Update tasks/audio.py

tasks/audio.py

@@ -39,7 +39,7 @@ async def evaluate_audio(request: AudioEvaluationRequest):
     logging.info("Données chargées")
     
     test_dataset = dataset["test"]
-    
+    del dataset
     # Start tracking emissions
     tracker.start()
     tracker.start_task("inference")
@@ -51,7 +51,9 @@ async def evaluate_audio(request: AudioEvaluationRequest):
         audio_arrays = [x["array"] for x in examples["audio"]]
         return feature_extractor(audio_arrays, sampling_rate=feature_extractor.sampling_rate, padding="longest", max_length=16000, truncation=True, return_tensors="pt")
     
-    encoded_data_test = test_dataset.map(preprocess_function, remove_columns="audio", batched=True, keep_in_memory=False)
+    encoded_data_test = test_dataset.map(preprocess_function, remove_columns="audio", batched=True, streaming=True, keep_in_memory=False)
+    del feature_extractor
+    del audio_arrays
     # Pipeline de classification
     classifier = pipeline("audio-classification", model="CindyDelage/Challenge_HuggingFace_DFG_FrugalAI", device=-1)
 
@@ -60,10 +62,13 @@ async def evaluate_audio(request: AudioEvaluationRequest):
 
     for data in encoded_data_test:
         # Récupérer les données audio et le label
-        result = classifier(np.asarray(data["input_values"]))
+        with torch.no_grad():
+            result = classifier(np.asarray(data["input_values"]))
         
         predicted_label = result[0]['label']
         predictions.append(1 if predicted_label == 'environment' else 0)
+        del result
+        del predicted_label
 
         # Nettoyage mémoire après chaque batch
         #del input_values