我正在尝试使用Keras使用暹罗CNN训练约200对图像，并注意验证精度在各个时期都不会改变。
Train on 144 samples, validate on 16 samples
Epoch 1/20
144/144 [==============================] - 51s 352ms/step - loss: 0.3041 - accuracy: 0.4375 - val_loss: 0.4816 - val_accuracy: 0.5000
Epoch 2/20
144/144 [==============================] - 56s 387ms/step - loss: 0.2819 - accuracy: 0.5208 - val_loss: 0.4816 - val_accuracy: 0.5000
Epoch 3/20
144/144 [==============================] - 47s 325ms/step - loss: 0.2784 - accuracy: 0.4861 - val_loss: 0.4816 - val_accuracy: 0.5000
Epoch 00003: ReduceLROnPlateau reducing learning rate to 0.0001500000071246177.
Epoch 4/20
144/144 [==============================] - 50s 349ms/step - loss: 0.2865 - accuracy: 0.4306 - val_loss: 0.4816 - val_accuracy: 0.5000
Epoch 5/20
144/144 [==============================] - 54s 377ms/step - loss: 0.2936 - accuracy: 0.4375 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 00005: ReduceLROnPlateau reducing learning rate to 4.500000213738531e-05.
Epoch 6/20
144/144 [==============================] - 50s 349ms/step - loss: 0.2980 - accuracy: 0.4097 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 7/20
144/144 [==============================] - 47s 324ms/step - loss: 0.2824 - accuracy: 0.4931 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 00007: ReduceLROnPlateau reducing learning rate to 1.3500000204658135e-05.
Epoch 8/20
144/144 [==============================] - 48s 336ms/step - loss: 0.2888 - accuracy: 0.4722 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 9/20
144/144 [==============================] - 45s 315ms/step - loss: 0.2572 - accuracy: 0.5417 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 00009: ReduceLROnPlateau reducing learning rate to 4.050000006827758e-06.
Epoch 10/20
144/144 [==============================] - 45s 313ms/step - loss: 0.2827 - accuracy: 0.5139 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 11/20
144/144 [==============================] - 46s 318ms/step - loss: 0.2660 - accuracy: 0.5764 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 00011: ReduceLROnPlateau reducing learning rate to 1.2149999747634864e-06.
Epoch 12/20
144/144 [==============================] - 58s 401ms/step - loss: 0.2869 - accuracy: 0.4583 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 13/20
144/144 [==============================] - 60s 417ms/step - loss: 0.2779 - accuracy: 0.5486 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 00013: ReduceLROnPlateau reducing learning rate to 3.644999992502562e-07.
Epoch 14/20
144/144 [==============================] - 51s 357ms/step - loss: 0.2959 - accuracy: 0.4722 - val_loss: 0.4815 - val_accuracy: 0.5000
Epoch 15/20
144/144 [==============================] - 49s 343ms/step - loss: 0.2729 - accuracy: 0.5069 - val_loss: 0.4815 - val_accuracy: 0.5000
我的神经网络如下所示：
input_shape = X_train.shape[1:]
model = Sequential()
model.add(Conv2D(nb_filters, nb_conv, border_mode = 'valid', input_shape=(1, img_rows, img_cols), data_format = 'channels_first'))
model.add(Conv2D(32, kernel_size=(3, 3),activation='relu',input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(nb_classes, activation='sigmoid'))
left_input = Input(input_shape)
right_input = Input(input_shape)
processed_a = model(left_input)
processed_b = model(right_input)
distance = Lambda(euclidean_distance,output_shape=eucl_dist_output_shape)([processed_a, processed_b])
siamese_net = Model([left_input, right_input], distance)
我尝试了不同的优化器，不同的学习率和正则化（辍学），但验证准确性/损失没有改变。
如何改进它？