自己实现的resnet18，使用pytorch 在数据集cifar10，miniimagmy-resnet-master.zip

import os import torch import torch.nn.functional as F import torchvision import torchvision.transforms as transforms from torchvision.datasets import CIFAR10 import matplotlib.pyplot as plt import numpy as np from MiniImagenet import MiniImagenet from resnet import Resnet18 from evaluation import eval start_lr = 0.01 train_iterations = 60000 save_per_iter = 1000 min_delta = 0.001 device = "cuda:0" dataset_name = "cifar10" record_name = "cifar_record_1.pth" def save_model(model, filename:str): if not os.path.exists("model_files"): os.makedirs("model_files") torch.save(model.state_dict(), os.path.join("model_files", filename)) print(f"Saved '{filename}'") def update_learing_rate(train_losses, optimizer): min_loss = min(train_losses[-save_per_iter:]) if train_losses[-1] > min_loss - min_delta: origin_lr = next(iter(optimizer.param_groups))['lr'] now_lr = origin_lr / 10.0 for param_group in optimizer.param_groups: param_group['lr'] = now_lr print(f"Change learning rate from {origin_lr} to {now_lr}") if dataset_name == "mini-imagenet": transform = transforms.Compose([transforms.Resize([224,224]), transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5])]) train_data = MiniImagenet('data/mini-imagenet', split="train", transform=transform) val_data = MiniImagenet('data/mini-imagenet', split="val", transform=transform) class_num = train_data.class_num if dataset_name == "cifar10": transform = transforms.Compose([transforms.ToTensor(), transforms.Resize([224,224]), transforms.Normalize([0.5,0.5,0.5], [0.5,0.5,0.5])]) train_data = CIFAR10('data', train=True, transform=transform) val_data = CIFAR10('data', train=False, transform=transform) class_num = len(train_data.classes) print(f"Dataset: {dataset_name}, class number: {class_num}") data_loader = torch.utils.data.DataLoader(train_data, shuffle=True, batch_size=32) device = torch.device(device) model = Resnet18(class_num).to(device) optimizer = torch.optim.SGD(model.parameters(), lr=start_lr, momentum=0.9, weight_decay=0.0001) train_losses = [] val_losses = [] val_accuracy = [] iteration = 0 model.train() while iteration < train_iterations: for X, label in data_loader: iteration += 1 X = X.to(device) label = label.to(device) pred = model(X) loss = F.nll_loss(pred, label) train_losses.append(loss.item()) optimizer.zero_grad() loss.backward() optimizer.step() print(f"Iteration {iteration} loss: {loss.item():.5f}") if iteration % save_per_iter == 0: save_model(model, f"model_{iteration}.pth") print("Evaluating ...") correct, val_loss = eval(model, val_data, device) total = len(val_data) accuracy = correct / total val_accuracy.append(accuracy.item()) print(f"Validation accuracy: {correct}/{total} {accuracy*100.0:.3f}%") print(f"Validation loss: {val_loss:.5f}") val_losses.append(val_loss) model.train() #update_learing_rate(train_losses, optimizer) pass if iteration >= train_iterations: break save_dict = {"start_lr": start_lr, "train_iterations": train_iterations, "save_per_iter": save_per_iter, "train_losses": train_losses, "val_losses": val_losses, "val_accuracy": val_accuracy} torch.save(save_dict, os.path.join("records", record_name))