在 keras 中同时训练神经网络并让它们在训练时共同分担损失？

Question

假设我想同时训练三个模型（模型 1、模型 2 和模型 3），并且在训练时让模型一和模型二与主网络（模型 1）共同共享损失。因此主模型可以从层间的其他两个模型中学习表示。损失总计 = (权重1)损失m1 + (权重2)(损失m1 - 损失m2) + (权重3)(损失m1 - 损失m3)到目前为止我有以下内容：def threemodel(num_nodes, num_class, w1, w2, w3):    #w1; w2; w3 are loss weights        in1 = Input((6373,))    enc1 = Dense(num_nodes)(in1)    enc1 = Dropout(0.3)(enc1)    enc1 = Dense(num_nodes, activation='relu')(enc1)    enc1 = Dropout(0.3)(enc1)    enc1 = Dense(num_nodes, activation='relu')(enc1)    out1 = Dense(units=num_class, activation='softmax')(enc1)        in2 = Input((512,))    enc2 = Dense(num_nodes, activation='relu')(in2)    enc2 = Dense(num_nodes, activation='relu')(enc2)        out2 = Dense(units=num_class, activation='softmax')(enc2)        in3 = Input((768,))    enc3 = Dense(num_nodes, activation='relu')(in3)    enc3 = Dense(num_nodes, activation='relu')(enc3)        out3 = Dense(units=num_class, activation='softmax')(enc3)        adam = Adam(lr=0.0001)        model = Model(inputs=[in1, in2, in3], outputs=[out1, out2, out3])        model.compile(loss='categorical_crossentropy', #continu together          optimizer='adam',          metrics=['accuracy'] not sure know what changes need to be made here)## I am confused on how to formulate the shared losses equation here to share the losses of out2 and out3 with out1.经过一番搜索后，似乎可以执行以下操作：loss_1 = tf.keras.losses.categorical_crossentropy(y_true_1, out1)  loss_2 = tf.keras.losses.categorical_crossentropy(y_true_2, out2)  loss_3 = tf.keras.losses.categorical_crossentropy(y_true_3, out3)  model.add_loss((w1)*loss_1 + (w2)*(loss_1 - loss_2) + (w3)*(loss_1 - loss_3))这可以吗？我觉得通过执行上面建议的操作并没有真正执行我想要的操作，即让主模型（mod1）从各层之间的其他两个模型（mod2 和 mod3）学习表示。有什么建议么？

Answer 1

由于您对使用可训练权重不感兴趣（我将它们标记为系数以将它们与可训练权重区分开），您可以连接输出并将它们作为单个输出传递给自定义损失函数。这意味着这些系数将在训练开始时可用。

您应该提供如上所述的自定义损失函数。损失函数预计只接受 2 个参数，因此您应该使用这样一个函数categorical_crossentropy，它也应该熟悉您感兴趣的参数，例如coeffs和num_class。因此，我使用所需的参数实例化一个包装函数，然后将内部实际损失函数作为主损失函数传递。

from tensorflow.keras.layers import Dense, Dropout, Input, Concatenate

from tensorflow.keras.optimizers import Adam

from tensorflow.keras.models import Model

from tensorflow.python.framework import ops

from tensorflow.python.framework import smart_cond

from tensorflow.python.ops import math_ops

from tensorflow.python.ops import array_ops

from tensorflow.python.keras import backend as K

def categorical_crossentropy_base(coeffs, num_class):

    def categorical_crossentropy(y_true, y_pred, from_logits=False, label_smoothing=0):

        """Computes the categorical crossentropy loss.

      Args:

        y_true: tensor of true targets.

        y_pred: tensor of predicted targets.

        from_logits: Whether `y_pred` is expected to be a logits tensor. By default,

          we assume that `y_pred` encodes a probability distribution.

        label_smoothing: Float in [0, 1]. If > `0` then smooth the labels.

      Returns:

        Categorical crossentropy loss value.

        https://github.com/tensorflow/tensorflow/blob/v1.15.0/tensorflow/python/keras/losses.py#L938-L966

      """

        y_pred1 = y_pred[:, :num_class]  # the 1st prediction

        y_pred2 = y_pred[:, num_class:2*num_class]  # the 2nd prediction

        y_pred3 = y_pred[:, 2*num_class:]  # the 3rd prediction

        # you should adapt the ground truth to contain all 3 ground truth of course

        y_true1 = y_true[:, :num_class]  # the 1st gt

        y_true2 = y_true[:, num_class:2*num_class]  # the 2nd gt

        y_true3 = y_true[:, 2*num_class:]  # the 3rd gt

        loss1 = K.categorical_crossentropy(y_true1, y_pred1, from_logits=from_logits)

        loss2 = K.categorical_crossentropy(y_true2, y_pred2, from_logits=from_logits)

        loss3 = K.categorical_crossentropy(y_true3, y_pred3, from_logits=from_logits)

        # combine the losses the way you like it

        total_loss = coeffs[0]*loss1 + coeffs[1]*(loss1 - loss2) + coeffs[2]*(loss2 - loss3)

        return total_loss

    return categorical_crossentropy

in1 = Input((6373,))

enc1 = Dense(num_nodes)(in1)

enc1 = Dropout(0.3)(enc1)

enc1 = Dense(num_nodes, activation='relu')(enc1)

enc1 = Dropout(0.3)(enc1)

enc1 = Dense(num_nodes, activation='relu')(enc1)

out1 = Dense(units=num_class, activation='softmax')(enc1)

in2 = Input((512,))

enc2 = Dense(num_nodes, activation='relu')(in2)

enc2 = Dense(num_nodes, activation='relu')(enc2)

out2 = Dense(units=num_class, activation='softmax')(enc2)

in3 = Input((768,))

enc3 = Dense(num_nodes, activation='relu')(in3)

enc3 = Dense(num_nodes, activation='relu')(enc3)

out3 = Dense(units=num_class, activation='softmax')(enc3)

adam = Adam(lr=0.0001)

total_out = Concatenate(axis=1)([out1, out2, out3])

model = Model(inputs=[in1, in2, in3], outputs=[total_out])

coeffs = [1, 1, 1]

model.compile(loss=categorical_crossentropy_base(coeffs=coeffs, num_class=num_class),  optimizer='adam', metrics=['accuracy'])

不过，我不确定有关准确性的指标。但我认为无需其他更改即可发挥作用。我也在使用K.categorical_crossentropy，但是您当然也可以自由地使用其他实现来更改它。