在上一次分類的基礎上將二分類擴展為九分類。
切割圖片為2000*2000，space=2000，按類別放在patch文件夾中。
要將整個數據集劃分為訓練集和測試集。
下面是兩種方法劃分數據集：
方法1：將源文件夾和目標文件夾相互對應起來，根據圖片在文件夾中的位置索引，來判斷是放在訓練集，還是測試集。

import os import shutilpath="/home/cad429/code/yue/Week2/patch" name_list=os.listdir(path) print(name_list)train_dir="/home/cad429/code/yue/Week2/train" if not os.path.exists(train_dir):os.makedirs(train_dir)test_dir="/home/cad429/code/yue/Week2/test" if not os.path.exists(test_dir):os.makedirs(test_dir)val_dir="/home/cad429/code/yue/Week2/val" if not os.path.exists(val_dir):os.makedirs(val_dir)for i in range(len(name_list)):patch_path=os.path.join(path,name_list[i])print(patch_path)patch_list=os.listdir(patch_path)print(patch_list)patch_name_list = [i.split('.')[0] for i in patch_list]print(patch_name_list)print(len(patch_name_list))for j in range(len(patch_name_list)):original_path = os.path.join(patch_path, patch_name_list[j] + ".png")# print(original_path)des_path = os.path.join(train_dir, name_list[i])des_path2=os.path.join(test_dir,name_list[i])if j%2==0:#print(des_path)shutil.copy(original_path,des_path)else:shutil.copy(original_path,des_path2)

方法2：利用sklearn中帶有的劃分數據集的方法，將每個類別中的圖片的名字按照一定的比例劃分到x_train,x_test.

from sklearn.model_selection import train_test_split import numpy as np import os import shutil split_train="/home/cad429/code/yue/Week2/splittrain" if not os.path.exists(split_train):os.makedirs(split_train)split_test="/home/cad429/code/yue/Week2/splittest" if not os.path.exists(split_test):os.makedirs(split_test)Image_path="/home/cad429/code/yue/Week2/patch" Image_list=os.listdir(Image_path) for i in Image_list:#對應的0，1，2，3，4...path=os.path.join(Image_path,i)imag=os.listdir(path)imag_name=[j.split('.')[0] for j in imag]x_train,x_test=train_test_split(imag_name,test_size=0.3)print("i:",i)# print("x_train list:",x_train)# print("x_test list:",x_test)#將圖片歸到對應的train文件夾for er in x_train:train_image_path=os.path.join(split_train,i,er+".png")train_image_original_path=os.path.join(path,er+".png")#print("destination",train_image_path)#print("original:",train_image_original_path)shutil.copy(train_image_original_path,train_image_path)for es in x_test:test_image_path=os.path.join(split_test,i,es+".png")test_image_original_path=os.path.join(path,es+".png")shutil.copy(test_image_original_path,test_image_path)# print("original:",test_image_original_path)# print("destination:",test_image_path)

圖片大小為2000*2000，batch_size=2,learnning_rate=0.01，epoch=100,損失函數為CrossEntropyLoss，優化 函數為Adam時的訓練結果：

測試的結果：

控制變量法做對比試驗：將圖片切為2000*2000，spacing=1000,learnning_rate=0.01,損失函數為CrossEntropyLoss，優化函數為Adam的
訓練 結果：

測試結果：

測試結果顯示，過擬合，要增加一個驗證集，監控模型 是否發生過擬合，來決定是否停止訓練。

將train的訓練集的圖片劃分一部分到驗證集。

import os import shutil import random from sklearn.model_selection import train_test_split def move_train_to_val(train_path,val_path,train2_path):class_list=os.listdir(train_path)print(class_list)for i in class_list:print(i)class_path=os.path.join(train_path,i)Image_list=os.listdir(class_path)Image_len=len(Image_list)Image_name=[j.split('.')[0] for j in Image_list]train_name,val_name=train_test_split(Image_name)for valname in val_name:original=os.path.join(class_path,valname+".png")val_image_path=os.path.join(val_path,i,valname+".png")shutil.move(original,val_image_path)# print("original:", original)# print("vallllll:", val_image_path)for trainname in train_name:trainoriginal=os.path.join(class_path,trainname+".png")#print("trainoriginal:",trainoriginal)train2_image_path=os.path.join(train2_path,i,trainname+".png")shutil.move(trainoriginal,train2_image_path)# print("original",trainoriginal)# print("train2",train2_image_path)if __name__=="__main__":train_path="/home/cad429/code/yue/Week2/train"val_path="/home/cad429/code/yue/Week2/val"train2_path="/home/cad429/code/yue/Week2/train2"move_train_to_val(train_path,val_path,train2_path)

將添加的新的驗證集用來判斷是否結束訓練，然后用測試集來測試訓練的模型。
結果：

總結：對圖片預處理的部分較少，test準確率較低，評判標準只有acc和loss，評判標準較少。有些代碼較冗余，重復功能的代碼需要反復寫。
下一次實驗，減少冗余代碼，評判標準多一點。

總結

以上是生活随笔為你收集整理的医学图像九分类的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。