predict.py

import torch
import torch.nn as nn
import librosa
from transformers import WhisperProcessor, WhisperModel

class WhisperClassifier(nn.Module):
    def __init__(self, model_name="openai/whisper-small"):
        super(WhisperClassifier, self).__init__()
        self.whisper = WhisperModel.from_pretrained(model_name)
        self.fc = nn.Linear(self.whisper.config.d_model, 1)
        self.sigmoid = nn.Sigmoid()

    def forward(self, input_features):
        hidden_states = self.whisper.encoder(input_features).last_hidden_state
        pooled_output = hidden_states.mean(dim=1)
        logits = self.fc(pooled_output)
        return self.sigmoid(logits).squeeze(1)

def predict(model, audio_path, processor, device='cpu'):
    audio, sr = librosa.load(audio_path, sr=16000)
    input_features = processor(audio, return_tensors="pt", sampling_rate=16000).input_features
    input_features = input_features.to(device)

    model.eval()
    with torch.no_grad():
        output = model(input_features)
        prediction = 1 if output.item() > 0.5 else 0  # Convert probability to 0 or 1
    return prediction

def load_model(model_path):
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = WhisperClassifier().to(device)
    model.load_state_dict(torch.load(model_path, map_location=device))
    return model, device