Tensorflow Federated 对象不可订阅

Question

我有这样的 run_one_round 函数：

def run_one_round(server_state, federated_dataset):
    """Orchestration logic for one round of computation.
    Args:
      server_state: A `ServerState`.
      federated_dataset: A federated `tf.data.Dataset` with placement
        `tff.CLIENTS`.
    Returns:
      A tuple of updated `ServerState` and `tf.Tensor` of average loss.
    """
    tf.print("run_one_round")
    server_message = tff.federated_map(server_message_fn, server_state)
    server_message_at_client = tff.federated_broadcast(server_message)

    client_outputs = tff.federated_map(
        client_update_fn, (federated_dataset, server_message_at_client))

    weight_denom = client_outputs.client_weight

    from tensorflow_federated.python.core.impl.federated_context import value_impl
    value = value_impl.to_value(client_outputs.test, None)
    from tensorflow_federated.python.core.impl.types import placements
    from tensorflow_federated.python.core.impl.federated_context import value_utils
    value = value_utils.ensure_federated_value(value, placements.CLIENTS,
                                               'value to be averaged')

    value_comp = value.comp
    testing = []
    import sparse_ternary_compression
    for index in range(len(value_comp[0])):
        testing.append(
            sparse_ternary_compression.stc_decompression(value_comp[0][index][0], value_comp[0][index][1],
                                                         value_comp[0][index][2], value_comp[0][index][3],
                                                         value_comp[0][index][4]))

    # round_model_delta indica i pesi che vengono usati su server_update. Quindi è quello che va cambiato
    round_model_delta = tff.federated_mean(
        client_outputs.weights_delta, weight=weight_denom)

    server_state = tff.federated_map(server_update_fn, (server_state, round_model_delta))
    round_loss_metric = tff.federated_mean(client_outputs.model_output, weight=weight_denom)

    return server_state, round_loss_metric, value.comp

但是当我尝试这样做时：

value_comp = value.comp
testing = []
import sparse_ternary_compression
for index in range(len(value_comp[0])):
    testing.append(
        sparse_ternary_compression.stc_decompression(value_comp[0][index][0], value_comp[0][index][1],
                                                     value_comp[0][index][2], value_comp[0][index][3],
                                                     value_comp[0][index][4]))

我收到此错误："

File "/mnt/d/Davide/Uni/TesiMagistrale/ProgettoTesi/simple_fedavg_tff.py", line 137, in run_one_round
    for index in range(len(value_comp[0])):
TypeError: 'Call' object is not subscriptable

虽然如果我 return 值 value.comp 然后我在 main 中执行相同的操作它工作正常。

    for round_num in range(FLAGS.total_rounds):
        print("--------------------------------------------------------")
        sampled_clients = np.random.choice(train_data.client_ids, size=FLAGS.train_clients_per_round, replace=False)
        sampled_train_data = [train_data.create_tf_dataset_for_client(client) for client in sampled_clients]

代码是一样的，为什么我不能在 run_one_round 函数中使用 for 循环？

        server_state, train_metrics, value_comp = iterative_process.next(server_state, sampled_train_data)

        testing = []
        import sparse_ternary_compression
        for index in range(len(value_comp[0])):
            testing.append(sparse_ternary_compression.stc_decompression(value_comp[0][index][0], value_comp[0][index][1],
                                                                   value_comp[0][index][2], value_comp[0][index][3],
                                                                   value_comp[0][index][4]))
        print(testing)
        print(f'Round {round_num}')
        print(f'\tTraining loss: {train_metrics:.4f}')
        if round_num % FLAGS.rounds_per_eval == 0:
            server_state.model_weights.assign_weights_to(keras_model)
            accuracy = evaluate(keras_model, test_data)
            print(f'\tValidation accuracy: {accuracy * 100.0:.2f}%')
            tf.print(tf.compat.v2.summary.scalar("Accuracy", accuracy * 100.0, step=round_num))

基本上我只想访问客户端使用 client_update 发送的 test 变量，并在 tff.federated_mean 函数之前对该列表进行一些操作。

问题可能是 run_one_round 是 tff.federated_computation?

Answer 1

先完成 Building Your Own Federated Learning Algorithm 教程可能会有帮助。

我认为您应该记住的主要事情是，任何 TensorFlow 代码或任何结构操作都应该在 tff.tf_computation 修饰的方法中。然后在 tff.federated_computation 装饰方法的范围内使用 tff.federated_* 运算符连接这些构建块。

我假设您的代码片段中的 stc_decompression 是某种 TensorFlow 逻辑。但是，您传递给它的不是任何 TF 可理解的值，而是计算结果的抽象表示，这些表示主要是 TFF 的内部实现细节。

因此，无论您想用这些方法做什么，都可以在 tff.tf_computation 装饰方法中进行，您可以在其中编写任何 TF 代码。您将通过使用 tff.federated_map 运算符调用该方法来获取您的价值。

Tensorflow Federated 对象不可订阅

Tensorflow Federated object is not subscriptable

python

tensorflow

tensorflow-federated