困惑度先降后升，但损失率下降

Question

import keras as K
def perplexity(y_true, y_pred):
    cross_entropy = K.backend.mean(K.backend.sparse_categorical_crossentropy(y_true, y_pred))
    perplexity = K.backend.exp(cross_entropy)
    return perplexity

model.compile(
    optimizer=optimizer,
    loss='sparse_categorical_crossentropy',
    metrics=[perplexity]
)

我检查了张量板，损耗不断减小，但困惑度先减小然后增大。