与timedistributedlayer相关的tf2.0内存泄漏（急切）

Question

每次运行tape.gradients部分时，内存使用量都会翻倍。我认为它与时间分布层有关...？有什么想法吗？

###### create model 

    inputs = tf.keras.Input(shape=(6, *data_loader_train[0][0][0].shape), name='img') ## (108, 192, 3)
    x = layers.TimeDistributed(layers.Conv2D(16, 3, activation='relu'))(inputs)
    x = layers.TimeDistributed(layers.Conv2D(16, 3, activation='relu'))(x)
    block_1_output = layers.TimeDistributed(layers.MaxPooling2D(2))(x)

    x = layers.TimeDistributed(layers.Conv2D(16, 3, activation='relu', padding='same'))(block_1_output)
    block_3_output = layers.add([x, block_1_output])
    block_3_output = layers.TimeDistributed(layers.MaxPooling2D(2))(block_3_output)

    x = layers.TimeDistributed(layers.Conv2D(16, 3, activation='relu'))(block_3_output)
    x = layers.TimeDistributed(layers.GlobalAveragePooling2D())(x)

    x = layers.Flatten()(x)
    x = layers.Dense(16, activation='relu')(x)
    x = layers.Dense(1)(x)
    counts = tf.keras.activations.softplus(x)

    model = tf.keras.Model(inputs, counts, name='toy_resnet')
    model.summary()

    ### run model
####### running this part doubles memory every two times ##########
for x_ in batch(np.random.uniform(size=(100,6,108,192,3)).astype(np.float32), 10):
     with tf.GradientTape() as tape:
             count_ = tf.reduce_sum(model(x_))