禹州做网站的,网站后台cms,怎么用自己的主机做网站服务器吗,专业制作广告本文主要针对交通标识图片进行分类#xff0c;包含62类#xff0c;这个就是当前科大讯飞比赛#xff0c;目前准确率在0.94左右#xff0c;难点如下#xff1a;
1 类别不均衡#xff0c;有得种类图片2百多#xff0c;有个只有10个不到#xff1b;
2 像素大小不同…
本文主要针对交通标识图片进行分类包含62类这个就是当前科大讯飞比赛目前准确率在0.94左右难点如下
1 类别不均衡有得种类图片2百多有个只有10个不到
2 像素大小不同导致有的图片很清晰有的很模糊
直接上代码 import os
import torch
import torchvision
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import random_splitfrom torchvision import models, datasets, transforms
import torch.utils.data as tud
import numpy as np
from torch.utils.data import Dataset, DataLoader, SubsetRandomSampler
from PIL import Image
import matplotlib.pyplot as plt
import warnings
import pandas as pd
from torch.utils.data import random_splitwarnings.filterwarnings(ignore)# 检测能否使用GPU
print(#labels
torch.device(cuda:0 if torch.cuda.is_available() else cpu)
)device torch.device(cuda:0 if torch.cuda.is_available() else cpu)
n_classes 62 # 几种分类的
preteain False # 是否下载使用训练参数 有网true 没网false
epoches 10 # 训练的轮次
traindataset datasets.ImageFolder(root../all/data/train_set/, transformtransforms.Compose([transforms.Resize((224,224)),#transforms.RandomHorizontalFlip(),transforms.ToTensor(),transforms.Normalize(mean[0.485, 0.456, 0.406], std[0.229, 0.224, 0.225])]))# 分割比例比如80%的数据用于训练20%用于验证
train_val_ratio 0.8
train_size int(len(traindataset) * train_val_ratio)
val_size len(traindataset) - train_size
train_dataset, val_dataset random_split(traindataset, [train_size, val_size])classes traindataset.classes
print(classes)model models.resnext50_32x4d(pretrainedpreteain)
#model models.resnet34(pretrainedpreteain)if preteain True:for param in model.parameters():param.requires_grad Falsemodel.fc nn.Linear(in_features2048, out_featuresn_classes, biasTrue)
model model.to(device)def train_model(model, train_loader, loss_fn, optimizer, epoch):model.train()total_loss 0.total_corrects 0.total 0.for idx, (inputs, labels) in enumerate(train_loader):inputs inputs.to(device)labels labels.to(device)outputs model(inputs)loss loss_fn(outputs, labels)optimizer.zero_grad()loss.backward()optimizer.step()preds outputs.argmax(dim1)total_corrects torch.sum(preds.eq(labels))total_loss loss.item() * inputs.size(0)total labels.size(0)total_loss total_loss / totalacc 100 * total_corrects / totalprint(轮次:%4d|训练集损失:%.5f|训练集准确率:%6.2f%% % (epoch 1, total_loss, acc))return total_loss, accdef test_model(model, test_loader, loss_fn, optimizer, epoch):model.train()total_loss 0.total_corrects 0.total 0.with torch.no_grad():for idx, (inputs, labels) in enumerate(test_loader):inputs inputs.to(device)labels labels.to(device)outputs model(inputs)loss loss_fn(outputs, labels)preds outputs.argmax(dim1)total labels.size(0)total_loss loss.item() * inputs.size(0)total_corrects torch.sum(preds.eq(labels))loss total_loss / totalaccuracy 100 * total_corrects / totalprint(轮次:%4d|测试集损失:%.5f|测试集准确率:%6.2f%% % (epoch 1, loss, accuracy))return loss, accuracyloss_fn nn.CrossEntropyLoss().to(device)optimizer optim.Adam(model.parameters(), lr0.0001)
train_loader DataLoader(train_dataset, batch_size32, shuffleTrue)
test_loader DataLoader(val_dataset, batch_size32, shuffleTrue)
for epoch in range(0, epoches):loss1, acc1 train_model(model, train_loader, loss_fn, optimizer, epoch)loss2, acc2 test_model(model, test_loader, loss_fn, optimizer, epoch)模型预测 sub pd.read_csv(../all/data/example.csv) transform transforms.Compose([ transforms.Resize((224, 224)), transforms.ToTensor(), transforms.Normalize(mean[0.485, 0.456, 0.406], std[0.229, 0.224, 0.225]) ]) model.eval() for path in os.listdir(../all/data/test_set/): try: img Image.open(../all/data/test_set/path) img_p transform(img).unsqueeze(0).to(device) output model(img_p) pred output.argmax(dim1).item() if img.size[0] * img.size[1]2000: plt.imshow(img) plt.show() p 100 * nn.Softmax(dim1)(output).detach().cpu().numpy()[0] sub.loc[sub[ImageID] path,label] classes[pred] print(f{path} size {img.size}, 该图像预测类别为:, classes[pred]) except: print(ferror {path}) sub.loc[sub[ImageID]e57471de-6527-4b9b-90a8-4f1d93909216.png,label] Under Construction sub.loc[sub[ImageID]ff38d59e-9a11-41e4-901b-67097bb0e960.png,label] Keep Left sub.columns [ImageID,Sign Name] label_map pd.read_excel(../all/data/label_map.xlsx) sub_all pd.merge(leftsub,rightlabel_map,onSign Name,howleft) #sub_all[[ImageID,label]].to_csv(./sub_resnet34_add_img_ratio_drop_dire.csv,indexFalse) 个人的心得
1 如何进行图片增强图片增强应该注意什么方向问题
2 模型大小如何进行选择
