hubert-base-960h-itw-deepfake（昇腾NPU适配版）

# 1. Clone this repository
git clone <repo-url>
cd hubert-npu-adapted

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run inference on NPU
python inference_npu.py --model_path ./hubert-base-960h-itw-deepfake

# 4. Run accuracy verification (CPU vs NPU)
python verify_accuracy.py

import torch
import torch_npu
import numpy as np
from transformers import AutoConfig, Wav2Vec2FeatureExtractor, HubertForSequenceClassification

# Use NPU
device = torch.device("npu" if torch.npu.is_available() else "cpu")

model_path = "./hubert-base-960h-itw-deepfake"
config = AutoConfig.from_pretrained(model_path)
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(model_path)
model = HubertForSequenceClassification.from_pretrained(model_path, config=config).to(device)
model.eval()

# Dummy audio (replace with real waveform)
audio = np.random.randn(16000).astype(np.float32) * 0.1
inputs = feature_extractor(audio, sampling_rate=16000, return_tensors="pt", padding=True)
input_values = inputs.input_values.to(device)

with torch.no_grad():
    logits = model(input_values).logits
    probs = torch.softmax(logits, dim=-1)

print(probs)  # [[bona-fide_prob, spoof_prob]]

audio_batch = [np.random.randn(16000).astype(np.float32) for _ in range(4)]
inputs = feature_extractor(audio_batch, sampling_rate=16000, return_tensors="pt", padding=True)
input_values = inputs.input_values.to(device)
with torch.no_grad():
    logits = model(input_values).logits
    probs = torch.softmax(logits, dim=-1)