如何在model.compile里的loss='categorical_crossentropy'的基础上利用loss_weights加上源域为Xs，目标域为Xt的MMD损失

可以在 `model.compile` 中使用 `loss_weights` 参数来加入 MMD 损失：

首先，需要定义一个 MMD 损失函数：

import tensorflow as tf
import numpy as np

def mmd_loss(source, target, kernel=rbf_kernel, sigma=1.0):
    """
    Computes the Maximum Mean Discrepancy (MMD) between source and target domains.

    :param source: Tensor, the source domain data.
    :param target: Tensor, the target domain data.
    :param kernel: Function, the kernel function to use (default: RBF kernel).
    :param sigma: Float, the sigma parameter for the RBF kernel.
    :return: Tensor, the MMD loss between source and target domains.
    """
    source_kernel = kernel(source, source, sigma=sigma)
    target_kernel = kernel(target, target, sigma=sigma)
    source_target_kernel = kernel(source, target, sigma=sigma)
    loss = tf.reduce_mean(source_kernel) + tf.reduce_mean(target_kernel) - 2 * tf.reduce_mean(source_target_kernel)
    return loss

其中，`rbf_kernel` 是高斯核函数，可以使用 `numpy` 实现：

def rbf_kernel(X, Y, sigma=1.0):
    """
    Computes the RBF kernel between two matrices.

    :param X: Tensor, the first matrix.
    :param Y: Tensor, the second matrix.
    :param sigma: Float, the sigma parameter for the RBF kernel.
    :return: Tensor, the RBF kernel between X and Y.
    """
    XX = tf.matmul(X, tf.transpose(X))
    XY = tf.matmul(X, tf.transpose(Y))
    YY = tf.matmul(Y, tf.transpose(Y))
    X_sqnorms = tf.linalg.diag_part(XX)
    Y_sqnorms = tf.linalg.diag_part(YY)
    return tf.exp(-0.5 / sigma ** 2 * (
            tf.reshape(X_sqnorms, (-1, 1)) - 2 * XY + tf.reshape(Y_sqnorms, (1, -1))))

然后，在 `model.compile` 中加入 MMD 损失：

model.compile(optimizer=optimizer,
              loss={'output': 'categorical_crossentropy', 'mmd': mmd_loss},
              loss_weights={'output': 1.0, 'mmd': lambda _: mmd_weight})

其中，`mmd_weight` 是一个超参数，用于调整 MMD 损失的权重。

在训练模型时，需要将源域和目标域的数据分别传入模型，然后将 MMD 损失乘以权重加入总损失中：

history = model.fit({'input': Xs, 'input_tgt': Xt},
                    {'output': Ys, 'mmd': tf.zeros(1)},
                    batch_size=batch_size,
                    epochs=epochs,
                    verbose=1,
                    callbacks=callbacks_list,
                    validation_data=({'input': Xs_val, 'input_tgt': Xt_val},
                                     {'output': Ys_val, 'mmd': tf.zeros(1)}))

for epoch in history.history.keys():
    if epoch.startswith('val_'):
        continue
    mmd_loss_train = mmd_weight * mmd_loss(Xs, Xt)
    mmd_loss_val = mmd_weight * mmd_loss(Xs_val, Xt_val)
    print(f'{epoch}: {history.history[epoch][-1]:.4f} (train), {history.history["val_" + epoch][-1]:.4f} (val), '
          f'{mmd_loss_train:.4f} (mmd train), {mmd_loss_val:.4f} (mmd val)')

这样就可以在训练模型时加入源域和目标域的 MMD 损失了。