Spaces:
Sleeping
Sleeping
Upload folder using huggingface_hub
Browse files
app.py
CHANGED
@@ -6,14 +6,14 @@ import gradio as gr
|
|
6 |
import torchaudio
|
7 |
import torchvision
|
8 |
|
9 |
-
# Import Gradio Spaces GPU decorator
|
10 |
-
try:
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
except ImportError:
|
15 |
-
|
16 |
-
|
17 |
|
18 |
# Add parent directory to path to import preprocess functions
|
19 |
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
|
@@ -78,291 +78,151 @@ def app_process_audio_data(waveform, sample_rate):
|
|
78 |
# Similarly for images, but let's import the original one
|
79 |
from preprocess import process_image_data
|
80 |
|
81 |
-
# Apply GPU decorator directly to the function if available
|
82 |
-
if HAS_SPACES:
|
83 |
# Using the decorator directly on the function definition
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
198 |
else:
|
199 |
-
|
200 |
-
|
201 |
-
# Format the result
|
202 |
-
if sweetness is not None:
|
203 |
-
result = f"Predicted Sweetness: {sweetness.item():.2f}/13"
|
204 |
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
return result
|
216 |
-
else:
|
217 |
-
return "Error: Could not predict sweetness. Please try again with different inputs."
|
218 |
-
|
219 |
-
except Exception as e:
|
220 |
-
import traceback
|
221 |
-
error_msg = f"Error: {str(e)}\n\n"
|
222 |
-
error_msg += traceback.format_exc()
|
223 |
-
print(f"\033[91mERR!\033[0m: {error_msg}")
|
224 |
-
return error_msg
|
225 |
|
226 |
print("\033[92mINFO\033[0m: GPU-accelerated prediction function created with @spaces.GPU decorator")
|
227 |
-
|
228 |
-
# Regular version without GPU decorator for non-Spaces environments
|
229 |
-
def predict_sweetness(audio, image, model_path):
|
230 |
-
"""Predict sweetness of a watermelon from audio and image input"""
|
231 |
-
try:
|
232 |
-
# Check for device - will be CPU in this case
|
233 |
-
device = torch.device("cpu")
|
234 |
-
print(f"\033[92mINFO\033[0m: Using device: {device}")
|
235 |
-
|
236 |
-
# Load model inside the function
|
237 |
-
model = WatermelonModel().to(device)
|
238 |
-
model.load_state_dict(torch.load(model_path, map_location=device))
|
239 |
-
model.eval()
|
240 |
-
print(f"\033[92mINFO\033[0m: Loaded model from {model_path}")
|
241 |
-
|
242 |
-
# Rest of function identical - processing code
|
243 |
-
# Debug information about input types
|
244 |
-
print(f"\033[92mDEBUG\033[0m: Audio input type: {type(audio)}")
|
245 |
-
print(f"\033[92mDEBUG\033[0m: Audio input shape/length: {len(audio)}")
|
246 |
-
print(f"\033[92mDEBUG\033[0m: Image input type: {type(image)}")
|
247 |
-
if isinstance(image, np.ndarray):
|
248 |
-
print(f"\033[92mDEBUG\033[0m: Image input shape: {image.shape}")
|
249 |
-
|
250 |
-
# Handle different audio input formats
|
251 |
-
if isinstance(audio, tuple) and len(audio) == 2:
|
252 |
-
# Standard Gradio format: (sample_rate, audio_data)
|
253 |
-
sample_rate, audio_data = audio
|
254 |
-
print(f"\033[92mDEBUG\033[0m: Audio sample rate: {sample_rate}")
|
255 |
-
print(f"\033[92mDEBUG\033[0m: Audio data shape: {audio_data.shape}")
|
256 |
-
elif isinstance(audio, tuple) and len(audio) > 2:
|
257 |
-
# Sometimes Gradio returns (sample_rate, audio_data, other_info...)
|
258 |
-
sample_rate, audio_data = audio[0], audio[-1]
|
259 |
-
print(f"\033[92mDEBUG\033[0m: Audio sample rate: {sample_rate}")
|
260 |
-
print(f"\033[92mDEBUG\033[0m: Audio data shape: {audio_data.shape}")
|
261 |
-
elif isinstance(audio, str):
|
262 |
-
# Direct path to audio file
|
263 |
-
audio_data, sample_rate = torchaudio.load(audio)
|
264 |
-
print(f"\033[92mDEBUG\033[0m: Loaded audio from path with shape: {audio_data.shape}")
|
265 |
-
else:
|
266 |
-
return f"Error: Unsupported audio format. Got {type(audio)}"
|
267 |
-
|
268 |
-
# Create a temporary file path for the audio and image
|
269 |
-
temp_dir = "temp"
|
270 |
-
os.makedirs(temp_dir, exist_ok=True)
|
271 |
-
|
272 |
-
temp_audio_path = os.path.join(temp_dir, "temp_audio.wav")
|
273 |
-
temp_image_path = os.path.join(temp_dir, "temp_image.jpg")
|
274 |
-
|
275 |
-
# Import necessary libraries
|
276 |
-
from PIL import Image
|
277 |
-
|
278 |
-
# Audio handling - direct processing from the data in memory
|
279 |
-
if isinstance(audio_data, np.ndarray):
|
280 |
-
# Convert numpy array to tensor
|
281 |
-
print(f"\033[92mDEBUG\033[0m: Converting numpy audio with shape {audio_data.shape} to tensor")
|
282 |
-
audio_tensor = torch.tensor(audio_data).float()
|
283 |
-
|
284 |
-
# Handle different audio dimensions
|
285 |
-
if audio_data.ndim == 1:
|
286 |
-
# Single channel audio
|
287 |
-
audio_tensor = audio_tensor.unsqueeze(0)
|
288 |
-
elif audio_data.ndim == 2:
|
289 |
-
# Ensure channels are first dimension
|
290 |
-
if audio_data.shape[0] > audio_data.shape[1]:
|
291 |
-
# More rows than columns, probably (samples, channels)
|
292 |
-
audio_tensor = torch.tensor(audio_data.T).float()
|
293 |
-
else:
|
294 |
-
# Already a tensor
|
295 |
-
audio_tensor = audio_data.float()
|
296 |
-
|
297 |
-
print(f"\033[92mDEBUG\033[0m: Audio tensor shape before processing: {audio_tensor.shape}")
|
298 |
-
|
299 |
-
# Skip saving/loading and process directly
|
300 |
-
mfcc = app_process_audio_data(audio_tensor, sample_rate)
|
301 |
-
print(f"\033[92mDEBUG\033[0m: MFCC tensor shape after processing: {mfcc.shape if mfcc is not None else None}")
|
302 |
-
|
303 |
-
# Image handling
|
304 |
-
if isinstance(image, np.ndarray):
|
305 |
-
print(f"\033[92mDEBUG\033[0m: Converting numpy image with shape {image.shape} to PIL")
|
306 |
-
pil_image = Image.fromarray(image)
|
307 |
-
pil_image.save(temp_image_path)
|
308 |
-
print(f"\033[92mDEBUG\033[0m: Saved image to {temp_image_path}")
|
309 |
-
elif isinstance(image, str):
|
310 |
-
# If image is already a path
|
311 |
-
temp_image_path = image
|
312 |
-
print(f"\033[92mDEBUG\033[0m: Using provided image path: {temp_image_path}")
|
313 |
-
else:
|
314 |
-
return f"Error: Unsupported image format. Got {type(image)}"
|
315 |
-
|
316 |
-
# Process image
|
317 |
-
print(f"\033[92mDEBUG\033[0m: Loading and preprocessing image from {temp_image_path}")
|
318 |
-
image_tensor = torchvision.io.read_image(temp_image_path)
|
319 |
-
print(f"\033[92mDEBUG\033[0m: Loaded image shape: {image_tensor.shape}")
|
320 |
-
image_tensor = image_tensor.float()
|
321 |
-
processed_image = process_image_data(image_tensor)
|
322 |
-
print(f"\033[92mDEBUG\033[0m: Processed image shape: {processed_image.shape if processed_image is not None else None}")
|
323 |
-
|
324 |
-
# Add batch dimension for inference and move to device
|
325 |
-
if mfcc is not None:
|
326 |
-
mfcc = mfcc.unsqueeze(0).to(device)
|
327 |
-
print(f"\033[92mDEBUG\033[0m: Final MFCC shape with batch dimension: {mfcc.shape}")
|
328 |
-
|
329 |
-
if processed_image is not None:
|
330 |
-
processed_image = processed_image.unsqueeze(0).to(device)
|
331 |
-
print(f"\033[92mDEBUG\033[0m: Final image shape with batch dimension: {processed_image.shape}")
|
332 |
-
|
333 |
-
# Run inference
|
334 |
-
print(f"\033[92mDEBUG\033[0m: Running inference on device: {device}")
|
335 |
-
if mfcc is not None and processed_image is not None:
|
336 |
-
with torch.no_grad():
|
337 |
-
sweetness = model(mfcc, processed_image)
|
338 |
-
print(f"\033[92mDEBUG\033[0m: Prediction successful: {sweetness.item()}")
|
339 |
-
else:
|
340 |
-
return "Error: Failed to process inputs. Please check the debug logs."
|
341 |
-
|
342 |
-
# Format the result
|
343 |
-
if sweetness is not None:
|
344 |
-
result = f"Predicted Sweetness: {sweetness.item():.2f}/13"
|
345 |
-
|
346 |
-
# Add a qualitative description
|
347 |
-
if sweetness.item() < 9:
|
348 |
-
result += "\n\nThis watermelon is not very sweet. You might want to choose another one."
|
349 |
-
elif sweetness.item() < 10:
|
350 |
-
result += "\n\nThis watermelon has moderate sweetness."
|
351 |
-
elif sweetness.item() < 11:
|
352 |
-
result += "\n\nThis watermelon is sweet! A good choice."
|
353 |
-
else:
|
354 |
-
result += "\n\nThis watermelon is very sweet! Excellent choice!"
|
355 |
-
|
356 |
-
return result
|
357 |
-
else:
|
358 |
-
return "Error: Could not predict sweetness. Please try again with different inputs."
|
359 |
-
|
360 |
-
except Exception as e:
|
361 |
-
import traceback
|
362 |
-
error_msg = f"Error: {str(e)}\n\n"
|
363 |
-
error_msg += traceback.format_exc()
|
364 |
-
print(f"\033[91mERR!\033[0m: {error_msg}")
|
365 |
-
return error_msg
|
366 |
|
367 |
def create_app(model_path):
|
368 |
"""Create and launch the Gradio interface"""
|
|
|
6 |
import torchaudio
|
7 |
import torchvision
|
8 |
|
9 |
+
# # Import Gradio Spaces GPU decorator
|
10 |
+
# try:
|
11 |
+
# from gradio import spaces
|
12 |
+
# HAS_SPACES = True
|
13 |
+
# print("\033[92mINFO\033[0m: Gradio Spaces detected, GPU acceleration will be enabled")
|
14 |
+
# except ImportError:
|
15 |
+
# HAS_SPACES = False
|
16 |
+
# print("\033[93mWARN\033[0m: gradio.spaces not available, running without GPU optimization")
|
17 |
|
18 |
# Add parent directory to path to import preprocess functions
|
19 |
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
|
|
|
78 |
# Similarly for images, but let's import the original one
|
79 |
from preprocess import process_image_data
|
80 |
|
|
|
|
|
81 |
# Using the decorator directly on the function definition
|
82 |
+
@spaces.GPU
|
83 |
+
def predict_sweetness(audio, image, model_path):
|
84 |
+
"""Function with GPU acceleration"""
|
85 |
+
try:
|
86 |
+
# Now check CUDA availability inside the GPU-decorated function
|
87 |
+
if torch.cuda.is_available():
|
88 |
+
device = torch.device("cuda")
|
89 |
+
print(f"\033[92mINFO\033[0m: CUDA is available. Using device: {device}")
|
90 |
+
else:
|
91 |
+
device = torch.device("cpu")
|
92 |
+
print(f"\033[92mINFO\033[0m: CUDA is not available. Using device: {device}")
|
93 |
+
|
94 |
+
# Load model inside the function to ensure it's on the correct device
|
95 |
+
model = WatermelonModel().to(device)
|
96 |
+
model.load_state_dict(torch.load(model_path, map_location=device))
|
97 |
+
model.eval()
|
98 |
+
print(f"\033[92mINFO\033[0m: Loaded model from {model_path}")
|
99 |
+
|
100 |
+
# Debug information about input types
|
101 |
+
print(f"\033[92mDEBUG\033[0m: Audio input type: {type(audio)}")
|
102 |
+
print(f"\033[92mDEBUG\033[0m: Audio input shape/length: {len(audio)}")
|
103 |
+
print(f"\033[92mDEBUG\033[0m: Image input type: {type(image)}")
|
104 |
+
if isinstance(image, np.ndarray):
|
105 |
+
print(f"\033[92mDEBUG\033[0m: Image input shape: {image.shape}")
|
106 |
+
|
107 |
+
# Handle different audio input formats
|
108 |
+
if isinstance(audio, tuple) and len(audio) == 2:
|
109 |
+
# Standard Gradio format: (sample_rate, audio_data)
|
110 |
+
sample_rate, audio_data = audio
|
111 |
+
print(f"\033[92mDEBUG\033[0m: Audio sample rate: {sample_rate}")
|
112 |
+
print(f"\033[92mDEBUG\033[0m: Audio data shape: {audio_data.shape}")
|
113 |
+
elif isinstance(audio, tuple) and len(audio) > 2:
|
114 |
+
# Sometimes Gradio returns (sample_rate, audio_data, other_info...)
|
115 |
+
sample_rate, audio_data = audio[0], audio[-1]
|
116 |
+
print(f"\033[92mDEBUG\033[0m: Audio sample rate: {sample_rate}")
|
117 |
+
print(f"\033[92mDEBUG\033[0m: Audio data shape: {audio_data.shape}")
|
118 |
+
elif isinstance(audio, str):
|
119 |
+
# Direct path to audio file
|
120 |
+
audio_data, sample_rate = torchaudio.load(audio)
|
121 |
+
print(f"\033[92mDEBUG\033[0m: Loaded audio from path with shape: {audio_data.shape}")
|
122 |
+
else:
|
123 |
+
return f"Error: Unsupported audio format. Got {type(audio)}"
|
124 |
+
|
125 |
+
# Create a temporary file path for the audio and image
|
126 |
+
temp_dir = "temp"
|
127 |
+
os.makedirs(temp_dir, exist_ok=True)
|
128 |
+
|
129 |
+
temp_audio_path = os.path.join(temp_dir, "temp_audio.wav")
|
130 |
+
temp_image_path = os.path.join(temp_dir, "temp_image.jpg")
|
131 |
+
|
132 |
+
# Import necessary libraries
|
133 |
+
from PIL import Image
|
134 |
+
|
135 |
+
# Audio handling - direct processing from the data in memory
|
136 |
+
if isinstance(audio_data, np.ndarray):
|
137 |
+
# Convert numpy array to tensor
|
138 |
+
print(f"\033[92mDEBUG\033[0m: Converting numpy audio with shape {audio_data.shape} to tensor")
|
139 |
+
audio_tensor = torch.tensor(audio_data).float()
|
140 |
+
|
141 |
+
# Handle different audio dimensions
|
142 |
+
if audio_data.ndim == 1:
|
143 |
+
# Single channel audio
|
144 |
+
audio_tensor = audio_tensor.unsqueeze(0)
|
145 |
+
elif audio_data.ndim == 2:
|
146 |
+
# Ensure channels are first dimension
|
147 |
+
if audio_data.shape[0] > audio_data.shape[1]:
|
148 |
+
# More rows than columns, probably (samples, channels)
|
149 |
+
audio_tensor = torch.tensor(audio_data.T).float()
|
150 |
+
else:
|
151 |
+
# Already a tensor
|
152 |
+
audio_tensor = audio_data.float()
|
153 |
+
|
154 |
+
print(f"\033[92mDEBUG\033[0m: Audio tensor shape before processing: {audio_tensor.shape}")
|
155 |
+
|
156 |
+
# Skip saving/loading and process directly
|
157 |
+
mfcc = app_process_audio_data(audio_tensor, sample_rate)
|
158 |
+
print(f"\033[92mDEBUG\033[0m: MFCC tensor shape after processing: {mfcc.shape if mfcc is not None else None}")
|
159 |
+
|
160 |
+
# Image handling
|
161 |
+
if isinstance(image, np.ndarray):
|
162 |
+
print(f"\033[92mDEBUG\033[0m: Converting numpy image with shape {image.shape} to PIL")
|
163 |
+
pil_image = Image.fromarray(image)
|
164 |
+
pil_image.save(temp_image_path)
|
165 |
+
print(f"\033[92mDEBUG\033[0m: Saved image to {temp_image_path}")
|
166 |
+
elif isinstance(image, str):
|
167 |
+
# If image is already a path
|
168 |
+
temp_image_path = image
|
169 |
+
print(f"\033[92mDEBUG\033[0m: Using provided image path: {temp_image_path}")
|
170 |
+
else:
|
171 |
+
return f"Error: Unsupported image format. Got {type(image)}"
|
172 |
+
|
173 |
+
# Process image
|
174 |
+
print(f"\033[92mDEBUG\033[0m: Loading and preprocessing image from {temp_image_path}")
|
175 |
+
image_tensor = torchvision.io.read_image(temp_image_path)
|
176 |
+
print(f"\033[92mDEBUG\033[0m: Loaded image shape: {image_tensor.shape}")
|
177 |
+
image_tensor = image_tensor.float()
|
178 |
+
processed_image = process_image_data(image_tensor)
|
179 |
+
print(f"\033[92mDEBUG\033[0m: Processed image shape: {processed_image.shape if processed_image is not None else None}")
|
180 |
+
|
181 |
+
# Add batch dimension for inference and move to device
|
182 |
+
if mfcc is not None:
|
183 |
+
mfcc = mfcc.unsqueeze(0).to(device)
|
184 |
+
print(f"\033[92mDEBUG\033[0m: Final MFCC shape with batch dimension: {mfcc.shape}")
|
185 |
+
|
186 |
+
if processed_image is not None:
|
187 |
+
processed_image = processed_image.unsqueeze(0).to(device)
|
188 |
+
print(f"\033[92mDEBUG\033[0m: Final image shape with batch dimension: {processed_image.shape}")
|
189 |
+
|
190 |
+
# Run inference
|
191 |
+
print(f"\033[92mDEBUG\033[0m: Running inference on device: {device}")
|
192 |
+
if mfcc is not None and processed_image is not None:
|
193 |
+
with torch.no_grad():
|
194 |
+
sweetness = model(mfcc, processed_image)
|
195 |
+
print(f"\033[92mDEBUG\033[0m: Prediction successful: {sweetness.item()}")
|
196 |
+
else:
|
197 |
+
return "Error: Failed to process inputs. Please check the debug logs."
|
198 |
+
|
199 |
+
# Format the result
|
200 |
+
if sweetness is not None:
|
201 |
+
result = f"Predicted Sweetness: {sweetness.item():.2f}/13"
|
202 |
+
|
203 |
+
# Add a qualitative description
|
204 |
+
if sweetness.item() < 9:
|
205 |
+
result += "\n\nThis watermelon is not very sweet. You might want to choose another one."
|
206 |
+
elif sweetness.item() < 10:
|
207 |
+
result += "\n\nThis watermelon has moderate sweetness."
|
208 |
+
elif sweetness.item() < 11:
|
209 |
+
result += "\n\nThis watermelon is sweet! A good choice."
|
210 |
else:
|
211 |
+
result += "\n\nThis watermelon is very sweet! Excellent choice!"
|
|
|
|
|
|
|
|
|
212 |
|
213 |
+
return result
|
214 |
+
else:
|
215 |
+
return "Error: Could not predict sweetness. Please try again with different inputs."
|
216 |
+
|
217 |
+
except Exception as e:
|
218 |
+
import traceback
|
219 |
+
error_msg = f"Error: {str(e)}\n\n"
|
220 |
+
error_msg += traceback.format_exc()
|
221 |
+
print(f"\033[91mERR!\033[0m: {error_msg}")
|
222 |
+
return error_msg
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
223 |
|
224 |
print("\033[92mINFO\033[0m: GPU-accelerated prediction function created with @spaces.GPU decorator")
|
225 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
226 |
|
227 |
def create_app(model_path):
|
228 |
"""Create and launch the Gradio interface"""
|