tf.contrib.tpu.keras_to_tpu_model

Копирует model вместе с весами на TPU. (устарело)

tf.contrib.tpu.keras_to_tpu_model(
    model, strategy=None
)

Возвращает модель TPU.

Использование:

a = Input(shape=(32,))
b = Dense(32)(a)
model = Model(inputs=a, outputs=b)

# If `num_cores_per_host` is greater than one, batch parallelism will be used
# to run on multiple TPU cores.
strategy = keras_support.TPUDistributionStrategy(tpu_cluster_resolver)
model = keras_support.tpu_model(model, strategy)
model.compile(
    optimizer=tf.compat.v1.train.GradientDescentOptimizer(learning_rate=1.0),
    ...)