怎么做手机网站,服务器网站绑定域名,抖音小程序定制开发,wordpress标记已读文章目录 FashionMNIST数据集需求库导入、数据迭代器生成设备选择样例图片展示日志写入评估—计数器模型构建训练函数整体代码训练过程日志 FashionMNIST数据集
FashionMNIST#xff08;时尚 MNIST#xff09;是一个用于图像分类的数据集#xff0c;旨在替代传统的手写数字… 文章目录 FashionMNIST数据集需求库导入、数据迭代器生成设备选择样例图片展示日志写入评估—计数器模型构建训练函数整体代码训练过程日志 FashionMNIST数据集
FashionMNIST时尚 MNIST是一个用于图像分类的数据集旨在替代传统的手写数字MNIST数据集。它由 Zalando Research 创建适用于深度学习和计算机视觉的实验。 FashionMNIST 包含 10 个类别分别对应不同的时尚物品。这些类别包括 T恤/上衣、裤子、套头衫、裙子、外套、凉鞋、衬衫、运动鞋、包和踝靴。每个类别有 6,000 张训练图像和 1,000 张测试图像总计 70,000 张图像。每张图像的尺寸为 28x28 像素与MNIST数据集相同。数据集中的每个图像都是灰度图像像素值在0到255之间。
需求库导入、数据迭代器生成
import os
import random
import numpy as np
import datetime
import torch
import torch.nn as nn
from torch.utils.data import DataLoaderimport torchvision
from torchvision import transformsimport argparse
from tqdm import tqdmimport matplotlib.pyplot as plt
from torch.utils.tensorboard import SummaryWriterdef _load_data():download the data, and generate the dataloadertrans transforms.Compose([transforms.ToTensor()])train_dataset torchvision.datasets.FashionMNIST(root./data/, trainTrue, downloadTrue, transformtrans)test_dataset torchvision.datasets.FashionMNIST(root./data/, trainFalse, downloadTrue, transformtrans)# print(len(train_dataset), len(test_dataset))train_loader DataLoader(train_dataset, shuffleTrue, batch_sizeargs.batch_size, num_workersargs.num_works)test_loader DataLoader(test_dataset, shuffleTrue, batch_sizeargs.batch_size, num_workersargs.num_works)return (train_loader, test_loader)设备选择
def _device():device torch.device(cuda if torch.cuda.is_available() else cpu)return device样例图片展示
display data examples
def _image_label(labels):text_labels [t-shirt, trouser, pullover, dress, coat,sandal, shirt, sneaker, bag, ankle boot]return [text_labels[int(i)] for i in labels]def _show_images(imgs, rows, columns, titlesNone, scale1.5):figsize (rows * scale, columns * 1.5)fig, axes plt.subplots(rows, columns, figsizefigsize)axes axes.flatten()for i, (img, ax) in enumerate(zip(imgs, axes)):ax.imshow(img)ax.axes.get_xaxis().set_visible(False)ax.axes.get_yaxis().set_visible(False)if titles:ax.set_title(titles[i])plt.show()return axesdef _show_examples():train_loader, test_loader _load_data()for images, labels in train_loader:images images.squeeze(1)_show_images(images, 3, 3, _image_label(labels))break日志写入
class _logger():def __init__(self, log_dir, log_historyTrue):if log_history:log_dir os.path.join(log_dir, datetime.datetime.now().strftime(%Y_%m_%d__%H_%M_%S))self.summary SummaryWriter(log_dir)def scalar_summary(self, tag, value, step):self.summary.add_scalars(tag, value, step)def images_summary(self, tag, image_tensor, step):self.summary.add_images(tag, image_tensor, step)def figure_summary(self, tag, figure, step):self.summary.add_figure(tag, figure, step)def graph_summary(self, model):self.summary.add_graph(model)def close(self):self.summary.close()评估—计数器
class AverageMeter():def __init__(self):self.reset()def reset(self):self.val 0self.avg 0self.sum 0self.count 0def update(self, val, n1):self.val valself.sum val * nself.count nself.avg self.sum / self.count模型构建
class Conv3x3(nn.Module):def __init__(self, in_channels, out_channels, down_sampleFalse):super(Conv3x3, self).__init__()self.conv nn.Sequential(nn.Conv2d(in_channels, out_channels, 3, 1, 1),nn.BatchNorm2d(out_channels),nn.ReLU(inplaceTrue),nn.Conv2d(out_channels, out_channels, 3, 1, 1),nn.BatchNorm2d(out_channels),nn.ReLU(inplaceTrue))if down_sample:self.conv[3] nn.Conv2d(out_channels, out_channels, 2, 2, 0)def forward(self, x):return self.conv(x)class SimpleNet(nn.Module):def __init__(self, in_channels, out_channels):super(SimpleNet, self).__init__()self.conv1 Conv3x3(in_channels, 32)self.conv2 Conv3x3(32, 64, down_sampleTrue)self.conv3 Conv3x3(64, 128)self.conv4 Conv3x3(128, 256, down_sampleTrue)self.fc nn.Linear(256*7*7, out_channels)def forward(self, x):x self.conv1(x)x self.conv2(x)x self.conv3(x)x self.conv4(x)x torch.flatten(x, 1)out self.fc(x)return out训练函数
def train(model, train_loader, test_loader, criterion, optimizor, epochs, device, writer, save_weightFalse):train_loss AverageMeter()test_loss AverageMeter()train_precision AverageMeter()test_precision AverageMeter()time_tick datetime.datetime.now().strftime(%Y_%m_%d__%H_%M_%S)for epoch in range(epochs):print(\nEpoch: [%d | %d] LR: %f % (epoch 1, args.epochs, args.lr))model.train()for input, label in tqdm(train_loader):input, label input.to(device), label.to(device)output model(input)# backwardloss criterion(output, label)optimizor.zero_grad()loss.backward()optimizor.step()# loggerpredict torch.argmax(output, dim1)train_pre sum(predict label) / len(label)train_loss.update(loss.item(), input.size(0))train_precision.update(train_pre.item(), input.size(0))model.eval()with torch.no_grad():for X, y in tqdm(test_loader):X, y X.to(device), y.to(device)y_hat model(X)loss_te criterion(y_hat, y)predict_ torch.argmax(y_hat, dim1)test_pre sum(predict_ y) / len(y)test_loss.update(loss_te.item(), X.size(0))test_precision.update(test_pre.item(), X.size(0))if save_weight:best_dice args.best_diceweight_dir os.path.join(args.weight_dir, args.model, time_tick)os.makedirs(weight_dir, exist_okTrue)monitor_dice test_precision.avgif monitor_dice best_dice:best_dice max(monitor_dice, best_dice)name os.path.join(weight_dir, args.model _ str(epoch) \_test_loss- str(round(test_loss.avg, 4)) \_test_dice- str(round(best_dice, 4)) .pt)torch.save(model.state_dict(), name)print(train ---Loss: {loss:.4f} | Dice: {dice:.4f}.format(losstrain_loss.avg, dicetrain_precision.avg))print(test ---Loss: {loss:.4f} | Dice: {dice:.4f}.format(losstest_loss.avg, dicetest_precision.avg))# summarywriter.scalar_summary(Loss/loss, {train: train_loss.avg, test: test_loss.avg}, epoch)writer.scalar_summary(Loss/precision, {train: train_precision.avg, test: test_precision.avg}, epoch)writer.close()整体代码
import os
import random
import numpy as np
import datetime
import torch
import torch.nn as nn
from torch.utils.data import DataLoaderimport torchvision
from torchvision import transformsimport argparse
from tqdm import tqdmimport matplotlib.pyplot as plt
from torch.utils.tensorboard import SummaryWriterReproduction experiment
def setup_seed(seed):random.seed(seed)np.random.seed(seed)torch.manual_seed(seed)torch.cuda.manual_seed(seed)torch.cuda.manual_seed_all(seed)# torch.backends.cudnn.benchmark False# torch.backends.cudnn.enabled False# torch.backends.cudnn.deterministic Truedata related
def _base_options():parser argparse.ArgumentParser(descriptionTrain setting for FashionMNIST)# about datasetparser.add_argument(--batch_size, default8, typeint, helpthe batch size of dataset)parser.add_argument(--num_works, default4, typeint, helpthe num_works used)# trainparser.add_argument(--epochs, default100, typeint, helptrain iterations)parser.add_argument(--lr, default0.001, typefloat, helplearning rate)parser.add_argument(--model, defaultSimpleNet, choices[SimpleNet], helpthe model choosed)# log dirparser.add_argument(--log_dir, default./logger/, helpthe path of log file)#parser.add_argument(--best_dice, default-100, typeint, helpfor save weight)parser.add_argument(--weight_dir, default./weight/, helpthe dir for save weight)args parser.parse_args()return argsdef _load_data():download the data, and generate the dataloadertrans transforms.Compose([transforms.ToTensor()])train_dataset torchvision.datasets.FashionMNIST(root./data/, trainTrue, downloadTrue, transformtrans)test_dataset torchvision.datasets.FashionMNIST(root./data/, trainFalse, downloadTrue, transformtrans)# print(len(train_dataset), len(test_dataset))train_loader DataLoader(train_dataset, shuffleTrue, batch_sizeargs.batch_size, num_workersargs.num_works)test_loader DataLoader(test_dataset, shuffleTrue, batch_sizeargs.batch_size, num_workersargs.num_works)return (train_loader, test_loader)def _device():device torch.device(cuda if torch.cuda.is_available() else cpu)return devicedisplay data examples
def _image_label(labels):text_labels [t-shirt, trouser, pullover, dress, coat,sandal, shirt, sneaker, bag, ankle boot]return [text_labels[int(i)] for i in labels]def _show_images(imgs, rows, columns, titlesNone, scale1.5):figsize (rows * scale, columns * 1.5)fig, axes plt.subplots(rows, columns, figsizefigsize)axes axes.flatten()for i, (img, ax) in enumerate(zip(imgs, axes)):ax.imshow(img)ax.axes.get_xaxis().set_visible(False)ax.axes.get_yaxis().set_visible(False)if titles:ax.set_title(titles[i])plt.show()return axesdef _show_examples():train_loader, test_loader _load_data()for images, labels in train_loader:images images.squeeze(1)_show_images(images, 3, 3, _image_label(labels))breaklog
class _logger():def __init__(self, log_dir, log_historyTrue):if log_history:log_dir os.path.join(log_dir, datetime.datetime.now().strftime(%Y_%m_%d__%H_%M_%S))self.summary SummaryWriter(log_dir)def scalar_summary(self, tag, value, step):self.summary.add_scalars(tag, value, step)def images_summary(self, tag, image_tensor, step):self.summary.add_images(tag, image_tensor, step)def figure_summary(self, tag, figure, step):self.summary.add_figure(tag, figure, step)def graph_summary(self, model):self.summary.add_graph(model)def close(self):self.summary.close()evaluate the result
class AverageMeter():def __init__(self):self.reset()def reset(self):self.val 0self.avg 0self.sum 0self.count 0def update(self, val, n1):self.val valself.sum val * nself.count nself.avg self.sum / self.countdefine the Net
class Conv3x3(nn.Module):def __init__(self, in_channels, out_channels, down_sampleFalse):super(Conv3x3, self).__init__()self.conv nn.Sequential(nn.Conv2d(in_channels, out_channels, 3, 1, 1),nn.BatchNorm2d(out_channels),nn.ReLU(inplaceTrue),nn.Conv2d(out_channels, out_channels, 3, 1, 1),nn.BatchNorm2d(out_channels),nn.ReLU(inplaceTrue))if down_sample:self.conv[3] nn.Conv2d(out_channels, out_channels, 2, 2, 0)def forward(self, x):return self.conv(x)class SimpleNet(nn.Module):def __init__(self, in_channels, out_channels):super(SimpleNet, self).__init__()self.conv1 Conv3x3(in_channels, 32)self.conv2 Conv3x3(32, 64, down_sampleTrue)self.conv3 Conv3x3(64, 128)self.conv4 Conv3x3(128, 256, down_sampleTrue)self.fc nn.Linear(256*7*7, out_channels)def forward(self, x):x self.conv1(x)x self.conv2(x)x self.conv3(x)x self.conv4(x)x torch.flatten(x, 1)out self.fc(x)return outprogress of train/test
def train(model, train_loader, test_loader, criterion, optimizor, epochs, device, writer, save_weightFalse):train_loss AverageMeter()test_loss AverageMeter()train_precision AverageMeter()test_precision AverageMeter()time_tick datetime.datetime.now().strftime(%Y_%m_%d__%H_%M_%S)for epoch in range(epochs):print(\nEpoch: [%d | %d] LR: %f % (epoch 1, args.epochs, args.lr))model.train()for input, label in tqdm(train_loader):input, label input.to(device), label.to(device)output model(input)# backwardloss criterion(output, label)optimizor.zero_grad()loss.backward()optimizor.step()# loggerpredict torch.argmax(output, dim1)train_pre sum(predict label) / len(label)train_loss.update(loss.item(), input.size(0))train_precision.update(train_pre.item(), input.size(0))model.eval()with torch.no_grad():for X, y in tqdm(test_loader):X, y X.to(device), y.to(device)y_hat model(X)loss_te criterion(y_hat, y)predict_ torch.argmax(y_hat, dim1)test_pre sum(predict_ y) / len(y)test_loss.update(loss_te.item(), X.size(0))test_precision.update(test_pre.item(), X.size(0))if save_weight:best_dice args.best_diceweight_dir os.path.join(args.weight_dir, args.model, time_tick)os.makedirs(weight_dir, exist_okTrue)monitor_dice test_precision.avgif monitor_dice best_dice:best_dice max(monitor_dice, best_dice)name os.path.join(weight_dir, args.model _ str(epoch) \_test_loss- str(round(test_loss.avg, 4)) \_test_dice- str(round(best_dice, 4)) .pt)torch.save(model.state_dict(), name)print(train ---Loss: {loss:.4f} | Dice: {dice:.4f}.format(losstrain_loss.avg, dicetrain_precision.avg))print(test ---Loss: {loss:.4f} | Dice: {dice:.4f}.format(losstest_loss.avg, dicetest_precision.avg))# summarywriter.scalar_summary(Loss/loss, {train: train_loss.avg, test: test_loss.avg}, epoch)writer.scalar_summary(Loss/precision, {train: train_precision.avg, test: test_precision.avg}, epoch)writer.close()if __name__ __main__:# configargs _base_options()device _device()# datatrain_loader, test_loader _load_data()# loggerwriter _logger(log_diros.path.join(args.log_dir, args.model))# modelmodel SimpleNet(in_channels1, out_channels10).to(device)optimizor torch.optim.Adam(model.parameters(), lrargs.lr)criterion nn.CrossEntropyLoss()train(model, train_loader, test_loader, criterion, optimizor, args.epochs, device, writer, save_weightTrue) args _base_options()_show_examples() # ——— 样例图片显示训练过程 日志
