diff --git a/ebus_movie_classify.py b/ebus_movie_classify.py
index 2477ce3..a86e22c 100644
--- a/ebus_movie_classify.py
+++ b/ebus_movie_classify.py
@@ -13,6 +13,7 @@
 import torchvision
 from torchvision import models, transforms
 import sys
+import datetime
 
 # 入力画像の前処理クラス
 # 訓練時と推論時で処理を変える
@@ -75,7 +76,7 @@
         データへのパスを格納したリスト
     """
 
-    rootpath = "/mnt/d/usr/DL/EBUS/"
+    rootpath = "/data2/EBUS/EBUS動画20220124/ExtractedFrames/"
     target_path = osp.join(rootpath + phase + "/" + label + '/**/*.png')
     # print(target_path)
     # print(len(glob.glob(target_path, recursive=True)))
@@ -137,18 +138,13 @@
     net.to(device)
     # ネットワークがある程度固定なら高速化
     torch.backends.cudnn.benchmark = True
-
-    # ログ出力
-    f = open('result.csv', 'w')
-    writer = csv.writer(f, lineterminator='\n')
-    writer.writerow(['epoch', 'train_loss', 'train_acc', 'val_loss', 'val_acc'])
+    training_result = []
     
     # epochのループ
     for epoch in range(num_epochs):
-        csvitems = []
-        csvitems.append(epoch+1)
         print('')
         print('Epoch {}/{}'.format(epoch+1, num_epochs))
+        epoch_result = [epoch + 1]
 
         # epochごとの学習と検証のループ
         for phase in ['train', 'val']:
@@ -162,8 +158,7 @@
 
             # 未学習時の性能を確かめるため epoch=0 の訓練は省略
             if (epoch == 0) and (phase == 'train'):
-                csvitems.append(0)
-                csvitems.append(0)
+                epoch_result.extend([0, 0])
                 continue
 
             # データローダーからミニバッチを取り出すループ
@@ -196,15 +191,12 @@
             epoch_loss = epoch_loss / len(dataloaders_dict[phase].dataset)
             epoch_acc = epoch_corrects.double() / len(dataloaders_dict[phase].dataset)
             
-            print('{} Loss: {:.4f} Acc: {:.4f}'.format(
-                phase, epoch_loss, epoch_acc))
-            csvitems.append('{:.4f}'.format(epoch_loss))
-            csvitems.append('{:.4f}'.format(epoch_acc))
+            print('{} Loss: {:.4f} Acc: {:.4f}'.format(phase, epoch_loss, epoch_acc))
+            epoch_result.extend([epoch_loss, float(epoch_acc)])
         
-        writer.writerow(csvitems)
-    
-    f.close()
+        training_result.append(epoch_result)
 
+    return training_result
 
 if __name__ == "__main__":
 
@@ -225,7 +217,7 @@
     val_malignant_list = make_datapath_list(phase="val", label="Malignant")
     print("Dataset(orig): train benign:%d malignant:%d  val benign:%d malignant:%d" %
         (len(train_benign_list), len(train_malignant_list), len(val_benign_list), len(val_malignant_list) ))
-    num_traindata = 200
+    num_traindata = 1000
     num_valdata = 100
     train_benign_list = random.sample(train_benign_list, num_traindata)
     train_malignant_list = random.sample(train_malignant_list, num_traindata)
@@ -256,16 +248,23 @@
 
     use_pretrained = True
     # ResNet18
-    # net = models.resnet18(pretrained=use_pretrained)
-    # net.fc = nn.Linear(in_features=512, out_features=2, bias=True)
-    net = models.resnet101(pretrained=use_pretrained)
-    net.fc = nn.Linear(in_features=2048, out_features=2, bias=True)
+    net = models.resnet18(pretrained=use_pretrained)
+    model_name = "ResNet18"
+    net.fc = nn.Linear(in_features=512, out_features=2, bias=True)
+    # net = models.resnet101(pretrained=use_pretrained)
+    # model_name = "ResNet101"
+    # net.fc = nn.Linear(in_features=2048, out_features=2, bias=True)
     # print(net)
     # sys.exit()
-    update_param_names = ["fc.weight", "fc.bias"]   # 学習させるパラメータ名
+    # 学習するレイヤーの指定
+    update_param_names = [name for name, param in net.named_parameters() if "fc." in name]
+    update_layer = "FC"
+    # update_param_names = [name for name, param in net.named_parameters() if "fc." in name or "layer4" in name]
+    # update_layer = "FC and Layer4"
 
     # VGG16
     # net = models.vgg16(pretrained=use_pretrained)
+    # model_name = "VGG16"
     # net.classifier[6] = nn.Linear(in_features=4096, out_features=2)
     # update_param_names = ["classifier.3.weight", "classifier.3.bias", "classifier.6.weight", "classifier.6.bias"]
     # update_param_names = ["classifier.6.weight", "classifier.6.bias"]
@@ -300,9 +299,34 @@
     # print(params_to_update)
 
     # 最適化手法の設定
-    optimizer = optim.SGD(params=params_to_update, lr=0.001, momentum=0.9)
+    learning_rate = 0.001
+    optimizer = optim.SGD(params=params_to_update, lr=learning_rate, momentum=0.9)
 
     # 学習・検証を実行する
-    num_epochs = 50
-    train_model(net, dataloaders_dict, criterion, optimizer, num_epochs=num_epochs)
+    num_epochs = 10
+    training_result = train_model(net, dataloaders_dict, criterion, optimizer, num_epochs)
+
+    # 学習カーブを解析
+    result_max = np.amax(training_result, axis=0)[1:]
+    result_min = np.amin(training_result, axis=0)[1:]
+    result_median = np.median(training_result, axis=0)[1:]
+
+    # ログ出力
+    f = open("{0:result%Y%m%d_%H%M%S.csv}".format(datetime.datetime.now()), 'w')
+    writer = csv.writer(f, lineterminator='\n')
+    writer.writerow(['Model','Update Layer','#Training','#Validation','Batch Size','Learning Rate','#Epoch','ValAccMax','ValAccMedian'])
+    writer.writerow([model_name, update_layer, num_traindata, num_valdata, batch_size, learning_rate, num_epochs, result_max[3], result_median[3]])
+    writer.writerow([])
+    writer.writerow(['epoch', 'train_loss', 'train_acc', 'val_loss', 'val_acc'])
+    np.savetxt(f, training_result, fmt=['%.0f', '%.4f', '%.4f', '%.4f', '%.4f'], delimiter=',', newline='\n')
+    writer.writerow([])
+    f.write('max,')
+    np.savetxt(f, result_max.reshape((1, 4)), fmt='%.4f', delimiter=',', newline='\n')
+    f.write('min,')
+    np.savetxt(f, result_min.reshape((1, 4)), fmt='%.4f', delimiter=',', newline='\n')
+    f.write('median,')
+    np.savetxt(f, result_median.reshape((1, 4)), fmt='%.4f', delimiter=',', newline='\n')
+
+    # 終了処理
+    f.close()
     print('done.')