ZeroGPU to CPU

app.py

@@ -1,5 +1,5 @@
 import gradio as gr
-import spaces
+#import spaces ## For ZeroGPU
 import torch
 import torchaudio
 from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2ForSequenceClassification
@@ -9,29 +9,27 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
 model_name = "Hatman/audio-emotion-detection"
 feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(model_name)
 model = Wav2Vec2ForSequenceClassification.from_pretrained(model_name)
-print(device)
-
 
 def preprocess_audio(audio):
     waveform, sampling_rate = torchaudio.load(audio)
     resampled_waveform = torchaudio.transforms.Resample(orig_freq=sampling_rate, new_freq=16000)(waveform)
     return {'speech': resampled_waveform.numpy().flatten(), 'sampling_rate': 16000}
 
-@spaces.GPU
+#@spaces.GPU ## For ZeroGPU
 def inference(audio):
     example = preprocess_audio(audio)
     inputs = feature_extractor(example['speech'], sampling_rate=16000, return_tensors="pt", padding=True)
-    inputs = inputs # Move inputs to GPU
+    inputs = {k: v.to('cpu') for k, v in inputs.items()} # Not necessary on ZeroGPU
     with torch.no_grad():
         logits = model(**inputs).logits
     predicted_ids = torch.argmax(logits, dim=-1)
-    return model.config.id2label[predicted_ids.item()], logits, predicted_ids   # Move tensors back to CPU for further processing
+    return model.config.id2label[predicted_ids.item()], logits, predicted_ids   
     
-@spaces.GPU
+#@spaces.GPU ## For ZeroGPU
 def inference_label(audio):
     example = preprocess_audio(audio)
     inputs = feature_extractor(example['speech'], sampling_rate=16000, return_tensors="pt", padding=True)
-    inputs = inputs  # Move inputs to GPU
+    inputs = {k: v.to('cpu') for k, v in inputs.items()} # Not necessary on ZeroGPU
     with torch.no_grad():
         logits = model(**inputs).logits
     predicted_ids = torch.argmax(logits, dim=-1)