语音分类问题处理，使用keras

import pandas as pd import os os.environ['CUDA_LAUNCH_BLOCKING'] = '1' from tqdm import tqdm dir_path = './data/train' dirs = os.listdir(dir_path) tmp_path, tmp_label = [], [] for path in dirs: label=0 if path.split('.')[0]=='daisy': label = 1 if path.split('.')[0] == 'dandelion': label = 2 if path.split('.')[0] == 'roses': label = 3 if path.split('.')[0] == 'sunflowers': label = 4 if path.split('.')[0] == 'tulips': label = 5 tmp_label.append(label) tmp_path.append(os.path.join(dir_path, path)) csv_data = {'path': tmp_path, 'label':tmp_label} data = pd.DataFrame(csv_data) data.to_csv('./data/data.csv') data = data.sample(frac=1) train_data = data[:int(len(data)*0.85)] val_data = data[int(len(data)*0.85):] print('共划分训练集样本{}个，验证集样本{}个'.format(len(train_data), len(val_data))) import torch from torch import nn from torch.utils.data import Dataset, DataLoader import torch.nn.functional as F from torchvision import models from torchvision import transforms from PIL import Image import cv2 import numpy as np vgg16 = models.vgg16(pretrained=True) vgg16.classifier[-1] = nn.Linear(4096, 6, bias=True) vgg16.classifier for param in vgg16.features.parameters(): param.requires_grad = False class VggDataset(Dataset): def __init__(self, data): super(VggDataset, self).__init__() self.path = data['path'] self.label = data['label'] self.len = len(self.path) self.transform = transforms.Compose([ transforms.Resize((224, 224)), # 重置图像分辨率 transforms.ToTensor(), # 转化为张量并归一化至[0-1] transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) def __getitem__(self, index): img = self.transform(Image.open(self.path.iloc[index])) return img, torch.LongTensor([self.label.iloc[index]]).squeeze() def __len__(self): return self.len BATCH_SIZE = 8 EPOCHS = 10 lr = 0.001 # device = torch.device("cuda" if torch.cuda.is_available() else "cpu") device = torch.device("cuda" ) train_set = VggDataset(train_data) val_set = VggDataset(val_data) train_loader = DataLoader(train_set, batch_size=BATCH_SIZE, shuffle=True) val_loader = DataLoader(val_set, batch_size=BATCH_SIZE, shuffle=False) loss_stack = [] val_acc = [] optimizer = torch.optim.Adam(vgg16.parameters(), lr=lr) loss_fn = torch.nn.CrossEntropyLoss() vgg16 = vgg16.to(device) for epoch in tqdm(range(EPOCHS)): vgg16.train() all_loss = 0 for data, labels in train_loader: data, labels = data.to(device), labels.to(device) preds = vgg16(data) loss = loss_fn(preds, labels) all_loss += loss.detach().cpu().numpy() optimizer.zero_grad() loss.backward() optimizer.step() loss_stack.append(all_loss) vgg16.eval() count = 0 data_num = 0 for data, labels in val_loader: data, labels = data.to(device), labels.to(device) preds = vgg16(data) pred_labels = torch.argmax(preds, dim=1) for i, p in enumerate(pred_labels): if p == labels[i]: count += 1 data_num += data.shape[0] val_acc.append(count / data_num) torch.save(vgg16, "vgg16_model.pth") import matplotlib.pyplot as plt plt.subplot(1, 2, 1) plt.plot(val_acc, c='red') plt.subplot(1, 2, 2) plt.plot(loss_stack, c='red') plt.show() print('val_acc',val_acc[-1]) print('loss_stack',loss_stack[-1])