Create handler.py

handler.py

@@ -0,0 +1,38 @@
+# handler.py
+from models import StudentForAudioClassification
+import torch
+import torchaudio
+
+# 1. W2V extractor: 실제 사용한 W2V로 교체! (아래는 예시)
+bundle = torchaudio.pipelines.WAV2VEC2_BASE
+w2v_model = bundle.get_model()
+w2v_model.eval()
+
+def preprocess(audio_bytes):
+    # 1. 오디오 파일을 waveform으로 로드 (16kHz 변환)
+    import io
+    waveform, orig_sr = torchaudio.load(io.BytesIO(audio_bytes))
+    waveform = waveform.mean(dim=0, keepdim=True)  # mono 변환
+    if orig_sr != 16000:
+        resampler = torchaudio.transforms.Resample(orig_sr, 16000)
+        waveform = resampler(waveform)
+
+    # 2. W2V embedding 추출 (여기선 mean pooling)
+    with torch.no_grad():
+        features = w2v_model(waveform)[0]  # (1, T, 512)
+        x_w2v = features.mean(dim=1)      # (1, 512)
+    return x_w2v
+
+def inference(model, inputs):
+    with torch.no_grad():
+        outputs = model(inputs)
+        probs = torch.softmax(outputs.logits, dim=-1)
+    return {
+        "probabilities": probs.squeeze(0).tolist(),  # [not_stressed_prob, stressed_prob]
+        "label": int(probs.argmax(dim=-1)[0])
+    }
+
+def init():
+    model = StudentForAudioClassification.from_pretrained(".", trust_remote_code=True)
+    model.eval()
+    return model