import os
import time
import requests
import re
def imgdata_set(save_path,word,epoch):
    q=0     #停止爬取圖片條件
    a=0     #圖片名稱
    while(True):
        time.sleep(1)
        url="https://image.baidu.com/search/flip?tn=baiduimage&ie=utf-8&word={}&pn={}&ct=&ic=0&lm=-1&width=0&height=0".format(word,q)
        #word=需要搜索的名字
        headers={
            'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/88.0.4324.96 Safari/537.36 Edg/88.0.705.56'
        }
        response=requests.get(url,headers=headers)
        # print(response.request.headers)
        html=response.text
        # print(html)
        urls=re.findall('"objURL":"(.*?)"',html)
        # print(urls)
        for url in urls:
            print(a)    #圖片的名字
            response = requests.get(url, headers=headers)
            image=response.content
            with open(os.path.join(save_path,"{}.jpg".format(a)),'wb') as f:
                f.write(image)
            a=a+1
        q=q+20
        if (q/20)>=int(epoch):
            break
if __name__=="__main__":
    save_path = input('你想保存的路徑：')
    word = input('你想要下載什么圖片？請(qǐng)輸入:')
    epoch = input('你想要下載幾輪圖片？請(qǐng)輸入(一輪為60張左右圖片):')  # 需要迭代幾次圖片
    imgdata_set(save_path, word, epoch)

import os
from PIL import Image
root_path=r"C:\Users\Lenovo\Desktop\data"   #待處理文件夾絕對(duì)路徑（可按‘Ctrl+Shift+c'復(fù)制）
root_names=os.listdir(root_path)

for root_name in root_names:
    path=os.path.join(root_path,root_name)
    print("正在刪除文件夾：",path)
    names=os.listdir(path)
    names_path=[]
    for name in names:
        # print(name)
        img=Image.open(os.path.join(path,name))
        name_path=os.path.join(path,name)
        if img==None:           #篩選無法打開的圖片
            names_path.append(name_path)
            print('成功保存錯(cuò)誤圖片路徑：{}'.format(name))
        else:
            w,h=img.size
            if w<50 or h<50:    #篩選錯(cuò)誤圖片
                names_path.append(name_path)
                print('成功保存特小圖片路徑：{}'.format(name))
    print("開始刪除需刪除的圖片")
    for r in names_path:
        os.remove(r)
        print("已刪除：",r)

import os
root_dir=r"C:\Users\Lenovo\Desktop\pzlh"    #原文件夾路徑
save_path=r"C:\Users\Lenovo\Desktop\pzlh2"  #新建文件夾路徑
img_path=os.listdir(root_dir)
a=0
for i in img_path:
    a+=1
    i= os.path.join(os.path.abspath(root_dir), i)
    new_name=os.path.join(os.path.abspath(save_path), str(a) + '_pzlh.jpg')    #此處可以修改圖片名稱
    os.rename(i,new_name)       #特別注意：rename會(huì)刪除原圖

import os
root_path=r"C:\Users\Lenovo\Desktop\data"
save_path=r"C:\Users\Lenovo\Desktop\data_label" #對(duì)應(yīng)的label文件夾下也要建好相應(yīng)的空子文件夾
names=os.listdir(root_path) #得到images文件夾下的子文件夾的名稱
for name in names:
    path=os.path.join(root_path,name)
    img_names=os.listdir(path)  #得到子文件夾下的圖片的名稱
    for img_name in img_names:
        save_name = img_name.split(".jpg")[0]+'.txt'    #得到相應(yīng)的lable名稱
        txt_path=os.path.join(save_path,name)           #得到label的子文件夾的路徑
        with open(os.path.join(txt_path,save_name), "w") as f:  #結(jié)合子文件夾路徑和相應(yīng)子文件夾下圖片的名稱生成相應(yīng)的子文件夾txt文件
            f.write(name)       #將label寫入對(duì)應(yīng)txt文件夾
            print(f.name)

import time
from torch.utils.tensorboard import SummaryWriter
from torchvision.datasets import ImageFolder
from torchvision import transforms
from torch.utils.data import DataLoader
import torchvision.models as models
import torch.nn as nn
import torch

print("是否使用GPU訓(xùn)練：{}".format(torch.cuda.is_available()))    #打印是否采用gpu訓(xùn)練
if torch.cuda.is_available:
    print("GPU名稱為：{}".format(torch.cuda.get_device_name()))  #打印相應(yīng)的gpu信息
#數(shù)據(jù)增強(qiáng)太多也可能造成訓(xùn)練出不好的結(jié)果，而且耗時(shí)長，宜增強(qiáng)兩三倍即可。
normalize=transforms.Normalize(mean=[.5,.5,.5],std=[.5,.5,.5])  #規(guī)范化
transform=transforms.Compose([                                  #數(shù)據(jù)處理
    transforms.Resize((64,64)),
    transforms.ToTensor(),
    normalize
])
dataset_train=ImageFolder('data/train',transform=transform)     #訓(xùn)練數(shù)據(jù)集
# print(dataset_tran[0])
dataset_valid=ImageFolder('data/valid',transform=transform)     #驗(yàn)證或測(cè)試數(shù)據(jù)集
# print(dataset_train.classer)#返回類別
print(dataset_train.class_to_idx)                               #返回類別及其索引
# print(dataset_train.imgs)#返回圖片路徑
print(dataset_valid.class_to_idx)
train_data_size=len(dataset_train)                              #放回?cái)?shù)據(jù)集長度
test_data_size=len(dataset_valid)
print("訓(xùn)練數(shù)據(jù)集的長度為：{}".format(train_data_size))
print("測(cè)試數(shù)據(jù)集的長度為：{}".format(test_data_size))
#torch自帶的標(biāo)準(zhǔn)數(shù)據(jù)集加載函數(shù)
dataloader_train=DataLoader(dataset_train,batch_size=4,shuffle=True,num_workers=0,drop_last=True)
dataloader_test=DataLoader(dataset_valid,batch_size=4,shuffle=True,num_workers=0,drop_last=True)

#2.模型加載
model_ft=models.resnet18(pretrained=True)#使用遷移學(xué)習(xí)，加載預(yù)訓(xùn)練權(quán)重
# print(model_ft)

in_features=model_ft.fc.in_features
model_ft.fc=nn.Sequential(nn.Linear(in_features,36),
                          nn.Linear(36,6))#將最后的全連接改為（36，6），使輸出為六個(gè)小數(shù)，對(duì)應(yīng)六種植物的置信度
#凍結(jié)卷積層函數(shù)
# for i,para in enumerate(model_ft.parameters()):
#     if i<18:
#         para.requires_grad=False

# print(model_ft)


# model_ft.half()#可改為半精度，加快訓(xùn)練速度，在這里不適用

model_ft=model_ft.cuda()#將模型遷移到gpu
#3.優(yōu)化器
loss_fn=nn.CrossEntropyLoss()

loss_fn=loss_fn.cuda()  #將loss遷移到gpu
learn_rate=0.01         #設(shè)置學(xué)習(xí)率
optimizer=torch.optim.SGD(model_ft.parameters(),lr=learn_rate,momentum=0.01)#可調(diào)超參數(shù)

total_train_step=0
total_test_step=0
epoch=50                #迭代次數(shù)
writer=SummaryWriter("logs_train_yaopian")
best_acc=-1
ss_time=time.time()

for i in range(epoch):
    start_time = time.time()
    print("--------第{}輪訓(xùn)練開始---------".format(i+1))
    model_ft.train()
    for data in dataloader_train:
        imgs,targets=data
        # if torch.cuda.is_available():
        # imgs.float()
        # imgs=imgs.float()#為上述改為半精度操作，在這里不適用
        imgs=imgs.cuda()
        targets=targets.cuda()
        # imgs=imgs.half()
        outputs=model_ft(imgs)
        loss=loss_fn(outputs,targets)

        optimizer.zero_grad()   #梯度歸零
        loss.backward()         #反向傳播計(jì)算梯度
        optimizer.step()        #梯度優(yōu)化

        total_train_step=total_train_step+1
        if total_train_step%100==0:#一輪時(shí)間過長可以考慮加一個(gè)
            end_time=time.time()
            print("使用GPU訓(xùn)練100次的時(shí)間為：{}".format(end_time-start_time))
            print("訓(xùn)練次數(shù)：{},loss:{}".format(total_train_step,loss.item()))
            # writer.add_scalar("valid_loss",loss.item(),total_train_step)
    model_ft.eval()
    total_test_loss=0
    total_accuracy=0
    with torch.no_grad():       #驗(yàn)證數(shù)據(jù)集時(shí)禁止反向傳播優(yōu)化權(quán)重
        for data in dataloader_test:
            imgs,targets=data
            # if torch.cuda.is_available():
            # imgs.float()
            # imgs=imgs.float()
            imgs = imgs.cuda()
            targets = targets.cuda()
            # imgs=imgs.half()
            outputs=model_ft(imgs)
            loss=loss_fn(outputs,targets)
            total_test_loss=total_test_loss+loss.item()
            accuracy=(outputs.argmax(1)==targets).sum()
            total_accuracy=total_accuracy+accuracy
        print("整體測(cè)試集上的loss：{}(越小越好,與上面的loss無關(guān)此為測(cè)試集的總loss)".format(total_test_loss))
        print("整體測(cè)試集上的正確率：{}(越大越好)".format(total_accuracy / len(dataset_valid)))

        writer.add_scalar("valid_loss",(total_accuracy/len(dataset_valid)),(i+1))#選擇性使用哪一個(gè)
        total_test_step = total_test_step + 1
        if total_accuracy > best_acc:   #保存迭代次數(shù)中最好的模型
            print("已修改模型")
            best_acc = total_accuracy
            torch.save(model_ft, "best_model_yaopian.pth")
ee_time=time.time()
zong_time=ee_time-ss_time
print("訓(xùn)練總共用時(shí):{}h:{}m:{}s".format(int(zong_time//3600),int((zong_time%3600)//60),int(zong_time%60))) #打印訓(xùn)練總耗時(shí)
writer.close()

import os
import torch
import torchvision
from PIL import Image
from torch import nn
i=0 #識(shí)別圖片計(jì)數(shù)
root_path="測(cè)試_data"         #待測(cè)試文件夾
names=os.listdir(root_path)
for name in names:
    print(name)
    i=i+1
    data_class=['滴水觀音','發(fā)財(cái)樹','非洲茉莉','君子蘭','盆栽蘆薈','文竹']   #按文件索引順序排列
    image_path=os.path.join(root_path,name)             
    image=Image.open(image_path)
    print(image)
    transforms=torchvision.transforms.Compose([torchvision.transforms.Resize((64,64)),
                                              torchvision.transforms.ToTensor()])
    image=transforms(image)
    print(image.shape)

    model_ft=torchvision.models.resnet18()      #需要使用訓(xùn)練時(shí)的相同模型
    # print(model_ft)
    in_features=model_ft.fc.in_features
    model_ft.fc=nn.Sequential(nn.Linear(in_features,36),
                              nn.Linear(36,6))     #此處也要與訓(xùn)練模型一致

    model=torch.load("best_model_yaopian.pth",map_location=torch.device("cpu")) #選擇訓(xùn)練后得到的模型文件
    # print(model)
    image=torch.reshape(image,(1,3,64,64))      #修改待預(yù)測(cè)圖片尺寸，需要與訓(xùn)練時(shí)一致
    model.eval()
    with torch.no_grad():
        output=model(image)
    print(output)               #輸出預(yù)測(cè)結(jié)果
    # print(int(output.argmax(1)))
    print("第{}張圖片預(yù)測(cè)為：{}".format(i,data_class[int(output.argmax(1))]))   #對(duì)結(jié)果進(jìn)行處理，使直接顯示出預(yù)測(cè)的植物種類