+
"image": pil_image,
|
285 |
+
},
|
286 |
+
{
|
287 |
+
"type": "text",
|
288 |
+
"text": data.instruction,
|
289 |
+
},
|
290 |
+
],
|
291 |
+
},
|
292 |
+
]
|
293 |
|
294 |
+
# Run inference
|
295 |
+
pred = cpu_inference(conversation, model, tokenizer, processor)
|
296 |
+
px, py = pred["topk_points"][0]
|
297 |
+
|
298 |
+
return JSONResponse(content={
|
299 |
+
"x": round(px, 4),
|
300 |
+
"y": round(py, 4),
|
301 |
+
"response": pred["response"],
|
302 |
+
"success": pred["success"]
|
303 |
+
})
|
304 |
+
|
305 |
+
except HTTPException:
|
306 |
+
raise
|
307 |
+
except Exception as e:
|
308 |
+
logger.error(f"Prediction error: {e}")
|
309 |
+
raise HTTPException(
|
310 |
+
status_code=500,
|
311 |
+
detail=f"Internal server error: {str(e)}"
|
312 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
313 |
|
314 |
+
@app.get("/health")
|
315 |
+
async def health_check():
|
316 |
+
return {
|
317 |
+
"status": "healthy" if model_loaded else "unhealthy",
|
318 |
+
"model": model_name,
|
319 |
+
"device": "cpu",
|
320 |
+
"torch_dtype": "float32",
|
321 |
+
"model_loaded": model_loaded
|
322 |
+
}
|
323 |
+
|
324 |
+
@app.get("/debug")
|
325 |
+
async def debug_info():
|
326 |
+
"""Debug endpoint to check model loading status"""
|
327 |
+
import transformers
|
328 |
+
available_classes = [attr for attr in dir(transformers) if 'Qwen' in attr or 'VL' in attr]
|
329 |
+
|
330 |
+
return {
|
331 |
+
"model_loaded": model_loaded,
|
332 |
+
"processor_type": type(processor).__name__ if processor else None,
|
333 |
+
"model_type": type(model).__name__ if model else None,
|
334 |
+
"available_qwen_classes": available_classes,
|
335 |
+
"transformers_version": transformers.__version__
|
336 |
+
}
|