损失没有变化，准确性保持不变。

left_input = Input((3,224,224))
right_input = Input((3,224,224))
encoded_l = encoder2(left_input)
encoded_r = encoder2(right_input)

# Getting the L1 Distance between the 2 encodings
L1_layer = L.Lambda(lambda tensor:K.abs(tensor[0] - tensor[1]))

L1_distance = L1_layer([encoded_l, encoded_r])
prediction = L.Dense(1,activation='sigmoid')(L1_distance)
siamese_net = Model(inputs=[left_input,right_input],outputs=prediction)
optimizer = Adam(0.0005, decay=2.5e-4)
siamese_net.compile(loss="binary_crossentropy",optimizer=optimizer,metrics=['accuracy'])

siamese_net.fit([Xl_train.reshape((-1,3,224,224)),Xr_train.reshape((-1,3,224,224))], to_categorical(y_train),
          batch_size=32,
          epochs=50,
          verbose=1,
          validation_data=([Xl_test.reshape((-1,3,224,224)),Xr_test.reshape((-1,3,224,224))],to_categorical(y_test)),
          shuffle=True)

我只设置了正对，因此我必须通过在对中的一组成员上使用 np.roll() 来设置负对，以产生新的组合。因此，由于不平衡问题，它不能预测 1（因为正样本和负样本具有相同的对数）。我尝试过改组数据集，在最后一层使用带有 2 个节点的 softmax。我在带有 rbf 内核的 SVM 上进行了尝试。虽然准确度很差，但它仍然至少预测为 0。这是编码器的代码-

def encoder():
    model=Sequential()
    model.add(Convolution2D(128, (7, 7), activation='relu'))
    model.add(Dropout(0.2))
    model.add(Convolution2D(64, (5, 5), activation='relu'))
    model.add(Dropout(0.2))
    model.add(Convolution2D(64, (5, 5), activation='relu'))
    model.add(Dropout(0.2))
    model.add(Convolution2D(16, (5, 5), activation='relu'))
    model.add(Dropout(0.2))
    model.add(Convolution2D(1, (5, 5), activation='relu'))
    model.add(Flatten())
    model.build((None,3, 224, 224))

    #print(model.summary())
    return model

Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 128, 218, 218)     18944     
_________________________________________________________________
dropout_1 (Dropout)          (None, 128, 218, 218)     0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 64, 214, 214)      204864    
_________________________________________________________________
dropout_2 (Dropout)          (None, 64, 214, 214)      0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 64, 210, 210)      102464    
_________________________________________________________________
dropout_3 (Dropout)          (None, 64, 210, 210)      0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 16, 206, 206)      25616     
_________________________________________________________________
dropout_4 (Dropout)          (None, 16, 206, 206)      0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 1, 202, 202)       401       
_________________________________________________________________
flatten_1 (Flatten)          (None, 40804)             0         
=================================================================
Total params: 352,289
Trainable params: 352,289
Non-trainable params: 0

编辑 1- 我添加了- kernel_initializer='random_uniform',bias_initializer='zeros'

到所有层，以便它停止将所有内容预测为 0。训练损失现在正在减少，但验证损失和准确性没有改变。

编辑2-解决了。这是我的数据集的问题。