简易语音分类模型

m0_57987491

已于 2023-03-25 10:09:41 修改

阅读量126

点赞数 1

分类专栏：鸟声识别文章标签：分类 python 深度学习

于 2023-03-25 10:06:44 首次发布

本文链接：https://blog.youkuaiyun.com/m0_57987491/article/details/129763471

版权

鸟声识别专栏收录该内容

2 篇文章

订阅专栏

该代码定义了一个AudioClassifier类，它包含卷积神经网络结构用于音频分类。在主函数中，模型被训练在名为2s_wav_mini的数据集上，使用Adam优化器和交叉熵损失函数。训练过程包括前向传播、损失计算、反向传播和优化步骤。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

import torch
import torch.nn as nn
import torch.nn.functional as F


class AudioClassifier(nn.Module):
    def __init__(self, num_classes):
        super(AudioClassifier, self).__init__()

        self.features = nn.Sequential(
            nn.Conv2d(2, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2)),
            nn.Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2)),
            nn.Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2))
        )

        self.classifier = nn.Sequential(
            nn.Linear(128 * 25 * 22, 512),
            nn.ReLU(),
            nn.Linear(512, num_classes)
        )

    def forward(self, x):
        x = self.features(x)
        x = x.view(-1, 128 * 25 * 22)
        x = self.classifier(x)
        return x


if __name__ == '__main__':
    net = AudioClassifier(num_classes=5)
    temp = torch.randn(32, 1, 200, 176)
    out = net(temp)
    print(out)
    print('out', out.shape)

主函数：

import torch
from torch import nn, optim
from torch.utils.data import Dataset, DataLoader
from conv_audio import AudioClassifier
from dataset import WavDataset
from tqdm import tqdm, trange


def main():
    batchsz = 32
    num_classes = 5
    train_Data = DataLoader(WavDataset('2s_wav_mini'), batch_size=batchsz, shuffle=True)
    test_Data = DataLoader(WavDataset('2s_wav_mini'), batch_size=batchsz, shuffle=True)

    device = torch.device('cuda')
    model = AudioClassifier(num_classes).to(device)
    criteon = nn.CrossEntropyLoss().to(device)
    optimizer = optim.Adam(model.parameters(), lr=1e-3)
    for epoch in range(100):
        loss = 0
        train_total = 0
        train_correct = 0
        tqdm(total=100)
        for batchidx, (x, label) in enumerate(train_Data):
            # x = [b , 2, 200, 176 ]
            # label = [b]
            x, label = x.to(device), label.to(device)

            logits = model(x)
            # logits = [b,5]
            # label = [b]
            # loss = tensor scalar
            loss = criteon(logits, label)
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            _, predicted = torch.max(logits.data, 1)
            train_total += label.size(0)
            train_correct += (predicted == label).sum().item()
        # print('epoch', epoch, "loss", loss.item())
        print('epoch:', epoch, 'train_correct_rate:', train_correct / train_total, 'train_loss:', loss.item())


if __name__ == '__main__':
    main()