simulations.ipynb

Kernel: Python 3

Copyright 2019 The TensorFlow Authors.

#@title Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

High-performance simulations with TFF

本教程将介绍如何在各种常见场景中使用 TFF 设置高性能模拟。

TODO(b/134543154)：填充内容，此处涉及以下几点：

• using GPUs in a single-machine setup,

• multi-machine setup on GCP/GKE, with and without TPUs,

• interfacing MapReduce-like backends,

• current limitations and when/how they will be relaxed.

在 TensorFlow.org 上查看

在 Google Colab 中运行

在 Github 上查看源代码

下载笔记本

准备工作

首先，请确保您的笔记本电脑连接到具有已编译相关组件（包括用于多机场景的 gRPC 依赖关系）的后端。

现在，我们从加载 TFF 网站上的 MNIST 示例开始，然后声明 Python 函数，该函数将在一组 10 个客户端上运行一个小型实验循环。

#@test {"skip": true}
!pip install --quiet --upgrade tensorflow-federated

/bin/sh: pip: command not found

import collections
import time

import tensorflow as tf

import tensorflow_federated as tff

source, _ = tff.simulation.datasets.emnist.load_data()


def map_fn(example):
  return collections.OrderedDict(
      x=tf.reshape(example['pixels'], [-1, 784]), y=example['label'])


def client_data(n):
  ds = source.create_tf_dataset_for_client(source.client_ids[n])
  return ds.repeat(10).shuffle(500).batch(20).map(map_fn)


train_data = [client_data(n) for n in range(10)]
element_spec = train_data[0].element_spec


def model_fn():
  model = tf.keras.models.Sequential([
      tf.keras.layers.InputLayer(input_shape=(784,)),
      tf.keras.layers.Dense(units=10, kernel_initializer='zeros'),
      tf.keras.layers.Softmax(),
  ])
  return tff.learning.models.from_keras_model(
      model,
      input_spec=element_spec,
      loss=tf.keras.losses.SparseCategoricalCrossentropy(),
      metrics=[tf.keras.metrics.SparseCategoricalAccuracy()])


trainer = tff.learning.algorithms.build_weighted_fed_avg(
    model_fn, client_optimizer_fn=lambda: tf.keras.optimizers.SGD(0.02))


def evaluate(num_rounds=10):
  state = trainer.initialize()
  for _ in range(num_rounds):
    t1 = time.time()
    result = trainer.next(state, train_data)
    state = result.state
    train_metrics = result.metrics['client_work']['train']
    t2 = time.time()
    print('train metrics {m}, round time {t:.2f} seconds'.format(
        m=train_metrics, t=t2 - t1))

单机模拟

现在默认打开。

evaluate()

train metrics OrderedDict([('sparse_categorical_accuracy', 0.15329218), ('loss', 2.918891), ('num_examples', 9720), ('num_batches', 490)]), round time 4.64 seconds
train metrics OrderedDict([('sparse_categorical_accuracy', 0.18004115), ('loss', 2.7677088), ('num_examples', 9720), ('num_batches', 490)]), round time 2.37 seconds
train metrics OrderedDict([('sparse_categorical_accuracy', 0.21841563), ('loss', 2.511075), ('num_examples', 9720), ('num_batches', 490)]), round time 2.30 seconds
train metrics OrderedDict([('sparse_categorical_accuracy', 0.27160493), ('loss', 2.340346), ('num_examples', 9720), ('num_batches', 490)]), round time 2.25 seconds
train metrics OrderedDict([('sparse_categorical_accuracy', 0.34115225), ('loss', 2.0537064), ('num_examples', 9720), ('num_batches', 490)]), round time 2.27 seconds
train metrics OrderedDict([('sparse_categorical_accuracy', 0.3745885), ('loss', 1.9158486), ('num_examples', 9720), ('num_batches', 490)]), round time 2.21 seconds
train metrics OrderedDict([('sparse_categorical_accuracy', 0.41502059), ('loss', 1.7523248), ('num_examples', 9720), ('num_batches', 490)]), round time 2.19 seconds
train metrics OrderedDict([('sparse_categorical_accuracy', 0.47644034), ('loss', 1.6085855), ('num_examples', 9720), ('num_batches', 490)]), round time 2.20 seconds
train metrics OrderedDict([('sparse_categorical_accuracy', 0.5126543), ('loss', 1.5272282), ('num_examples', 9720), ('num_batches', 490)]), round time 2.27 seconds
train metrics OrderedDict([('sparse_categorical_accuracy', 0.5576132), ('loss', 1.393721), ('num_examples', 9720), ('num_batches', 490)]), round time 2.16 seconds

在 GCP/GKE、GPU、TPU 等上进行多机模拟…

敬请期待。