DeepLearningコード（Pytorch）

はじめに

大学院を卒業したら，AIの研究職に…
いかないので，何年後かに見ても思い出せるようにコードを置いておきます．

Kaggle記事

こちらは全コードを載せています．より実践的なものとしてKaggeleでのタスクをまとめてあります．
RNNはすごくわかりやすいKaggeleのNotebookを見つけたので，それを載せています．（自分でRNN解くのがめんどくさかった…とかじゃないよ…うん…）

• DNN
  • Classification
    • Kaggleに挑戦 Titanic🚢（Notebook🥉）
  • Regression
    • Kaggleに挑戦 Tabular Ordinal Regression with a Wine Quality Dataset🍷
• RNN
  • Time Series Forecast
    • Recurrent Neural N. (RNN) Tutorial for Beginners
  • Classification
    • メディカルAI専門コース　ディープラーニングを使ったモニタリングデータの時系列解析
• CNN
  • Classification
    • Kaggleに挑戦 DigitRecognizer
  • Segmentation
    • 準備中

※ RNNのClassificationはKaggeleで見つけられなかったので，心電図の不整脈種類分類タスクのURLを載せています．しかしChainerで書かれています…（私はChainerアンチ）まぁ参考程度に．実際に自分でタスクするならPytorchで書き直すことをお勧めします😀
※ SegmentationはTutorial的なコンペがなく，すぐまとめられそになかったので現段階(2023/02/11時点)では載せれてません．

DNN(Base）

基本的な部分はDNNで書きます．
CNN・RNN・Segmentationはそれぞれ変化がある部分だけ載せます．

必要なライブラリをimportします．
pipなりcondaなりで環境にインストールしておいてください．
これら2つはどっちかに統一してください．混同すると環境壊れるこもがあるらしいっす
俺はpip派です

import os
import cv2
import random
import numpy as np

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
from torch.utils.data import TensorDataset
from torch.utils.data import Dataset
from torch.utils.data import DataLoader

#import timm

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline

from sklearn.metrics import confusion_matrix
from sklearn.metrics import classification_report
from sklearn.metrics import accuracy_score
import pandas as pd
import seaborn as sn 

# Validation
from torchvision import datasets
from torch.utils.data.dataset import Subset

# n分割交差検証
# テスト・学習をn分割してn回の平均精度を出す
# Validationは学習に分類されたものから何割かを持ってくる
from sklearn.model_selection import KFold
from sklearn.model_selection import StratifiedKFold

# モデルの保存
import pickle

# グラフのスタイルを指定
plt.style.use('seaborn-darkgrid')

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

データセットの用意です．
今回は自前ではなく，IRISデータを使います．
Deepに入れるようにデータをTensorにしときます．

from sklearn.datasets import load_iris
iris = load_iris()

X_data = iris.data
y_data = iris.target
y_data = y_data.reshape(-1, 1)
OH_y_data = OneHotEncoder().fit_transform(y_data).toarray()
print(X_data.shape)
print(OH_y_data.shape)
X_tensor = torch.Tensor(X_data).to(device)
y_tensor = torch.Tensor(OH_y_data).to(device)
dataset = TensorDataset(X_tensor, y_tensor)


batch_size = 4

Train，Validation，Testの処理部分です．

def train_epoch(model, optimizer, criterion, dataloader, device):
    train_loss = 0
    model.train()
    
    for i, (images, labels) in enumerate(dataloader):
        #labels = labels.type(torch.LongTensor) 
        images, labels = images.to(device), labels.to(device)
        optimizer.zero_grad()
        outputs = model(images)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
        train_loss += loss.item()
    train_loss = train_loss / len(dataloader.dataset)
    
    return train_loss

def validation(model, optimizer, criterion, dataloader, device):
    model.eval()
    val_loss=0
    
    with torch.no_grad():
        for i, (images, labels) in enumerate(dataloader):
            #labels = labels.type(torch.LongTensor) 
            images, labels = images.to(device), labels.to(device)
            outputs = model(images)
            loss = criterion(outputs, labels)
            val_loss += loss.item()
        val_loss = val_loss / len(dataloader.dataset)
    return val_loss


def inference(model, optimizer, criterion, dataloader, device):
    model.eval()
    test_loss=0

    with torch.no_grad():
        for i, (images, labels) in enumerate(dataloader):
            #labels = labels.type(torch.LongTensor) 
            images, labels = images.to(device), labels.to(device)
            outputs = model(images)
            loss = criterion(outputs, labels)
            test_loss += loss.item()
        test_loss = test_loss / len(dataloader.dataset)
    return test_loss

EarlyStoppingの実装です．
Patience回連続でValidationLossが更新されなければ，学習を打ち止め．最もValidationLossが下がったEpochのモデルを採用します．

class EarlyStopping:
    def __init__(self, patience=10, verbose=False, path='checkpoint_model.pth'):
        self.patience = patience   
        self.verbose = verbose      
        self.counter = 0            
        self.best_score = None      
        self.early_stop = False     
        self.val_loss_min = np.Inf   
        self.path = path             

    def __call__(self, val_loss, model):
        score = -val_loss

        if self.best_score is None:  
            self.best_score = score   
            self.checkpoint(val_loss, model)  
        elif score < self.best_score:  
            self.counter += 1   
            if self.verbose:  
                print(f'EarlyStopping counter: {self.counter} out of {self.patience}')   
            if self.counter >= self.patience:  
                self.early_stop = True
        else:  #ベストスコアを更新した場合
            self.best_score = score  
            self.checkpoint(val_loss, model)  
            self.counter = 0  

    def checkpoint(self, val_loss, model):
        if self.verbose:  
            print(f'Validation loss decreased ({self.val_loss_min:.6f} --> {val_loss:.6f}).  Saving model ...')
        torch.save(model.state_dict(), self.path)  
        self.val_loss_min = val_loss  

RUN部分です．
Epoch分だけTrain・Validation・Testします．
EarlyStoppingで早く終了する場合もあります．

def run(num_epochs, optimizer, criterion, device, train_loader, val_loader, test_loader,model):
    train_loss_list = []
    test_loss_list = []
    val_loss_list = []
    
    earlystopping = EarlyStopping(verbose=True)

    for epoch in range(num_epochs):
        train_loss = train_epoch(model, optimizer, criterion, train_loader, device)
        val_loss = validation(model, optimizer, criterion, val_loader, device)
        test_loss = inference(model, optimizer, criterion, test_loader, device)

        print(f'Epoch [{epoch+1}], train_Loss : {train_loss:.4f}, test_Loss : {test_loss:.4f}')
        train_loss_list.append(train_loss)
        test_loss_list.append(test_loss)
        val_loss_list.append(val_loss)

        earlystopping(val_loss_list[-1], model)
        if earlystopping.early_stop: 
          print("Early Stopping!")
          break
      
    return train_loss_list, val_loss_list, test_loss_list

結果の混同行列です．

def graph(train_loss_list, val_loss_list, test_loss_list):
    num_epochs=len(train_loss_list)
    fig, ax = plt.subplots(figsize=(4, 3), dpi=100)
    ax.plot(range(num_epochs), train_loss_list, c='b', label='train loss')
    ax.plot(range(num_epochs), val_loss_list, c='r', label='test loss')
    ax.plot(range(num_epochs), test_loss_list, c='m', label='val loss')
    ax.set_xlabel('epoch', fontsize='10')
    ax.set_ylabel('loss', fontsize='10')
    ax.set_title('training and test loss', fontsize='10')
    ax.grid()
    ax.legend(fontsize='10')
    plt.show()


def print_confusion_matrix(test_loader,model,cv_y_true,cv_y_pred):
    
    model.eval()
    y_true,y_pred = [],[]
    
    with torch.no_grad():
        for i, (images, labels) in enumerate(test_loader):
            labels = labels.type(torch.LongTensor) 
            images, labels = images.to(device), labels.to(device)
            outputs = model(images)
            for nval in range(len(labels)):
                print(f"{outputs[nval]}\t{labels[nval]}")
                y_true.append(torch.argmax(labels[nval]))
                y_pred.append(torch.argmax(outputs[nval]))
                
                
    for leny in range(len(y_true)):
        y_true[leny] = y_true[leny].item()
        y_pred[leny] = y_pred[leny].item()
        
    ## CV ALL CONFUSION MATRIX
    cv_y_true.append(y_true)
    cv_y_pred.append(y_pred)
    
    target_names = ['0', '1', '2']
    cmx = confusion_matrix(y_true, y_pred)
    df_cmx = pd.DataFrame(cmx, index=target_names, columns=target_names)
    plt.figure(figsize = (6,3))
    sn.heatmap(df_cmx, annot=True, annot_kws={"size": 18}, fmt="d", cmap='Blues')
    plt.show()   
    
    print(classification_report(y_true, y_pred, target_names=target_names))
    print("accuracy: ", accuracy_score(y_true, y_pred))

モデルです．
今回は全結合だけの非常にシンプルなモデルです．活性化関数にReLUを挟んでいます．

'''モデルの定義'''
class RNN(nn.Module):
    def __init__(self):
        super(RNN, self).__init__()
        self.fc1 = nn.Linear(4, 100)
        self.fc2 = nn.Linear(100, 50)
        self.fc3 = nn.Linear(50, 3)
        self.lsf  = nn.LogSoftmax(dim=1)
        self.sf  = nn.Softmax(dim=1)
        
    def forward(self, x):
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        x = self.lsf(x)      
          
        return x

交差検証のための準備です．

# RUN
cv = StratifiedKFold(n_splits=5, random_state=0, shuffle=True)
cv_y_true,cv_y_pred = [],[]

メイン部分です．
上記で実装した関数を呼び出していきます．

for i,(train_index, test_index) in enumerate(cv.split(X_data,y_data)):    

    model = RNN().to(device)
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(),lr=0.001)
    #optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

    # train/test 分割
    cv_train_dataset = Subset(dataset, train_index)
    cv_test_dataset  = Subset(dataset, test_index)

    # validation作成
    n_samples = len(cv_train_dataset)             # train の大きさ
    train_size = round(n_samples * 0.75)          # train の1/4をバリデーションにする
    subset1_indices = list(range(0,train_size))
    subset2_indices = list(range(train_size,n_samples)) 
    strain_dataset = Subset(cv_train_dataset, subset1_indices)
    val_dataset   = Subset(cv_train_dataset, subset2_indices)

    train_loader = DataLoader(strain_dataset, batch_size=batch_size, shuffle=True)
    val_loader   = DataLoader(val_dataset, batch_size=batch_size, shuffle=True)
    test_loader  = DataLoader(cv_test_dataset, batch_size=batch_size, shuffle=False)


    # run
    print(f"***FOLD {i}")
    train_loss_list, val_loss_list, test_loss_list = run(500, optimizer, criterion, device, train_loader, val_loader, test_loader,model)
    model = pickle.load(open('check_point_model.sav', 'rb'))
    
    # PLOT
    graph(train_loss_list, val_loss_list, test_loss_list)
    print_confusion_matrix(test_loader,model,cv_y_true,cv_y_pred)
    print("-----------------\n")

cv_y_true = list(itertools.chain.from_iterable(cv_y_true))
cv_y_pred = list(itertools.chain.from_iterable(cv_y_pred))

# Deep
target_names = ['0', '1', '2']
cmx = confusion_matrix(cv_y_true, cv_y_pred)
df_cmx = pd.DataFrame(cmx, index=target_names, columns=target_names)
plt.figure(figsize = (6,3))
sn.heatmap(df_cmx, annot=True, annot_kws={"size": 18}, fmt="d", cmap='Blues')
plt.show()   
print(classification_report(cv_y_true, cv_y_pred, target_names=target_names))
print("accuracy: ", accuracy_score(cv_y_true, cv_y_pred))

CNN

フォルダの配置

• data
  • Cat
  • Dog

import timm

DATADIR = "C:/Users/~~/data" # Change Own DataPATH
# 各ラベル
CATEGORIES = ["Cat", "Dog"]

# リサイズ後のサイズ
IMG_SIZE = 300
training_data = []

def create_training_data():
  for class_num, category in enumerate(CATEGORIES):
    path = os.path.join(DATADIR, category)
    for image_name in os.listdir(path):
      try:
        # 画像読み込み
        img_array = cv2.imread(os.path.join(path, image_name),) 
        # 画像のリサイズ
        img_resize_array = cv2.resize(img_array, (IMG_SIZE, IMG_SIZE))
        # 画像データ、ラベル情報を追加
        training_data.append([img_resize_array, class_num])  
      except Exception as e:
          pass

create_training_data()

random.shuffle(training_data)  # データをシャッフル

X_data = []  # 画像データ
y_data = []  # ラベル情報

# データセット作成
for feature, label in training_data:
    X_data.append(feature)
    y_data.append(label)

# numpy配列に変換
X_data = np.array(X_data)
X_data = X_data.transpose(0,3,1,2)
y_data = np.array(y_data)

DataAugmentation用の関数です．

image_transform = transforms.RandomOrder([
    transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),
    transforms.RandomPosterize(bits=4),
])

share_transform = transforms.RandomOrder([
    transforms.RandomErasing(),
    transforms.RandomHorizontalFlip(),
    transforms.RandomResizedCrop(IMG_SIZE, scale=(0.8, 1.2)),
    transforms.RandomAffine(degrees=[-10, 10],translate=(0.2, 0.2)),
])

def fix_seed(seed):
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True

Trainの時のみDataAugmentationを行います．

def train_epoch(model, optimizer, criterion, dataloader, device):
    train_loss = 0
    model.train()
    
    for i, (images, labels) in enumerate(dataloader):
        # DataAugmentation ------#
        ### FIX SEED
        seed = random.randint(0, 2**32)
        ### IMAGE
        images = images.to(torch.uint8)
        fix_seed(seed)
        images = share_transform(images)
        images = image_transform(images)
        images = images.to(torch.float32)
        #------------------------#
        images, labels = images.to(device), labels.to(device)
        optimizer.zero_grad()
        outputs = model(images)
        outputs = torch.sigmoid(outputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
        train_loss += loss.item()
        del images,labels,loss,outputs
    train_loss = train_loss / len(dataloader.dataset)
    
    return train_loss

モデルはTIMMから，ImageNetで事前学習されているモデルをファインチューニングして使います．

for i,(train_index, test_index) in enumerate(cv.split(X_data,y_data)):    

    # モデル指定
    TIMM = timm.create_model('tf_efficientnetv2_s_in21ft1k', pretrained=True, num_classes=2)
    model = TIMM.to(device)
    ###############################################################
    # 重み変更
    weights0 = 50.0
    weights1 = 100.0
    w0 = float(weights0)
    w1 = float(weights1)
    all_zero = [float(0.0) for i in range(2)]
    all_zero[0] = w0
    all_zero[1] = w1
    weights = torch.tensor(all_zero)
    # 同デバイスに載せる
    weights = weights.to(device)
    ###############################################################
    criterion = nn.CrossEntropyLoss(weight=weights)
    optimizer = optim.Adam(model.parameters())

    # train/test 分割
    cv_train_dataset = Subset(dataset, train_index)
    cv_test_dataset  = Subset(dataset, test_index)
    # ↑保存
    """
    listname = 'train_index_'+str(i)+'.txt'
    pickle.dump(train_index, open(listname, 'wb'))
    listname = 'test_index_'+str(i)+'.txt'
    pickle.dump(test_index, open(listname, 'wb'))
    """

    # validation作成
    n_samples = len(cv_train_dataset)             # train の大きさ
    train_size = round(n_samples * 0.75)          # train の1/4をバリデーションにする
    subset1_indices = list(range(0,train_size))
    subset2_indices = list(range(train_size,n_samples)) 
    strain_dataset = Subset(cv_train_dataset, subset1_indices)
    val_dataset   = Subset(cv_train_dataset, subset2_indices)

    train_loader = DataLoader(strain_dataset, batch_size=batch_size, shuffle=True)
    val_loader   = DataLoader(val_dataset, batch_size=batch_size, shuffle=True)
    test_loader  = DataLoader(cv_test_dataset, batch_size=batch_size, shuffle=False)

    # run
    print(f"***FOLD {i}")
    train_loss_list, val_loss_list, test_loss_list = run(300, optimizer, criterion, device, train_loader, val_loader, test_loader)
    
    # PLOT
    graph(train_loss_list, val_loss_list, test_loss_list)
    print_confusion_matrix(test_loader,model)
    
    """
    # モデルを保存
    filename = 'timm_test_model_'+str(i)+'.sav'
    pickle.dump(model, open(filename, 'wb'))
    # train_lossの推移を保存
    listname = 'train_loss_list_'+str(i)+'.txt'
    pickle.dump(train_loss_list, open(listname, 'wb'))
    # val_lossの推移を保存
    listname = 'val_loss_list_'+str(i)+'.txt'
    pickle.dump(val_loss_list, open(listname, 'wb'))
    # test_lossの推移を保存
    listname = 'test_loss_list_'+str(i)+'.txt'
    pickle.dump(test_loss_list, open(listname, 'wb'))
    """
    
    # 各実行の最後のLossを保存
    fold_train_list.append(train_loss_list[-1])
    fold_val_list.append(val_loss_list[-1])
    fold_test_list.append(test_loss_list[-1])
    print("-----------------\n")

交差検証全体の混同行列を出力します．

import itertools

## CV ALL
target_names = ['anomaly', 'nomaly']
cv_y_true = list(itertools.chain.from_iterable(cv_y_true))
cv_y_pred = list(itertools.chain.from_iterable(cv_y_pred))
cmx = confusion_matrix(cv_y_true, cv_y_pred)
df_cmx = pd.DataFrame(cmx, index=target_names, columns=target_names)
plt.figure(figsize = (6,3))
sn.heatmap(df_cmx, annot=True, annot_kws={"size": 18}, fmt="d", cmap='Blues')
plt.show()   


print(classification_report(cv_y_true, cv_y_pred, target_names=target_names))
print("accuracy: ", accuracy_score(cv_y_true, cv_y_pred))

CNN-GradCAM

CNNの注目画素を可視化する手法です．
下記URLから"pytorch_grad_cam"をダウンロード．
https://github.com/jacobgil/pytorch-grad-cam/tree/master/pytorch_grad_cam
下記関数をTRAIN終了後に実行．

from pytorch_grad_cam import grad_cam
from pytorch_grad_cam.utils import image

def print_GradCAM(test_loader,model,cv,TEST_NAME_LIST):
    
    model.eval()

    for i, (images, labels) in enumerate(test_loader):
        #labels = labels.type(torch.LongTensor) 
        #images, labels = images.to(device), labels.to(device)
        target_layers = [model.conv_head] # Choice final Convolution Layer
        cam = grad_cam.GradCAM(model = model, target_layers = target_layers, use_cuda = torch.cuda.is_available())
        
        target_category = None
        grayscale_cam = cam(input_tensor = images, targets = target_category)
        grayscale_cam = grayscale_cam[0, :]
        
        # plot用のイメージ
        vis_image = images.cpu().numpy()

        for n in range(len(labels)):
            vis_image_plot = vis_image[n].transpose(1,2,0)
            vis_image_plot = (vis_image_plot/ 255.).astype(np.float32) 
            visualization = image.show_cam_on_image(vis_image_plot, grayscale_cam, use_rgb = True)
            visualization = cv2.bitwise_not(visualization)
            cv2.imwrite("GRADCAM/" + TEST_NAME_LIST[i*batch_size + n] + ".png", visualization)

RNN

RNNの入力形式にデータを合わせてくれれば，Okです．
RNN・LSTM・GRUはそれぞれ"nn."を前につけるだけで実装完了です．（Pytorchならね

'''モデルの定義'''
class RNN(nn.Module):
    def __init__(self):
        super(RNN, self).__init__()
        self.rnn = nn.LSTM(num_features, 1000, batch_first=True)
        self.fc1 = nn.Linear(1000, 200)
        self.fc2 = nn.Linear(200, 2)
        self.sf = nn.Softmax(dim=1)
        self.relu = nn.ReLU()
        
    def forward(self, x):
        x, hidden = self.rnn(x, None)
        #x = self.fc1(x[0])
        x = self.fc1(x[:,-1,:])
        x = self.relu(x)
        x = self.fc2(x)
        x = self.sf(x)
        
        return x

• RNN Classification
  
  

Segmentation

セグメンテーションは，Inputが画像，GTruthも画像，出力も画像です．
DataAugmentationは，IMAGEとGTruth両方に対して行います．GTruthは0,1の数値になっているので，色彩等をAugentationするとダメです．なので，GTruthにたいしては回転などの処理だけを行うようにしましょう．

def train_epoch(model, optimizer, criterion, dataloader, device):
    train_loss = 0
    model.train()
    
    for i, (images, labels) in enumerate(dataloader):
        # DataAugmentation ------#
        ### FIX SEED
        seed = random.randint(0, 2**32)
        ### IMAGE
        images = images.to(torch.uint8)
        #fix_seed(seed)
        images = share_transform(images)
        images = image_transform(images)
        images = images.to(torch.float32)
        ### LABEL
        fix_seed(seed)
        labels = share_transform(labels)
        #------------------------#
        images, labels = images.to(device), labels.to(device)
        optimizer.zero_grad()
        outputs = model(images)
        outputs = torch.sigmoid(outputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
        train_loss += loss.item()
        del images,labels,loss,outputs
    train_loss = train_loss / len(dataloader.dataset)
    
    return train_loss

画像を保存します．入力はそのまま，GTruthは(0,1)を(0,255)にしてから出力します．
出力は，(0,1)の範囲になっている筈なので，0.5以上の場合255，以下の場合0にします．

def save_image(image_number,pimg,plab,pout):
    input_path  = os.path.join(OUTPUT_DATADIR, "IMAGE")
    corect_path = os.path.join(OUTPUT_DATADIR, "MASK")
    output_path = os.path.join(OUTPUT_DATADIR, "OUTPUT")
    
    #　Original Image
    pltimg = pimg.transpose(1,2,0)
    cv2.imwrite(os.path.join(input_path, image_number),pltimg)
    
    # Ground Truth
    pltlab = plab.reshape(IMG_SIZE,IMG_SIZE)
    pltlab = np.where(pltlab == 1, 255,0)
    cv2.imwrite(os.path.join(corect_path, image_number),pltlab)
    
    # Model Output
    # Set a threshold(HyperParameter)
    pltout = pout.reshape(IMG_SIZE,IMG_SIZE)
    pltout = np.where(pltput > 0.5, 255, 0)
    cv2.imwrite(os.path.join(output_path, image_number),pltout)
    
    return pltlab,pltout

Segmentationの精度を出力する関数です．

def segmentation_score(plab_mask,plout_mask):
    smooth=1
    
    inputs  = np.ravel(plab_mask)
    targets = np.ravel(plout_mask)
    i_and_t = inputs & targets
                
    intersection = i_and_t.sum()
    FP = inputs.sum() - intersection
    FN = targets.sum() - intersection

    recall = (intersection + smooth) / (intersection + FN + smooth)
    precision = (intersection + smooth) / (intersection + FP + smooth)
    dice = (2*precision*recall) / (precision+recall)
    iou = (intersection + smooth) / (inputs.sum() + targets.sum() - intersection + smooth) 
    
    return recall,precision,dice,iou

def print_SCORE_IMAGE(test_loader,model,crossval,all_recall,all_precision,all_dice,all_iou):
    model.eval()
    recall_list    = []
    precision_list = []
    dice_list      = []
    iou_list       = []
    
    with torch.no_grad():
        for i, (images, labels) in enumerate(test_loader):
            images, labels = images.to(device), labels.to(device)
            outputs = model(images)
            
            pimg = images.to('cpu').detach().numpy().copy()
            pimg = np.vectorize(int)(pimg)
            plab = labels.to('cpu').detach().numpy().copy()
            plab = np.vectorize(int)(plab)
            pout = outputs.to('cpu').detach().numpy().copy()
            
            for li in range(len(pimg)):
                ###### IMAGE
                # image_numbe
                image_number = str(crossval)+"_"+str(i)+"_"+str(li)+".png"
                # save image
                plab_mask,plout_mask = save_image(image_number,pimg[li],plab[li],pout[li])
                ###### SCORE
                recall,precision,dice,iou = segmentation_score(plab_mask,plout_mask)
                recall_list.append(recall)
                precision_list.append(precision)
                dice_list.append(dice)
                iou_list.append(iou)
                
    # PLOT
    print(f"RECALL   :{statistics.mean(recall_list)}")
    print(f"PRECISION:{statistics.mean(precision_list)}")
    print(f"DICE     :{statistics.mean(dice_list)}")
    print(f"IOU      :{statistics.mean(iou_list)}")
    
    ## CV_ALL
    all_recall.extend(recall_list)
    all_precision.extend(precision_list)
    all_dice.extend(dice_list)
    all_iou.extend(iou_list)

メイン部分です．

# List for CV
all_recall,all_precision,all_dice,all_iou   = [],[],[],[]
# FolderName for CV
ListForCV = ["1","2","3","4","5"]
# Input Data Path
INPUT_DATADIR = "C:/Data/" # Change Your DataDir

for i in range(1,6):
    print(f"***FOLD_{i}")
    #* CreateDataset
    TRAIN_DATADIR1 = INPUT_DATADIR + ListForCV[0]
    TRAIN_DATADIR2 = INPUT_DATADIR + ListForCV[1]
    TRAIN_DATADIR3 = INPUT_DATADIR + ListForCV[2]
    VAL_DATADIR    = INPUT_DATADIR + ListForCV[3]
    TEST_DATADIR   = INPUT_DATADIR + ListForCV[4]
    train_dataset = TrainDataForCV(TRAIN_DATADIR1,TRAIN_DATADIR2,TRAIN_DATADIR3)
    val_dataset = DataForCV(VAL_DATADIR,0)
    test_dataset = DataForCV(TEST_DATADIR,1)
    train_loader   = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
    val_loader   = DataLoader(val_dataset, batch_size=batch_size, shuffle=False)
    test_loader  = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)
    #* MODEL
    model = UNet.UNet(n_channels=1, n_classes=1)
    model = model.to(device)
    #* Loss & Optimizer
    # UNet3plus => Adam
    # UTNet => SGD
    criterion = nn.BCELoss()
    #criterion = diceloss()
    optimizer = optim.Adam(model.parameters(), lr=0.0001)
    #optimizer = optim.SGD(model.parameters(), lr=0.0001, momentum=0.9)
    #* run
    train_loss_list, val_loss_list, test_loss_list = run(1, optimizer, criterion, device, train_loader, val_loader, test_loader,model)
    model = pickle.load(open('check_point_model.sav', 'rb'))
    pickle.dump(model, open('model1.sav', 'wb'))
    #* PLOT
    graph(train_loss_list, val_loss_list, test_loss_list)
    print_SCORE_IMAGE(test_loader,model,i,all_recall,all_precision,all_dice,all_iou)
    #* rotate list
    pop = ListForCV.pop(0)
    ListForCV.append(pop)
    #break

## Plot CV Mean Score
"""
print("\n\n\n*** ALL_RESULT ***")
print(f"RECALL    :{statistics.mean(all_recall)}")
print(f"PRECISION :{statistics.mean(all_precision)}")
print(f"DICE      :{statistics.mean(all_dice)}")
print(f"IOU       :{statistics.mean(all_iou)}")
"""