import gradio as gr
import spaces
import torch
import librosa
from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2ForSequenceClassification

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

model_name = "Hemg/human-emotion-detection"
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(model_name).to(device)
model = Wav2Vec2ForSequenceClassification.from_pretrained(model_name).to(device) 

def preprocess_audio(audio):
    audio_array, sampling_rate = librosa.load(audio, sr=16000)  # Load and resample to 16kHz
    return {'speech': audio_array, 'sampling_rate': sampling_rate}

@spaces.GPU
def inference(audio):
    example = preprocess_audio(audio)
    inputs = feature_extractor(example['speech'], sampling_rate=16000, return_tensors="pt", padding=True)
    inputs = inputs.to(device)  # Move inputs to GPU
    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

iface = gr.Interface(fn=predict_sentiment,
                     inputs=gr.inputs.Audio(source="microphone", type="filepath"),
                     outputs="text",
                     title="Audio Sentiment Analysis",
                     description="Upload an audio file or record one to analyze sentiment.")


iface.launch()