we can start now

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+root_env filter=lfs diff=lfs merge=lfs -text

app.py

@@ -1,11 +1,22 @@
-from fastapi import FastAPI
+from fastapi import FastAPI, File, UploadFile
+from utils.preprocessing import load_audio
+from utils.inference import get_embeddings, predict_risk
 
 app = FastAPI()
 
 @app.get("/")
-def greet_json():
+def hello():
     return {"Hello": "World!"}
 
 
-
-# To run the app, use the command:
+@app.post("/predict")
+async def predict(file: UploadFile = File(...)):
+    temp_path = f"/tmp/{file.filename}"
+    with open(temp_path, "wb") as f:
+        f.write(await file.read())
+    
+    waveform, sr = load_audio(temp_path)
+    embeddings = get_embeddings(waveform, sr)
+    score = predict_risk(embeddings)
+    
+    return {"risk_score": score}

requirements.txt

@@ -1,2 +1,6 @@
 fastapi
 uvicorn[standard]
+torch
+torchaudio
+numpy
+transformers

utils/inference.py

@@ -0,0 +1,27 @@
+import torch
+from transformers import AutoFeatureExtractor, AutoModel
+
+# Charger HeAR depuis Hugging Face
+MODEL_NAME = "google/hear"
+feature_extractor = AutoFeatureExtractor.from_pretrained(MODEL_NAME)
+hear_model = AutoModel.from_pretrained(MODEL_NAME)
+hear_model.eval()
+
+def get_embeddings(waveform, sr):
+    """
+    Transforme un audio en embeddings HeAR
+    """
+    # Transformer waveform en input pour HeAR
+    inputs = feature_extractor(waveform, sampling_rate=sr, return_tensors="pt")
+    with torch.no_grad():
+        embeddings = hear_model(**inputs).last_hidden_state
+    # embeddings : (batch, sequence, features)
+    return embeddings
+
+def predict_risk(embeddings):
+    """
+    Pour le moment : score factice basé sur la norme L2 des embeddings
+    Plus tard : utiliser ton classifieur pneumonie
+    """
+    score = torch.norm(embeddings, dim=-1).mean().item()
+    return score

utils/preprocessing.py

@@ -0,0 +1,7 @@
+import torchaudio
+
+def load_audio(file_path, target_sr=16000):
+    waveform, sr = torchaudio.load(file_path)
+    if sr != target_sr:
+        waveform = torchaudio.transforms.Resample(orig_freq=sr, new_freq=target_sr)(waveform)
+    return waveform, target_sr