Question

我有一个CNN的Keras功能模型。我正在尝试实现三重损失功能。我发现了一些有关谁使用“合并”执行此操作的帖子，现已弃用，但我无法像使用合并那样使用“连接”。

原始代码如下：

def triplet_loss(x):
    anchor, positive, negative = x
    pos_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, positive)), 1)
    neg_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, negative)), 1)

    basic_loss = tf.add(tf.subtract(pos_dist, neg_dist), 0.05)
    loss = tf.reduce_mean(tf.maximum(basic_loss, 0.0), 0)
    return loss



def build_model(img_x, img_y):
    input_shape = Input(shape=(img_x, img_y, 3))
    c0 = Conv2D(32, kernel_size=(3, 3), strides=(1, 1), activation='relu') (input_shape)
    m0 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2)) (c0)
    f = Flatten()(m0)
    d1 = Dense(4024, activation='relu')(f)
    d2 = Dense(512, activation='sigmoid')(d1)

    anchor = Input(shape=(128, 254, 3))
    positive = Input(shape=(128, 254, 3))
    negative = Input(shape=(128, 254, 3))

    reid_model = Model(inputs=[input_shape], outputs=[d2])

    anchor_embed = reid_model(anchor)
    positive_embed = reid_model(positive)
    negative_embed = reid_model(negative)

    loss = merge([anchor_embed, positive_embed, negative_embed],
             mode=triplet_loss, output_shape=(1,))

    model = Model(inputs=[anchor, positive, negative], outputs=loss)
    model.compile(optimizer='Adam', loss='mean_absolute_error')
    return model

我使用loss = merge([anchor_embed, positive_embed, negative_embed], mode=triplet_loss, output_shape=(1,))作为将函数triplet_loss的输出转换为keras层输出的一种方法（如https://codepad.co/snippet/F1uVDD5N中所建议）。函数concatenate没有参数“ mode”。他有什么方法可以适应我的代码，以将损失函数的结果作为Keras层输出？

Answer 1

我终于找到了一种方法，通过添加lambda层来计算triplet_loss函数的值，从而保持代码的原始体系结构。

def triplet_loss(x):
    anchor, positive, negative = x
    pos_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, positive)), 1)
    neg_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, negative)), 1)

    basic_loss = tf.add(tf.subtract(pos_dist, neg_dist), 0.05)
    loss = tf.reduce_mean(tf.maximum(basic_loss, 0.0), 0)
    return loss

def build_model(img_x, img_y):
    input_shape = Input(shape=(img_x, img_y, 3))
    c0 = Conv2D(32, kernel_size=(3, 3), strides=(1, 1), activation='relu') 
(input_shape)
    m0 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2)) (c0)
    f = Flatten()(m0)
    d1 = Dense(4024, activation='relu')(f)
    d2 = Dense(512, activation='sigmoid')(d1)

    anchor = Input(shape=(128, 254, 3))
    positive = Input(shape=(128, 254, 3))
    negative = Input(shape=(128, 254, 3))

    reid_model = Model(inputs=[input_shape], outputs=[d2])

    anchor_embed = reid_model(anchor)
    positive_embed = reid_model(positive)
    negative_embed = reid_model(negative)

    merged_output = concatenate([anchor_embed, positive_embed, 
negative_embed])
    loss = Lambda(triplet_loss, (1,))(merged_output)

    model = Model(inputs=[anchor, positive, negative], outputs=loss)
    model.compile(optimizer='Adam', loss='mse',
                  metrics=["mae"])
    return model

Keras合并VS级联，无法更新我的代码

1 个答案: