我正在嘗試將此tflearn DCNN示例（使用圖像預(yù)處理和augmemtation）轉(zhuǎn)換為keras：Tflearn示例：import tflearnfrom tflearn.data_utils import shuffle, to_categoricalfrom tflearn.layers.core import input_data, dropout, fully_connectedfrom tflearn.layers.conv import conv_2d, max_pool_2dfrom tflearn.layers.estimator import regressionfrom tflearn.data_preprocessing import ImagePreprocessingfrom tflearn.data_augmentation import ImageAugmentation# Data loading and preprocessingfrom tflearn.datasets import cifar10(X, Y), (X_test, Y_test) = cifar10.load_data()X, Y = shuffle(X, Y)Y = to_categorical(Y, 10)Y_test = to_categorical(Y_test, 10)# Real-time data preprocessingimg_prep = ImagePreprocessing()img_prep.add_featurewise_zero_center()img_prep.add_featurewise_stdnorm()# Real-time data augmentationimg_aug = ImageAugmentation()img_aug.add_random_flip_leftright()img_aug.add_random_rotation(max_angle=25.)# Convolutional network buildingnetwork = input_data(shape=[None, 32, 32, 3], data_preprocessing=img_prep, data_augmentation=img_aug)network = conv_2d(network, 32, 3, activation='relu')network = max_pool_2d(network, 2)network = conv_2d(network, 64, 3, activation='relu')network = conv_2d(network, 64, 3, activation='relu')network = max_pool_2d(network, 2)network = fully_connected(network, 512, activation='relu')network = dropout(network, 0.5)50個紀元后產(chǎn)生了以下結(jié)果：Training Step: 26050 | total loss: 0.35260 | time: 144.306s| Adam | epoch: 050 | loss: 0.35260 - acc: 0.8785 | val_loss: 0.64622 - val_acc: 0.8212 -- iter: 50000/50000然后，我嘗試使用相同的DCNN圖層，參數(shù)和圖像預(yù)處理/增強功能將其轉(zhuǎn)換為Keras這會產(chǎn)生差得多的驗證準確性結(jié)果：Epoch 50/50521/521 [==============================] - 84s 162ms/step - loss: 0.4723 - acc: 0.8340 - val_loss: 3.2970 - val_acc: 0.2729Test score: 3.2969648239135743Accuracy: 27.29%誰能幫我理解原因？我在Keras中是否誤用了/誤解了圖像預(yù)處理/增強功能？
1 回答
桃花長相依
TA貢獻1860條經(jīng)驗 獲得超8個贊
在Keras模型中，您也忘記了標準化驗證數(shù)據(jù)。您可以通過使用訓練數(shù)據(jù)datagen.mean并對其進行datagen.std計算來做到這一點：
# normalize test data; add a small constant to avoid division by zero,
# you can alternatively use `keras.backend.epsilon()`
X_test = (X_test - datagen.mean) / (datagen.std + 1e-8)
或者您可以使用該datagen.standardize()方法對測試數(shù)據(jù)進行規(guī)范化：
X_test = datagen.standardize(X_test)
有關(guān)更多信息，請參見SO上的這個問題：預(yù)測期間，數(shù)據(jù)標準化如何在keras中起作用？
別忘了您應(yīng)該通過對訓練數(shù)據(jù)計算出的統(tǒng)計數(shù)據(jù)對測試數(shù)據(jù)進行歸一化。永遠不要通過自己的統(tǒng)計數(shù)據(jù)對測試數(shù)據(jù)進行歸一化。
警告：似乎也standardize 改變了它的參數(shù)...是的，您可以在源代碼中確認這一點。
添加回答
舉報
0/150
提交
取消