相比方法一個人感覺這種方法更好

import librosa
import numpy as np
import utils
import torch
import matplotlib.pyplot as pltclass Hook:def __init__(self):self.features = Nonedef hook_fn(self, module, input, output):self.features = output# 創建鉤子的實例
hook = Hook()device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def extract_mbe(_y, _sr, _nfft, _nb_mel):#梅爾頻譜spec = librosa.core.spectrum._spectrogram(y=_y, n_fft=_nfft, hop_length=_nfft // 2, power=1)[0]mel_basis = librosa.filters.mel(sr=_sr, n_fft=_nfft, n_mels=_nb_mel)mel_spec = np.log(np.dot(mel_basis, spec).T)return mel_spec       #最后必須是[frames, dimensions]def preprocess_data(X, seq_len, nb_ch):# split into sequencesX = utils.split_in_seqs(X, seq_len)X = utils.split_multi_channels(X, nb_ch)# Convert to PyTorch tensorsX = torch.Tensor(X)X = X.permute(0,1,3,2)   #x形狀為[709,2,40,256]，【總樣本數，通道數，特征維度，像素寬度】return X# 提取梅爾頻譜特征
audio_path = "a011.wav"
y, sr = librosa.load(audio_path, sr=44100)
mel = extract_mbe(y, sr, 2048, 64)value = preprocess_data(mel, 256, 1).to(device)     #value 為輸入模型的樣本特征model = torch.load(f'best_model_2.pth')# 將鉤子注冊到需要的層
model.cnn1.register_forward_hook(hook.hook_fn)# 假設`input_data`是你的輸入張量
output = model(value)# 訪問存儲的特征
retnet_features = hook.features
#print(retnet_features.shape)
# 可視化特征（假設retnet_features是一個張量）
retnet_features = retnet_features.permute(0, 2, 1, 3)
#retnet_features = retnet_features.transpose(1, 2)
#print(retnet_features.shape)
retnet_features = torch.cat([retnet_features[i] for i in range(10)], dim=2)
#print(retnet_features.shape)# 可視化批次中第一個樣本的特定通道
plt.imshow(retnet_features.sum(1).detach().cpu().numpy(), cmap='viridis', origin='lower')   #[高，通道， 寬]
# plt.imshow(retnet_features.detach().cpu().numpy(), cmap='viridis', origin='lower')   #[高，寬]
plt.show()