File size: 13,982 Bytes
00826e1
d23e88c
aaa4ee3
 
00826e1
 
 
 
 
 
 
 
a46e52c
ed9b2d0
00826e1
a46e52c
 
ed9b2d0
a46e52c
00826e1
 
aaa4ee3
00826e1
 
aaa4ee3
 
00826e1
 
 
 
 
 
 
 
 
 
 
 
aaa4ee3
 
00826e1
 
 
 
 
 
 
 
 
 
a46e52c
 
aaa4ee3
 
00826e1
ed9b2d0
a46e52c
00826e1
a46e52c
ed9b2d0
a46e52c
 
 
 
 
ed9b2d0
 
a46e52c
 
 
 
 
 
ed9b2d0
00826e1
ed9b2d0
00826e1
ed9b2d0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
00826e1
 
ed9b2d0
a46e52c
 
 
 
 
 
 
ed9b2d0
a46e52c
ed9b2d0
a46e52c
 
ed9b2d0
a46e52c
ed9b2d0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a46e52c
ed9b2d0
a46e52c
00826e1
ed9b2d0
00826e1
 
ed9b2d0
00826e1
 
 
 
 
 
 
 
a46e52c
 
 
 
ed9b2d0
00826e1
ed9b2d0
00826e1
ed9b2d0
00826e1
 
 
 
ed9b2d0
a46e52c
 
 
 
 
 
00826e1
 
 
 
 
ed9b2d0
a46e52c
 
 
 
00826e1
a46e52c
 
 
 
 
 
 
 
 
 
 
00826e1
 
 
 
 
 
 
aaa4ee3
00826e1
ed9b2d0
a46e52c
 
 
 
aaa4ee3
a46e52c
 
 
 
 
 
 
 
 
 
 
 
ed9b2d0
 
a46e52c
 
ed9b2d0
 
a46e52c
 
 
ed9b2d0
a46e52c
00826e1
ed9b2d0
a46e52c
00826e1
 
 
 
 
 
 
 
aaa4ee3
a46e52c
 
 
 
 
 
 
 
 
aaa4ee3
a46e52c
 
 
 
 
00826e1
aaa4ee3
 
 
 
a46e52c
 
 
aaa4ee3
 
 
 
 
a46e52c
 
 
 
aaa4ee3
a46e52c
 
 
 
 
 
 
 
aaa4ee3
a46e52c
 
00826e1
a46e52c
 
 
00826e1
 
aaa4ee3
 
 
 
00826e1
ed9b2d0
00826e1
 
aaa4ee3
00826e1
 
ed9b2d0
00826e1
 
aaa4ee3
 
00826e1
aaa4ee3
00826e1
ed9b2d0
00826e1
aaa4ee3
00826e1
aaa4ee3
 
ed9b2d0
 
 
 
 
 
00826e1
 
 
 
 
 
 
aaa4ee3
00826e1
 
 
 
 
 
d23e88c
aaa4ee3
 
 
00826e1
d23e88c
 
aaa4ee3
d23e88c
ed9b2d0
d23e88c
 
 
a46e52c
 
 
 
aaa4ee3
a46e52c
aaa4ee3
a46e52c
 
 
 
 
 
d23e88c
ed9b2d0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
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
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
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
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
"""
Audio Preprocessing Module for Respiratory Symptom Analysis
Updated for 39% F1-Macro Model (128x431 mel-spectrograms)
Version: 3.0.0
"""

import librosa
import numpy as np
import torch
import warnings
from typing import Union, Tuple, Dict
import soundfile as sf
import os
from scipy import signal

# Fix for Numba caching issues in Docker containers
os.environ['NUMBA_CACHE_DIR'] = '/tmp'
os.environ['NUMBA_DISABLE_JIT'] = '0'

warnings.filterwarnings('ignore')


class RespiratoryAudioPreprocessor:
    """
    Audio preprocessor matching your 39% F1-Macro training pipeline
    Mel-spectrogram shape: (128, 431) to match training data
    """
    
    def __init__(self, 
                 target_sr: int = 22050,
                 n_mels: int = 128,
                 n_fft: int = 2048,
                 hop_length: int = 512,
                 win_length: int = None,
                 window: str = 'hann',
                 fmin: float = 0.0,
                 fmax: float = None,
                 power: float = 2.0,
                 duration: float = 10.0):  # Changed from 3.0 to 10.0 to match training
        """Initialize preprocessing parameters to match training"""
        self.target_sr = target_sr
        self.n_mels = n_mels
        self.n_fft = n_fft
        self.hop_length = hop_length
        self.win_length = win_length
        self.window = window
        self.fmin = fmin
        self.fmax = fmax or target_sr // 2
        self.power = power
        self.duration = duration
        self.target_length = int(target_sr * duration)
        
        # Expected output shape - UPDATED to match training (128, 431)
        self.expected_shape = (1, 1, 128, 431)
        
        # Pre-warm librosa
        self._warmup_librosa()
        
    def _warmup_librosa(self):
        """Pre-compile librosa functions"""
        try:
            dummy_audio = np.random.randn(1024).astype(np.float32)
            _ = librosa.feature.melspectrogram(
                y=dummy_audio,
                sr=self.target_sr,
                n_mels=32,
                n_fft=512,
                hop_length=256
            )
            print("✅ Librosa functions warmed up successfully")
        except Exception as e:
            print(f"⚠️ Librosa warmup warning: {str(e)}")
    
    def scipy_resample(self, audio_data: np.ndarray, orig_sr: int, target_sr: int) -> np.ndarray:
        """
        Custom resampling using scipy.signal instead of resampy
        """
        if orig_sr == target_sr:
            return audio_data
        
        try:
            # Calculate resampling ratio
            resample_ratio = target_sr / orig_sr
            
            # Use scipy.signal.resample for resampling
            target_length = int(len(audio_data) * resample_ratio)
            resampled_audio = signal.resample(audio_data, target_length)
            
            return resampled_audio.astype(np.float32)
            
        except Exception as e:
            print(f"⚠️ Scipy resampling failed: {e}, using original audio")
            return audio_data
    
    def load_and_normalize_audio(self, audio_input: Union[str, np.ndarray, tuple]) -> np.ndarray:
        """Load audio file without resampy dependency"""
        try:
            if isinstance(audio_input, str):
                # Load with soundfile first
                try:
                    audio_data, sr = sf.read(audio_input)
                    
                    # Convert to mono if stereo
                    if len(audio_data.shape) > 1:
                        audio_data = np.mean(audio_data, axis=1)
                    
                    # Resample using scipy if needed
                    if sr != self.target_sr:
                        audio_data = self.scipy_resample(audio_data, sr, self.target_sr)
                        
                except Exception as sf_error:
                    # Fallback: try loading without librosa resampling
                    try:
                        # Load with original sample rate first
                        audio_data, sr = librosa.load(audio_input, sr=None)
                        
                        # Convert to mono if stereo
                        if len(audio_data.shape) > 1:
                            audio_data = np.mean(audio_data, axis=1)
                        
                        # Manual resampling with scipy
                        if sr != self.target_sr:
                            audio_data = self.scipy_resample(audio_data, sr, self.target_sr)
                        
                        # Limit duration manually
                        if len(audio_data) > self.target_length:
                            audio_data = audio_data[:self.target_length]
                            
                    except Exception as librosa_error:
                        raise RuntimeError(f"Failed to load audio. SoundFile: {sf_error}. Librosa: {librosa_error}")
                    
            elif isinstance(audio_input, tuple):
                # (sample_rate, audio_array) from uploads
                sr, audio_data = audio_input
                
                # Convert to float32
                if audio_data.dtype != np.float32:
                    if audio_data.dtype == np.int16:
                        audio_data = audio_data.astype(np.float32) / 32767.0
                    elif audio_data.dtype == np.int32:
                        audio_data = audio_data.astype(np.float32) / 2147483647.0
                    else:
                        audio_data = audio_data.astype(np.float32)
                
                # Convert to mono if stereo
                if len(audio_data.shape) > 1:
                    audio_data = np.mean(audio_data, axis=1)
                
                # Resample using scipy
                if sr != self.target_sr:
                    audio_data = self.scipy_resample(audio_data, sr, self.target_sr)
                
                # Trim duration
                if len(audio_data) > self.target_length:
                    audio_data = audio_data[:self.target_length]
                    
            elif isinstance(audio_input, np.ndarray):
                # Raw audio array (assume target_sr)
                audio_data = audio_input.astype(np.float32)
                
                # Convert to mono if stereo
                if len(audio_data.shape) > 1:
                    audio_data = np.mean(audio_data, axis=1)
                    
                if len(audio_data) > self.target_length:
                    audio_data = audio_data[:self.target_length]
            else:
                raise ValueError(f"Unsupported audio input type: {type(audio_input)}")
            
            # Ensure 1D
            if len(audio_data.shape) > 1:
                audio_data = audio_data.flatten()
            
            # Pad if too short
            if len(audio_data) < self.target_length:
                audio_data = np.pad(
                    audio_data, 
                    (0, self.target_length - len(audio_data)), 
                    mode='constant', 
                    constant_values=0
                )
            
            # Normalize amplitude
            max_val = np.max(np.abs(audio_data))
            if max_val > 0:
                audio_data = audio_data / max_val
            
            return audio_data
            
        except Exception as e:
            raise RuntimeError(f"Failed to load audio: {str(e)}")
    
    def extract_mel_spectrogram(self, audio_data: np.ndarray) -> np.ndarray:
        """Extract mel spectrogram matching training configuration"""
        try:
            # Ensure proper format
            audio_data = np.asarray(audio_data, dtype=np.float32)
            if len(audio_data.shape) > 1:
                audio_data = audio_data.flatten()
            
            # Extract mel spectrogram with exact training parameters
            try:
                mel_spec = librosa.feature.melspectrogram(
                    y=audio_data,
                    sr=self.target_sr,
                    n_mels=self.n_mels,
                    n_fft=self.n_fft,
                    hop_length=self.hop_length,
                    win_length=self.win_length,
                    window=self.window,
                    fmin=self.fmin,
                    fmax=self.fmax,
                    power=self.power,
                    center=True,
                    pad_mode='constant'
                )
            except Exception as mel_error:
                # Simplified fallback
                print(f"⚠️ Using simplified mel spectrogram extraction: {mel_error}")
                mel_spec = librosa.feature.melspectrogram(
                    y=audio_data,
                    sr=self.target_sr,
                    n_mels=self.n_mels
                )
            
            # Convert to dB
            mel_spec = np.maximum(mel_spec, 1e-10)
            mel_spec_db = librosa.power_to_db(mel_spec, ref=np.max)
            
            return mel_spec_db
            
        except Exception as e:
            raise RuntimeError(f"Failed to extract mel spectrogram: {str(e)}")
    
    def normalize_spectrogram(self, mel_spec: np.ndarray) -> np.ndarray:
        """Normalize spectrogram to match training"""
        try:
            mean = np.mean(mel_spec)
            std = np.std(mel_spec)
            
            if std == 0:
                normalized = mel_spec - mean
            else:
                normalized = (mel_spec - mean) / (std + 1e-8)
            
            # Clip to prevent extreme values
            normalized = np.clip(normalized, -5.0, 5.0)
            return normalized
            
        except Exception as e:
            raise RuntimeError(f"Failed to normalize spectrogram: {str(e)}")
    
    def resize_spectrogram(self, mel_spec: np.ndarray, target_width: int = 431) -> np.ndarray:
        """
        Resize spectrogram to target dimensions (128, 431) to match training
        """
        try:
            current_height, current_width = mel_spec.shape
            
            # Handle height (should be 128 already)
            if current_height != 128:
                print(f"⚠️ Unexpected height: {current_height}, expected 128")
            
            # Handle width
            if current_width == target_width:
                return mel_spec
            
            if current_width < target_width:
                # Pad to target width
                pad_width = target_width - current_width
                mel_spec = np.pad(
                    mel_spec, 
                    ((0, 0), (0, pad_width)), 
                    mode='constant', 
                    constant_values=0
                )
            else:
                # Crop to target width
                mel_spec = mel_spec[:, :target_width]
            
            return mel_spec
            
        except Exception as e:
            raise RuntimeError(f"Failed to resize spectrogram: {str(e)}")
    
    def preprocess_audio(self, audio_input: Union[str, np.ndarray, tuple]) -> torch.Tensor:
        """
        Complete preprocessing pipeline matching your training
        Output: (1, 1, 128, 431) tensor
        """
        try:
            # Load audio
            audio_data = self.load_and_normalize_audio(audio_input)
            
            # Extract mel-spectrogram
            mel_spec = self.extract_mel_spectrogram(audio_data)
            
            # Normalize
            mel_spec_norm = self.normalize_spectrogram(mel_spec)
            
            # Resize to (128, 431)
            mel_spec_resized = self.resize_spectrogram(mel_spec_norm, target_width=431)
            
            # Convert to tensor (1, 1, 128, 431)
            tensor_input = torch.FloatTensor(mel_spec_resized)
            tensor_input = tensor_input.unsqueeze(0).unsqueeze(0)
            
            # Verify shape
            if tensor_input.shape != self.expected_shape:
                print(f"⚠️ Shape mismatch: got {tensor_input.shape}, expected {self.expected_shape}")
                # Force resize using interpolation as last resort
                tensor_input = torch.nn.functional.interpolate(
                    tensor_input, 
                    size=self.expected_shape[2:], 
                    mode='bilinear', 
                    align_corners=False
                )
            
            return tensor_input
            
        except Exception as e:
            raise RuntimeError(f"Preprocessing failed: {str(e)}")
    
    def get_preprocessing_info(self) -> Dict:
        """Get preprocessing configuration info"""
        return {
            'target_sr': self.target_sr,
            'n_mels': self.n_mels,
            'n_fft': self.n_fft,
            'hop_length': self.hop_length,
            'duration': self.duration,
            'output_shape': self.expected_shape,
            'resampling_method': 'scipy.signal',
            'normalization': 'z-score (mean=0, std=1)',
            'db_scale': True
        }
    
    def validate_audio_file(self, audio_path: str) -> Tuple[bool, str]:
        """Validate audio file before processing"""
        try:
            if not audio_path:
                return False, "No audio file provided"
            
            try:
                info = sf.info(audio_path)
                duration = info.duration
                
                if duration < 0.5:
                    return False, f"Audio too short ({duration:.1f}s). Minimum: 0.5s"
                if duration > 30.0:
                    return False, f"Audio too long ({duration:.1f}s). Maximum: 30s"
                    
                return True, "Audio file is valid"
                
            except Exception as e:
                return False, f"Error validating audio: {str(e)}"
                
        except Exception as e:
            return False, f"Validation error: {str(e)}"