# 印刷字體辨識模型訓練紀錄 準確度與誤差值計算方式 ```python3 model = load_model("TrainChar.h5") # TrainChar.h5為訓練好後儲存的模型 model.evaluate(test_img_normalize, test_label_onehot) # test_img_normalize 為經過前處理的測試資料 # test_label_onehot 為轉為one hot encoding 的測試資料標籤 ``` ---- ## 第一次訓練 acc: 0.533 loss: 30.989 - 卷積神經網路(CNN) - 2層卷積 - 2層池化 - 2層隱藏 ```python3 # 卷積層 filters = 100, kernel_size=(3,3), padding = 'same', input_shape = (28, 28, 1), activation = 'relu' filters = 200, kernel_size=(3,3), padding = 'same', activation = 'relu' # 池化層 pool_size = (2, 2) # 隱藏層 units = 512, activation = 'relu' units = 512, activation = 'relu' ``` - 結果 [30.989418029785156, 0.5333333353201548] ---- ## 第二次訓練 acc: 0.499 loss: 35.235 - 卷積神經網路(CNN) - 2層卷積 - 2層池化 - 2層隱藏 ```python3 # 卷積層 filters = 50, kernel_size=(3,3), padding = 'same', input_shape = (28, 28, 1), activation = 'relu' filters = 100, kernel_size=(3,3), padding = 'same', activation = 'relu' # 池化層 pool_size = (2, 2) # 隱藏層 units = 512, activation = 'relu' units = 512, activation = 'relu' ``` - 結果 [35.234667777565562, 0.4986774834588013] ---- ## 第三次訓練 acc: 0.667 loss: 15.201 - 卷積神經網路(CNN) - 3層卷積 - 3層池化 - 2層隱藏 ```python3 # 卷積層 filters = 200, kernel_size=(3,3), padding = 'same', input_shape = (28, 28, 1), activation = 'relu' filters = 400, kernel_size=(3,3), padding = 'same', activation = 'relu' filters = 800, kernel_size=(3,3), padding = 'same', activation = 'relu' # 池化層 pool_size = (2, 2) # 隱藏層 units = 512, activation = 'relu' units = 512, activation = 'relu' ``` - 結果 [15.201334381103516, 0.6666666690508525]