WinFunction commited on
Commit
aee40b4
·
verified ·
1 Parent(s): 943eccd

Upload weights and the model architecture and some info files

Browse files
Files changed (6) hide show
  1. README.md +191 -0
  2. best_model.pt +3 -0
  3. config.json +28 -0
  4. modeling_f1tone.py +304 -0
  5. models.py +124 -0
  6. requirements.txt +5 -0
README.md CHANGED
@@ -1,3 +1,194 @@
1
  ---
 
 
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ language:
3
+ - en
4
  license: apache-2.0
5
+ tags:
6
+ - audio
7
+ - audio-classification
8
+ - emotion-recognition
9
+ - speech-emotion-recognition
10
+ - f1
11
+ - formula1
12
+ - wavlm
13
+ base_model: microsoft/wavlm-base-plus
14
+ pipeline_tag: audio-classification
15
+ metrics:
16
+ - accuracy
17
+ - f1
18
  ---
19
+
20
+ # 🏎️ F1 Driver Tone & Emotion Detector
21
+
22
+ An advanced Speech Emotion Recognition (SER) deep learning model designed to detect the emotional tone of Formula 1 driver team radio communications in real-time.
23
+
24
+ The model processes raw audio input from driver-to-pit-wall communications, extracts context-rich speech representations using a pretrained **WavLM** encoder (`microsoft/wavlm-base-plus`), and passes them through a downstream **BiLSTM with Temporal Attention mechanism** to classify driver emotions.
25
+
26
+ ---
27
+
28
+ ## Model Overview
29
+
30
+ - **Base Audio Encoder:** [microsoft/wavlm-base-plus](https://huggingface.co/microsoft/wavlm-base-plus) (768-dimensional frozen embeddings)
31
+ - **Downstream Architecture:** Bidirectional LSTM (128 hidden size) + Temporal Attention Module + Linear Classifier
32
+ - **Target Emotions (6 Classes):**
33
+ - `0`: **Anger**
34
+ - `1`: **Disgust**
35
+ - `2`: **Fear**
36
+ - `3`: **Happy**
37
+ - `4`: **Neutral**
38
+ - `5`: **Sad**
39
+
40
+ ---
41
+
42
+ ## Model Architecture
43
+
44
+ ```text
45
+ RAW DRIVER AUDIO (.wav, .mp3, .flac)
46
+ |
47
+ v
48
+ Audio Preprocessing Pipeline
49
+ - Convert Multi-channel to Mono
50
+ - Resample to 16,000 Hz (16 kHz)
51
+ - Peak Amplitude Normalization
52
+ |
53
+ v
54
+ Frozen WavLM Encoder
55
+ (microsoft/wavlm-base-plus)
56
+ |
57
+ v
58
+ Frame Embeddings [B, T, 768]
59
+ |
60
+ v
61
+ Bidirectional LSTM
62
+ (hidden_size = 128 x 2 = 256)
63
+ |
64
+ v
65
+ Temporal Attention
66
+ (Aggregates sequence into fixed context vector [B, 256])
67
+ |
68
+ v
69
+ Linear Classifier
70
+ (256 -> 6 Logits)
71
+ |
72
+ v
73
+ Softmax Probabilities & Emotion Output
74
+ ```
75
+
76
+ ---
77
+
78
+ ## 🎛️ Input & Output Specifications
79
+
80
+ ### 1️⃣ Direct PyTorch Model I/O (`models.py` + `best_model.pt`)
81
+ For users directly loading the PyTorch `BiLSTMAttentionClassifier` model and `.pt` weights:
82
+
83
+ - **Model Input:**
84
+ - `x`: PyTorch Tensor of WavLM frame embeddings with shape **`[Batch_Size, Sequence_Length, 768]`**
85
+ - `mask`: Timestep mask tensor with shape **`[Batch_Size, Sequence_Length]`** (Optional: 1 for valid frame, 0 for padding)
86
+ - **Model Output:**
87
+ - `logits`: Tensor of shape **`[Batch_Size, 6]`** (Unnormalized raw class logits)
88
+ - `attn_weights`: Attention weight tensor of shape **`[Batch_Size, Sequence_Length]`** (Frame-level temporal attention scores)
89
+
90
+ ---
91
+
92
+ ### 2️⃣ Full Pipeline I/O (Using `modeling_f1tone.py`)
93
+ For users using the complete audio-to-prediction helper script:
94
+
95
+ - **Input Requirements:**
96
+ - Raw Audio File (`.wav`, `.mp3`, `.flac`, `.ogg`, `.m4a`)
97
+ - Automatically converted to 16 kHz Mono and peak-normalized.
98
+ - **Output Return Format:**
99
+ Structured Python dictionary:
100
+ ```json
101
+ {
102
+ "audio_file": "driver_radio_01.wav",
103
+ "predicted_emotion": "Anger",
104
+ "confidence": 94.25,
105
+ "probabilities": {
106
+ "Anger": 94.25,
107
+ "Neutral": 3.10,
108
+ "Disgust": 1.45,
109
+ "Fear": 0.70,
110
+ "Happy": 0.30,
111
+ "Sad": 0.20
112
+ },
113
+ "attention_weights": [0.012, 0.045, 0.180, ...]
114
+ }
115
+ ```
116
+
117
+ ---
118
+
119
+ ## How to Run Inference using `modeling_f1tone.py`
120
+
121
+ ### Directory Setup (`input/` folder)
122
+ To run predictions on audio files:
123
+ 1. Create an `input/` folder in the directory where `modeling_f1tone.py` is located.
124
+ 2. Place your team radio `.wav` or `.mp3` files inside `input/`.
125
+
126
+ ```text
127
+ ToneDetectorF1/
128
+ ├── best_model.pt
129
+ ├── modeling_f1tone.py
130
+ ├── config.json
131
+ └── input/
132
+ ├── driver_radio_1.wav
133
+ └── pit_communication_2.mp3
134
+ ```
135
+
136
+ ### Running via Command Line
137
+
138
+ **Option A: Process all files in `input/` folder**
139
+ ```bash
140
+ python modeling_f1tone.py
141
+ ```
142
+
143
+ **Option B: Pass an explicit audio file path**
144
+ ```bash
145
+ python modeling_f1tone.py --audio_path path/to/my_audio.wav
146
+ ```
147
+
148
+ ### Python Code Integration Example
149
+ ```python
150
+ from modeling_f1tone import FinalPredictor
151
+
152
+ # Initialize predictor with model checkpoint
153
+ predictor = FinalPredictor(checkpoint_path="best_model.pt")
154
+
155
+ # Perform inference on an audio clip
156
+ result = predictor.predict_single("input/driver_radio_1.wav")
157
+
158
+ print(f"Predicted Emotion : {result['predicted_emotion']}")
159
+ print(f"Confidence : {result['confidence']:.2f}%")
160
+ print("Full Probabilities:", result['probabilities'])
161
+ ```
162
+
163
+ ---
164
+
165
+ ## 📊 Verified Model Evaluation Results
166
+
167
+ Evaluation performed on holdout test set (1,117 audio evaluation samples):
168
+
169
+ | Metric | Score |
170
+ | :--- | :--- |
171
+ | **Test Set Accuracy** | **67.86%** |
172
+ | **Test Set Macro F1-Score** | **0.6794** |
173
+ | **Validation Accuracy (Epoch 8)** | **69.89%** |
174
+ | **Validation Macro F1-Score** | **0.7006** |
175
+ | **Feature Encoder** | `microsoft/wavlm-base-plus` (768-dim, Frozen) |
176
+ | **Classifier Head** | BiLSTM (128) + Temporal Attention + Linear (6) |
177
+
178
+ ### 📈 Per-Class Performance Breakdown (Test Set)
179
+
180
+ | Emotion Class | Precision | Recall | F1-Score | Evaluation Support |
181
+ | :--- | :--- | :--- | :--- | :--- |
182
+ | **Anger** 😡 | **76.71%** | **87.96%** | **0.8195** | 191 samples |
183
+ | **Neutral** 😐 | **77.85%** | **75.46%** | **0.7664** | 163 samples |
184
+ | **Happy** 😊 | **79.05%** | **61.26%** | **0.6903** | 191 samples |
185
+ | **Disgust** 🤢 | **64.65%** | **67.02%** | **0.6581** | 191 samples |
186
+ | **Fear** 😨 | **52.05%** | **66.84%** | **0.5853** | 190 samples |
187
+ | **Sad** 🥺 | **63.33%** | **49.74%** | **0.5572** | 191 samples |
188
+
189
+ ---
190
+
191
+ ## Citation & License
192
+
193
+ - **License:** Apache 2.0
194
+ - **Base Encoder Paper:** *WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing* (Chen et al., 2022)
best_model.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5e258e689054804b02a2c037d1181d939055eaeb756955508ebd3a996fb65c69
3
+ size 11068789
config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "BiLSTMAttentionClassifier"
4
+ ],
5
+ "model_type": "wavlm-bilstm-attention",
6
+ "base_encoder": "microsoft/wavlm-base-plus",
7
+ "target_sample_rate": 16000,
8
+ "input_size": 768,
9
+ "hidden_size": 128,
10
+ "num_classes": 6,
11
+ "dropout": 0.3,
12
+ "id2label": {
13
+ "0": "Anger",
14
+ "1": "Disgust",
15
+ "2": "Fear",
16
+ "3": "Happy",
17
+ "4": "Neutral",
18
+ "5": "Sad"
19
+ },
20
+ "label2id": {
21
+ "Anger": 0,
22
+ "Disgust": 1,
23
+ "Fear": 2,
24
+ "Happy": 3,
25
+ "Neutral": 4,
26
+ "Sad": 5
27
+ }
28
+ }
modeling_f1tone.py ADDED
@@ -0,0 +1,304 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import sys
3
+ from pathlib import Path
4
+ import torch
5
+ import torch.nn as nn
6
+ import torch.nn.functional as F
7
+ from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
8
+ import torchaudio
9
+ from transformers import WavLMModel
10
+
11
+ # Ensure UTF-8 output encoding for Windows terminal compatibility
12
+ if sys.stdout.encoding and sys.stdout.encoding.lower() != 'utf-8':
13
+ try:
14
+ sys.stdout.reconfigure(encoding='utf-8')
15
+ except Exception:
16
+ pass
17
+
18
+ # ============================================================
19
+ # CONFIGURATION & CONSTANTS
20
+ # ============================================================
21
+
22
+ WAVLM_MODEL_NAME = "microsoft/wavlm-base-plus"
23
+ CHECKPOINT_PATH = Path("checkpoints/best_model.pt")
24
+ INPUT_DIR = Path("input")
25
+ TARGET_SR = 16000
26
+
27
+ EMOTION_CLASSES = ["Anger", "Disgust", "Fear", "Happy", "Neutral", "Sad"]
28
+
29
+ # ============================================================
30
+ # MODEL ARCHITECTURE DEFINITION (Self-Contained Single File)
31
+ # ============================================================
32
+
33
+ class BiLSTMFeatureExtractor(nn.Module):
34
+ def __init__(self, input_size=768, hidden_size=128, num_layers=1, dropout=0.0):
35
+ super(BiLSTMFeatureExtractor, self).__init__()
36
+ self.input_size = input_size
37
+ self.hidden_size = hidden_size
38
+ self.num_layers = num_layers
39
+ self.bidirectional = True
40
+
41
+ self.bilstm = nn.LSTM(
42
+ input_size=input_size,
43
+ hidden_size=hidden_size,
44
+ num_layers=num_layers,
45
+ batch_first=True,
46
+ bidirectional=True,
47
+ dropout=dropout if num_layers > 1 else 0.0
48
+ )
49
+
50
+ def forward(self, x, mask=None):
51
+ batch_size, seq_len, _ = x.shape
52
+
53
+ if mask is not None:
54
+ lengths = mask.sum(dim=1).cpu()
55
+ packed_x = pack_padded_sequence(
56
+ x,
57
+ lengths,
58
+ batch_first=True,
59
+ enforce_sorted=False
60
+ )
61
+ packed_out, (hn, cn) = self.bilstm(packed_x)
62
+ out, _ = pad_packed_sequence(
63
+ packed_out,
64
+ batch_first=True,
65
+ total_length=seq_len
66
+ )
67
+ else:
68
+ out, (hn, cn) = self.bilstm(x)
69
+
70
+ return out
71
+
72
+
73
+ class TemporalAttention(nn.Module):
74
+ def __init__(self, input_dim=256):
75
+ super(TemporalAttention, self).__init__()
76
+ self.input_dim = input_dim
77
+ self.w = nn.Linear(input_dim, 1, bias=False)
78
+
79
+ def forward(self, h, mask=None):
80
+ scores = self.w(torch.tanh(h)).squeeze(-1) # [B, T]
81
+
82
+ if mask is not None:
83
+ scores = scores.masked_fill(mask == 0, -1e9)
84
+
85
+ attn_weights = torch.softmax(scores, dim=1) # [B, T]
86
+ context = torch.bmm(attn_weights.unsqueeze(1), h).squeeze(1) # [B, 256]
87
+
88
+ return context, attn_weights
89
+
90
+
91
+ class BiLSTMAttentionClassifier(nn.Module):
92
+ def __init__(self, input_size=768, hidden_size=128, num_classes=6, dropout=0.3):
93
+ super(BiLSTMAttentionClassifier, self).__init__()
94
+ self.bilstm = BiLSTMFeatureExtractor(
95
+ input_size=input_size,
96
+ hidden_size=hidden_size,
97
+ num_layers=1,
98
+ dropout=dropout
99
+ )
100
+ context_dim = hidden_size * 2 # 256
101
+ self.attention = TemporalAttention(input_dim=context_dim)
102
+ self.dropout = nn.Dropout(dropout)
103
+ self.classifier = nn.Linear(context_dim, num_classes)
104
+
105
+ def forward(self, x, mask=None):
106
+ bilstm_out = self.bilstm(x, mask=mask)
107
+ context, attn_weights = self.attention(bilstm_out, mask=mask)
108
+ dropped_context = self.dropout(context)
109
+ logits = self.classifier(dropped_context)
110
+ return logits, attn_weights
111
+
112
+ # ============================================================
113
+ # EXACT AUDIO PREPROCESSING LOGIC (Used During Training)
114
+ # ============================================================
115
+
116
+ def preprocess_audio(audio_path):
117
+ """
118
+ Load raw audio file (.wav, .mp3, .flac, .ogg), convert to 16 kHz mono, and peak-normalize amplitude.
119
+ """
120
+ audio_path = Path(audio_path)
121
+
122
+ # 1. Load audio with torchaudio, fallback to soundfile for mp3/flac if needed
123
+ try:
124
+ waveform, sample_rate = torchaudio.load(str(audio_path))
125
+ except Exception:
126
+ import soundfile as sf
127
+ data, sample_rate = sf.read(str(audio_path))
128
+ waveform = torch.tensor(data, dtype=torch.float32)
129
+ if waveform.ndim == 1:
130
+ waveform = waveform.unsqueeze(0)
131
+ elif waveform.ndim == 2:
132
+ waveform = waveform.T
133
+
134
+ # 2. Convert to mono if multi-channel
135
+ if waveform.shape[0] > 1:
136
+ waveform = waveform.mean(dim=0, keepdim=True)
137
+
138
+ # 3. Resample to 16 kHz if necessary
139
+ if sample_rate != TARGET_SR:
140
+ resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=TARGET_SR)
141
+ waveform = resampler(waveform)
142
+
143
+ # 4. Squeeze channel dimension [1, num_samples] -> [num_samples]
144
+ waveform = waveform.squeeze(0)
145
+
146
+ # 5. Amplitude peak normalization
147
+ max_val = waveform.abs().max()
148
+ if max_val > 0:
149
+ waveform = waveform / max_val
150
+
151
+ return waveform
152
+
153
+ # ============================================================
154
+ # FINAL PREDICTOR PIPELINE CLASS
155
+ # ============================================================
156
+
157
+ class FinalPredictor:
158
+ def __init__(self, checkpoint_path=CHECKPOINT_PATH):
159
+ self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
160
+ print("=" * 60)
161
+ print("INITIALIZING F1 DRIVER TONE DETECTOR PIPELINE")
162
+ print("=" * 60)
163
+ print(f"Compute Device : {self.device}")
164
+ if self.device.type == "cuda":
165
+ print(f"GPU : {torch.cuda.get_device_name(0)}")
166
+
167
+ print(f"\n[1] Loading WavLM Encoder ({WAVLM_MODEL_NAME})...")
168
+ self.wavlm = WavLMModel.from_pretrained(WAVLM_MODEL_NAME).to(self.device)
169
+ self.wavlm.eval()
170
+
171
+ print(f"[2] Loading Trained Downstream Model ({checkpoint_path})...")
172
+ if not checkpoint_path.exists():
173
+ raise FileNotFoundError(f"Checkpoint file not found at {checkpoint_path}. Please run training first!")
174
+
175
+ self.classifier = BiLSTMAttentionClassifier(
176
+ input_size=768,
177
+ hidden_size=128,
178
+ num_classes=6
179
+ ).to(self.device)
180
+
181
+ checkpoint = torch.load(checkpoint_path, map_location=self.device, weights_only=False)
182
+ self.classifier.load_state_dict(checkpoint["model_state_dict"])
183
+ self.classifier.eval()
184
+ print("Pipeline initialized and ready for inference!")
185
+ print("=" * 60)
186
+
187
+ def predict_single(self, audio_path, chunk_duration=2.5, hop_duration=1.5):
188
+ audio_path = Path(audio_path)
189
+ if not audio_path.exists():
190
+ raise FileNotFoundError(f"Audio file not found: {audio_path}")
191
+
192
+ # 1. Preprocess raw audio waveform [num_samples]
193
+ waveform = preprocess_audio(audio_path)
194
+ num_samples = waveform.size(0)
195
+ duration_sec = num_samples / float(TARGET_SR)
196
+
197
+ chunk_samples = int(chunk_duration * TARGET_SR)
198
+ hop_samples = int(hop_duration * TARGET_SR)
199
+
200
+ chunk_probs = []
201
+ last_attn_weights = None
202
+
203
+ with torch.no_grad():
204
+ # If audio is long (> 4.0s), use 2.5s sliding window chunking
205
+ if duration_sec > 4.0 and num_samples > chunk_samples:
206
+ starts = list(range(0, num_samples - chunk_samples + 1, hop_samples))
207
+ # Ensure the end of audio is covered
208
+ if starts[-1] + chunk_samples < num_samples:
209
+ starts.append(num_samples - chunk_samples)
210
+
211
+ for start in starts:
212
+ end = start + chunk_samples
213
+ chunk_wave = waveform[start:end].unsqueeze(0).to(self.device)
214
+
215
+ outputs = self.wavlm(input_values=chunk_wave)
216
+ embedding = outputs.last_hidden_state
217
+ mask = torch.ones((1, embedding.size(1)), dtype=torch.int64, device=self.device)
218
+
219
+ logits, attn_weights = self.classifier(embedding, mask=mask)
220
+ probs = F.softmax(logits, dim=-1).squeeze(0)
221
+ chunk_probs.append(probs)
222
+ last_attn_weights = attn_weights.squeeze(0).cpu().numpy()
223
+
224
+ probabilities = torch.stack(chunk_probs).mean(dim=0)
225
+ else:
226
+ input_values = waveform.unsqueeze(0).to(self.device)
227
+ outputs = self.wavlm(input_values=input_values)
228
+ embedding = outputs.last_hidden_state
229
+ mask = torch.ones((1, embedding.size(1)), dtype=torch.int64, device=self.device)
230
+ logits, attn_weights = self.classifier(embedding, mask=mask)
231
+ probabilities = F.softmax(logits, dim=-1).squeeze(0)
232
+ last_attn_weights = attn_weights.squeeze(0).cpu().numpy()
233
+
234
+ pred_id = torch.argmax(probabilities).item()
235
+ pred_emotion = EMOTION_CLASSES[pred_id]
236
+ confidence = probabilities[pred_id].item() * 100.0
237
+
238
+ probs_dict = {
239
+ EMOTION_CLASSES[i]: probabilities[i].item() * 100.0
240
+ for i in range(len(EMOTION_CLASSES))
241
+ }
242
+
243
+ return {
244
+ "audio_file": audio_path.name,
245
+ "duration_sec": round(duration_sec, 2),
246
+ "predicted_emotion": pred_emotion,
247
+ "confidence": confidence,
248
+ "probabilities": probs_dict,
249
+ "attention_weights": last_attn_weights,
250
+ "chunks_processed": len(chunk_probs) if chunk_probs else 1
251
+ }
252
+
253
+ # ============================================================
254
+ # MAIN CLI DRIVER
255
+ # ============================================================
256
+
257
+ def print_report(result):
258
+ print("\n" + "=" * 60)
259
+ print(f"PREDICTION REPORT: {result['audio_file']}")
260
+ print("=" * 60)
261
+ print(f"Predicted Emotion : {result['predicted_emotion']} ({result['confidence']:.2f}% confidence)")
262
+ print("-" * 60)
263
+ print("EMOTION PROBABILITY BREAKDOWN:")
264
+ for emotion, prob in result["probabilities"].items():
265
+ bar = "#" * int(prob / 5)
266
+ print(f" {emotion:10s} : {prob:6.2f}% | {bar}")
267
+ print("=" * 60)
268
+
269
+
270
+ def main():
271
+ parser = argparse.ArgumentParser(description="F1 Driver Tone Predictor (Single File Pipeline)")
272
+ parser.add_argument("--audio_path", type=str, default=None,
273
+ help="Path to specific raw driver audio file (.wav, .mp3, .flac, .ogg)")
274
+ args = parser.parse_args()
275
+
276
+ predictor = FinalPredictor()
277
+
278
+ if args.audio_path:
279
+ target_path = Path(args.audio_path)
280
+ result = predictor.predict_single(target_path)
281
+ print_report(result)
282
+ else:
283
+ # Scan input directory
284
+ INPUT_DIR.mkdir(parents=True, exist_ok=True)
285
+ valid_extensions = {".wav", ".mp3", ".flac", ".ogg", ".m4a"}
286
+ audio_files = [
287
+ f for f in INPUT_DIR.iterdir()
288
+ if f.is_file() and f.suffix.lower() in valid_extensions
289
+ ]
290
+
291
+ if not audio_files:
292
+ print(f"\n[INFO] No audio files found in input/ directory ({INPUT_DIR.resolve()}).")
293
+ print("Usage Options:")
294
+ print(" 1. Place .wav or .mp3 files into the 'input/' folder and re-run python finalpredictor.py")
295
+ print(" 2. Run with explicit file path: python finalpredictor.py --audio_path <path_to_audio_file>")
296
+ else:
297
+ print(f"\nFound {len(audio_files)} audio file(s) in {INPUT_DIR}/ directory. Processing...")
298
+ for audio_file in audio_files:
299
+ result = predictor.predict_single(audio_file)
300
+ print_report(result)
301
+
302
+
303
+ if __name__ == "__main__":
304
+ main()
models.py ADDED
@@ -0,0 +1,124 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
4
+
5
+
6
+ class BiLSTMFeatureExtractor(nn.Module):
7
+ """
8
+ BiLSTM Feature Extractor module.
9
+
10
+ Inputs:
11
+ x : Padded WavLM sequence embeddings [B, T, 768]
12
+ mask : Timestep mask [B, T] (optional, 1 for real timesteps, 0 for padding)
13
+
14
+ Outputs:
15
+ out : BiLSTM contextual representations [B, T, 256]
16
+ """
17
+ def __init__(self, input_size=768, hidden_size=128, num_layers=1, dropout=0.0):
18
+ super(BiLSTMFeatureExtractor, self).__init__()
19
+
20
+ self.input_size = input_size
21
+ self.hidden_size = hidden_size
22
+ self.num_layers = num_layers
23
+ self.bidirectional = True
24
+
25
+ # BiLSTM Layer: hidden_size=128, bidirectional=True -> Output dim = 128 * 2 = 256
26
+ self.bilstm = nn.LSTM(
27
+ input_size=input_size,
28
+ hidden_size=hidden_size,
29
+ num_layers=num_layers,
30
+ batch_first=True,
31
+ bidirectional=True,
32
+ dropout=dropout if num_layers > 1 else 0.0
33
+ )
34
+
35
+ def forward(self, x, mask=None):
36
+ batch_size, seq_len, _ = x.shape
37
+
38
+ if mask is not None:
39
+ lengths = mask.sum(dim=1).cpu()
40
+
41
+ packed_x = pack_padded_sequence(
42
+ x,
43
+ lengths,
44
+ batch_first=True,
45
+ enforce_sorted=False
46
+ )
47
+
48
+ packed_out, (hn, cn) = self.bilstm(packed_x)
49
+
50
+ out, _ = pad_packed_sequence(
51
+ packed_out,
52
+ batch_first=True,
53
+ total_length=seq_len
54
+ )
55
+ else:
56
+ out, (hn, cn) = self.bilstm(x)
57
+
58
+ return out
59
+
60
+
61
+ class TemporalAttention(nn.Module):
62
+ """
63
+ Temporal Attention Module.
64
+ Learns frame-wise attention scores over sequence length T to aggregate
65
+ BiLSTM output sequence [B, T, 256] into a fixed 256-dimensional context vector [B, 256].
66
+
67
+ Inputs:
68
+ h : BiLSTM outputs [B, T, input_dim] (input_dim=256)
69
+ mask : Timestep mask [B, T] (1 for real timesteps, 0 for padding)
70
+
71
+ Outputs:
72
+ context : Aggregated context vector [B, 256]
73
+ attn_weights : Normalized attention weights [B, T]
74
+ """
75
+ def __init__(self, input_dim=256):
76
+ super(TemporalAttention, self).__init__()
77
+ self.input_dim = input_dim
78
+ self.w = nn.Linear(input_dim, 1, bias=False)
79
+
80
+ def forward(self, h, mask=None):
81
+ scores = self.w(torch.tanh(h)).squeeze(-1) # [B, T]
82
+
83
+ if mask is not None:
84
+ scores = scores.masked_fill(mask == 0, -1e9)
85
+
86
+ attn_weights = torch.softmax(scores, dim=1) # [B, T]
87
+ context = torch.bmm(attn_weights.unsqueeze(1), h).squeeze(1) # [B, 256]
88
+
89
+ return context, attn_weights
90
+
91
+
92
+ class BiLSTMAttentionClassifier(nn.Module):
93
+ """
94
+ Full Downstream Model Architecture:
95
+ WavLM Embedding [B, T, 768] -> BiLSTM -> [B, T, 256] -> Temporal Attention -> [B, 256] -> Linear Classifier -> [B, 6] Logits
96
+ """
97
+ def __init__(self, input_size=768, hidden_size=128, num_classes=6, dropout=0.3):
98
+ super(BiLSTMAttentionClassifier, self).__init__()
99
+
100
+ self.bilstm = BiLSTMFeatureExtractor(
101
+ input_size=input_size,
102
+ hidden_size=hidden_size,
103
+ num_layers=1,
104
+ dropout=dropout
105
+ )
106
+
107
+ context_dim = hidden_size * 2 # 256
108
+ self.attention = TemporalAttention(input_dim=context_dim)
109
+
110
+ self.dropout = nn.Dropout(dropout)
111
+ self.classifier = nn.Linear(context_dim, num_classes)
112
+
113
+ def forward(self, x, mask=None):
114
+ # 1. BiLSTM: [B, T, 768] -> [B, T, 256]
115
+ bilstm_out = self.bilstm(x, mask=mask)
116
+
117
+ # 2. Temporal Attention: [B, T, 256] -> [B, 256]
118
+ context, attn_weights = self.attention(bilstm_out, mask=mask)
119
+
120
+ # 3. Dropout + Linear Classifier: [B, 256] -> [B, 6]
121
+ dropped_context = self.dropout(context)
122
+ logits = self.classifier(dropped_context)
123
+
124
+ return logits, attn_weights
requirements.txt ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ torch>=2.0.0
2
+ torchaudio>=2.0.0
3
+ transformers>=4.30.0
4
+ soundfile>=0.12.1
5
+ numpy>=1.22.0