我正在尝试在Siamese神经网络中使用基于余弦的相似度，以下是我的尝试
输入和标签
EXAMPLES=10000FEATURES=30LEFT=np.random.random((EXAMPLES,FEATURES))RIGHT=np.random.random((EXAMPLES,FEATURES))LABELS=[]for i in range(EXAMPLES): LABELS.append(np.random.randint(0,2))LABELS=np.asarray(LABELS)余弦相似度
def cosine_distance(vecs): #我对这个函数也不是很确定 y_true, y_pred = vecs y_true = K.l2_normalize(y_true, axis=-1) y_pred = K.l2_normalize(y_pred, axis=-1) return K.mean(1 - K.sum((y_true * y_pred), axis=-1))def cosine_dist_output_shape(shapes): shape1, shape2 = shapes print((shape1[0], 1)) return (shape1[0], 1)Siamese模型
inputShape=Input(shape=(FEATURES,))left_input = Input(shape=(FEATURES,))right_input = Input(shape=(FEATURES,)) model = Sequential()model.add(Dense(20, activation='relu', input_shape=(30,)))model.add(BatchNormalization())model.add(Dense(10, activation='relu')) encoded_l = model(left_input)encoded_r = model(right_input)L1_Distance = Lambda(cosine_distance, output_shape=cosine_dist_output_shape)([encoded_l, encoded_r])siamese_net = Model([left_input, right_input], L1_Distance)siamese_net.summary() siamese_net.compile(loss="mse",optimizer=Adam(lr=0.0001))siamese_net.fit(x=[LEFT,RIGHT],y=LABELS,batch_size=64,epochs=100)基于SoftMax的输出
model = Sequential()model.add(Dense(20, activation='relu', input_shape=(30,)))model.add(BatchNormalization())model.add(Dense(10, activation='relu'))#model.add(Dense(30, activation='relu')) encoded_l = model(left_input)encoded_r = model(right_input)L1_Layer = Lambda(cosine_distance, output_shape=cosine_dist_output_shape)([encoded_l, encoded_r])L1_Diatance = L1_layer([encoded_l, encoded_r])prediction = Dense(2,activation='softmax')(L1_Diatance)siamese_net = Model([left_input, right_input], prediction)siamese_net.compile(loss="binary_crossentropy",optimizer=Adam(lr=0.001))siamese_net.summary()Model: "model_26"__________________________________________________________________________________________________Layer (type) Output Shape Param # Connected to ==================================================================================================input_126 (InputLayer) (None, 30) 0 __________________________________________________________________________________________________input_127 (InputLayer) (None, 30) 0 __________________________________________________________________________________________________sequential_42 (Sequential) (None, 10) 910 input_126[0][0] input_127[0][0] __________________________________________________________________________________________________lambda_19 (Lambda) multiple 0 sequential_42[1][0] sequential_42[2][0] __________________________________________________________________________________________________dense_133 (Dense) (None, 2) 22 lambda_19[9][0] 我的模型运行正常，但我的问题是，在使用余弦相似度后，使用mse损失函数来拟合这个模型是否正确？
回答：
应该是Model([left_input, right_input], L1_Distance)而不是Model([left_input, left_input], L1_Distance)
编辑：如果你的任务是回归问题，mse可能是一个好的选择。如果你的任务是分类问题，可能需要更改损失函数（binary_crossentropy？）。还要注意，你的最后一层计算的是距离，但在分类问题的情况下，其输出必须被解释为概率分数
