526互联

神经网络中间层特征图可视化（输入为音频）（二）

发布时间 2023-11-22 16:20:41作者: 孜孜不倦fly

相比（一）个人感觉这种方法更好

import librosa
import numpy as np
import utils
import torch
import matplotlib.pyplot as plt

class Hook:
    def __init__(self):
        self.features = None

    def hook_fn(self, module, input, output):
        self.features = output

# 创建钩子的实例
hook = Hook()

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def extract_mbe(_y, _sr, _nfft, _nb_mel):
    #梅尔频谱
    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       #最后必须是[frames, dimensions]

def preprocess_data(X, seq_len, nb_ch):
    # split into sequences
    X = utils.split_in_seqs(X, seq_len)
    X = utils.split_multi_channels(X, nb_ch)
    # Convert to PyTorch tensors
    X = torch.Tensor(X)
    X = X.permute(0,1,3,2)   #x形状为[709,2,40,256]，【总样本数，通道数，特征维度，像素宽度】
    return X

# 提取梅尔频谱特征
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 为输入模型的样本特征


model = torch.load(f'best_model_2.pth')

# 将钩子注册到需要的层
model.cnn1.register_forward_hook(hook.hook_fn)

# 假设`input_data`是你的输入张量
output = model(value)

# 访问存储的特征
retnet_features = hook.features
#print(retnet_features.shape)
# 可视化特征（假设retnet_features是一个张量）
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)

# 可视化批次中第一个样本的特定通道
plt.imshow(retnet_features.sum(1).detach().cpu().numpy(), cmap='viridis', origin='lower')   #[高，通道， 宽]
# plt.imshow(retnet_features.detach().cpu().numpy(), cmap='viridis', origin='lower')   #[高，宽]
plt.show()