在 TensorFlow.org 上檢視

在 Google Colab 中執行

在 GitHub 上檢視原始碼

下載筆記本

本教學課程說明如何使用 TFF 訓練非常大型的模型，其中每個用戶端裝置僅下載和更新模型的一小部分，方法是使用 tff.federated_select 和稀疏聚合。雖然本教學課程相當獨立完整，但 tff.federated_select 教學課程和 自訂 FL 演算法教學課程針對此處使用的一些技術提供了良好的簡介。

具體來說，在本教學課程中，我們考量用於多標籤分類的邏輯迴歸，根據詞袋特徵表示法預測哪些「標籤」與文字字串相關聯。重要的是，通訊和用戶端運算成本由固定常數 (MAX_TOKENS_SELECTED_PER_CLIENT) 控制，並且不會隨著整體詞彙表大小而擴展，在實際設定中，整體詞彙表大小可能非常大。

pip install --quiet --upgrade tensorflow-federated

import collections
from collections.abc import Callable
import itertools

import numpy as np
import tensorflow as tf
import tensorflow_federated as tff

每個用戶端都會 federated_select 模型權重的列，最多用於這麼多個不重複的符記。這會為用戶端本機模型的大小以及伺服器 -> 用戶端 (federated_select) 和用戶端 -> 伺服器 (federated_aggregate) 執行的通訊量設定上限。

即使您將此值設定為小至 1 (確保並非每個用戶端的所有符記都已選取) 或設定為較大的值，本教學課程仍應能正確執行，但模型收斂可能會受到影響。

MAX_TOKENS_SELECTED_PER_CLIENT = 6

我們也為各種類型定義了一些常數。對於此 Colab，符記是在剖析資料集後，特定單字的整數識別碼。

# There are some constraints on types
# here that will require some explicit type conversions:
#    - `tff.federated_select` requires int32
#    - `tf.SparseTensor` requires int64 indices.
TOKEN_DTYPE = np.int64
SELECT_KEY_DTYPE = np.int32

# Type for counts of token occurences.
TOKEN_COUNT_DTYPE = np.int32

# A sparse feature vector can be thought of as a map
# from TOKEN_DTYPE to FEATURE_DTYPE. 
# Our features are {0, 1} indicators, so we could potentially
# use np.int8 as an optimization.
FEATURE_DTYPE = np.int32

設定問題：資料集與模型

我們在本教學課程中建構了微型玩具資料集，以便於進行實驗。然而，資料集的格式與 Federated StackOverflow 相容，且 預先處理和 模型架構皆採用自 Adaptive Federated Optimization 的 StackOverflow 標籤預測問題。

資料集剖析和預先處理

NUM_OOV_BUCKETS = 1

BatchType = collections.namedtuple('BatchType', ['tokens', 'tags'])

def build_to_ids_fn(word_vocab: list[str],
                    tag_vocab: list[str]) -> Callable[[tf.Tensor], tf.Tensor]:
  """Constructs a function mapping examples to sequences of token indices."""
  word_table_values = np.arange(len(word_vocab), dtype=np.int64)
  word_table = tf.lookup.StaticVocabularyTable(
      tf.lookup.KeyValueTensorInitializer(word_vocab, word_table_values),
      num_oov_buckets=NUM_OOV_BUCKETS)

  tag_table_values = np.arange(len(tag_vocab), dtype=np.int64)
  tag_table = tf.lookup.StaticVocabularyTable(
      tf.lookup.KeyValueTensorInitializer(tag_vocab, tag_table_values),
      num_oov_buckets=NUM_OOV_BUCKETS)

  def to_ids(example):
    """Converts a Stack Overflow example to a bag-of-words/tags format."""
    sentence = tf.strings.join([example['tokens'], example['title']],
                               separator=' ')

    # We represent that label (output tags) densely.
    raw_tags = example['tags']
    tags = tf.strings.split(raw_tags, sep='|')
    tags = tag_table.lookup(tags)
    tags, _ = tf.unique(tags)
    tags = tf.one_hot(tags, len(tag_vocab) + NUM_OOV_BUCKETS)
    tags = tf.reduce_max(tags, axis=0)

    # We represent the features as a SparseTensor of {0, 1}s.
    words = tf.strings.split(sentence)
    tokens = word_table.lookup(words)
    tokens, _ = tf.unique(tokens)
    # Note:  We could choose to use the word counts as the feature vector
    # instead of just {0, 1} values (see tf.unique_with_counts).
    tokens = tf.reshape(tokens, shape=(tf.size(tokens), 1))
    tokens_st = tf.SparseTensor(
        tokens,
        tf.ones(tf.size(tokens), dtype=FEATURE_DTYPE),
        dense_shape=(len(word_vocab) + NUM_OOV_BUCKETS,))
    tokens_st = tf.sparse.reorder(tokens_st)

    return BatchType(tokens_st, tags)

  return to_ids

def build_preprocess_fn(word_vocab, tag_vocab):

  @tf.function
  def preprocess_fn(dataset):
    to_ids = build_to_ids_fn(word_vocab, tag_vocab)
    # We *don't* shuffle in order to make this colab deterministic for
    # easier testing and reproducibility.
    # But real-world training should use `.shuffle()`.
    return dataset.map(to_ids, num_parallel_calls=tf.data.experimental.AUTOTUNE)

  return preprocess_fn

微型玩具資料集

我們建構了微型玩具資料集，其中包含 12 個單字的整體詞彙表和 3 個用戶端。這個微型範例有助於測試邊緣案例 (例如，我們有兩個用戶端的不重複符記少於 MAX_TOKENS_SELECTED_PER_CLIENT = 6 個，而另一個用戶端則更多) 並開發程式碼。

然而，此方法的實際使用案例將是數千萬或更多個整體詞彙表，每個用戶端可能出現數千個不重複的符記。由於資料格式相同，因此擴展到更實際的測試平台問題應該很簡單，例如 tff.simulation.datasets.stackoverflow.load_data() 資料集。

首先，我們定義單字和標籤詞彙表。

# Features
FRUIT_WORDS = ['apple', 'orange', 'pear', 'kiwi']
VEGETABLE_WORDS = ['carrot', 'broccoli', 'arugula', 'peas']
FISH_WORDS = ['trout', 'tuna', 'cod', 'salmon']
WORD_VOCAB = FRUIT_WORDS + VEGETABLE_WORDS + FISH_WORDS

# Labels
TAG_VOCAB = ['FRUIT', 'VEGETABLE', 'FISH']

現在，我們建立 3 個具有小型本機資料集的用戶端。如果您在 Colab 中執行本教學課程，使用「在分頁中鏡射儲存格」功能釘選此儲存格及其輸出，以便解譯/檢查下方開發函式的輸出可能很有用。

preprocess_fn = build_preprocess_fn(WORD_VOCAB, TAG_VOCAB)


def make_dataset(raw):
  d = tf.data.Dataset.from_tensor_slices(
      # Matches the StackOverflow formatting
      collections.OrderedDict(
          tokens=tf.constant([t[0] for t in raw]),
          tags=tf.constant([t[1] for t in raw]),
          title=['' for _ in raw]))
  d = preprocess_fn(d)
  return d


# 4 distinct tokens
CLIENT1_DATASET = make_dataset([
    ('apple orange apple orange', 'FRUIT'),
    ('carrot trout', 'VEGETABLE|FISH'),
    ('orange apple', 'FRUIT'),
    ('orange', 'ORANGE|CITRUS')  # 2 OOV tag
])

# 6 distinct tokens
CLIENT2_DATASET = make_dataset([
    ('pear cod', 'FRUIT|FISH'),
    ('arugula peas', 'VEGETABLE'),
    ('kiwi pear', 'FRUIT'),
    ('sturgeon', 'FISH'),  # OOV word
    ('sturgeon bass', 'FISH')  # 2 OOV words
])

# A client with all possible words & tags (13 distinct tokens).
# With MAX_TOKENS_SELECTED_PER_CLIENT = 6, we won't download the model
# slices for all tokens that occur on this client.
CLIENT3_DATASET = make_dataset([
    (' '.join(WORD_VOCAB + ['oovword']), '|'.join(TAG_VOCAB)),
    # Mathe the OOV token and 'salmon' occur in the largest number
    # of examples on this client:
    ('salmon oovword', 'FISH|OOVTAG')
])

print('Word vocab')
for i, word in enumerate(WORD_VOCAB):
  print(f'{i:2d} {word}')

print('\nTag vocab')
for i, tag in enumerate(TAG_VOCAB):
  print(f'{i:2d} {tag}')

Word vocab
 0 apple
 1 orange
 2 pear
 3 kiwi
 4 carrot
 5 broccoli
 6 arugula
 7 peas
 8 trout
 9 tuna
10 cod
11 salmon

Tag vocab
 0 FRUIT
 1 VEGETABLE
 2 FISH

定義輸入特徵 (符記/單字) 和標籤 (文章標籤) 原始數量的常數。我們的實際輸入/輸出空間會多 NUM_OOV_BUCKETS = 1，因為我們新增了 OOV 符記/標籤。

NUM_WORDS = len(WORD_VOCAB) 
NUM_TAGS = len(TAG_VOCAB)

WORD_VOCAB_SIZE = NUM_WORDS + NUM_OOV_BUCKETS
TAG_VOCAB_SIZE = NUM_TAGS + NUM_OOV_BUCKETS

建立資料集的批次版本和個別批次，這在測試程式碼時會很有用。

batched_dataset1 = CLIENT1_DATASET.batch(2)
batched_dataset2 = CLIENT2_DATASET.batch(3)
batched_dataset3 = CLIENT3_DATASET.batch(2)

batch1 = next(iter(batched_dataset1))
batch2 = next(iter(batched_dataset2))
batch3 = next(iter(batched_dataset3))

定義具有稀疏輸入的模型

我們為每個標籤使用簡單的獨立邏輯迴歸模型。

def create_logistic_model(word_vocab_size: int, vocab_tags_size: int):

  model = tf.keras.models.Sequential([
      tf.keras.layers.InputLayer(input_shape=(word_vocab_size,), sparse=True),
      tf.keras.layers.Dense(
          vocab_tags_size,
          activation='sigmoid',
          kernel_initializer=tf.keras.initializers.zeros,
          # For simplicity, don't use a bias vector; this means the model
          # is a single tensor, and we only need sparse aggregation of
          # the per-token slices of the model. Generalizing to also handle
          # other model weights that are fully updated 
          # (non-dense broadcast and aggregate) would be a good exercise.
          use_bias=False),
  ])

  return model

讓我們確定它能運作，首先進行預測

model = create_logistic_model(WORD_VOCAB_SIZE, TAG_VOCAB_SIZE)
p = model.predict(batch1.tokens)
print(p)

[[0.5 0.5 0.5 0.5]
 [0.5 0.5 0.5 0.5]]

以及一些簡單的集中式訓練

model.compile(optimizer=tf.keras.optimizers.Adagrad(learning_rate=0.001),
              loss=tf.keras.losses.BinaryCrossentropy())
model.train_on_batch(batch1.tokens, batch1.tags)

聯邦式運算的建構區塊

我們將實作 聯邦式平均演算法的簡單版本，主要差異在於每個裝置僅下載模型的相關子集，且僅針對該子集貢獻更新。

我們使用 M 作為 MAX_TOKENS_SELECTED_PER_CLIENT 的簡寫。在高層次上，一輪訓練包含下列步驟

1. 每個參與的用戶端都會掃描其本機資料集，剖析輸入字串並將其對應至正確的符記 (int 索引)。這需要存取整體 (大型) 字典 (使用 特徵雜湊技術可能會避免此情況)。然後，我們會稀疏地計算每個符記出現的次數。如果裝置上出現 U 個不重複的符記，我們會選擇 num_actual_tokens = min(U, M) 個最常出現的符記進行訓練。

2. 用戶端使用 federated_select 從伺服器擷取 num_actual_tokens 個選取的符記的模型係數。每個模型切片都是形狀為 (TAG_VOCAB_SIZE, ) 的張量，因此傳輸到用戶端的資料總量最多為大小 TAG_VOCAB_SIZE * M (請參閱下方注意事項)。

3. 用戶端建構對應 global_token -> local_token，其中本機符記 (int 索引) 是選取符記清單中整體符記的索引。

4. 用戶端使用「小型」版本的整體模型，該模型僅具有最多 M 個符記的係數，範圍為 [0, num_actual_tokens)。global -> local 對應會用於從選取的模型切片初始化此模型的密集參數。

5. 用戶端使用 SGD 在使用 global -> local 對應預先處理的資料上訓練其本機模型。

6. 用戶端使用 local -> global 對應將其本機模型的參數轉換為 IndexedSlices 更新，以索引列。伺服器使用稀疏總和聚合來聚合這些更新。

7. 伺服器會取得上述聚合的 (密集) 結果，將其除以參與的用戶端數量，並將產生的平均更新套用至整體模型。

在本節中，我們建構這些步驟的建構區塊，然後將這些區塊合併到最終 federated_computation 中，以擷取一輪訓練的完整邏輯。

計算用戶端符記並決定要 federated_select 哪些模型切片

每個裝置都需要決定模型的哪些「切片」與其本機訓練資料集相關。針對我們的問題，我們透過 (稀疏地！) 計算用戶端訓練資料集中每個符記包含的範例數量來執行此操作。

@tf.function
def token_count_fn(token_counts, batch):
  """Adds counts from `batch` to the running `token_counts` sum."""
  # Sum across the batch dimension.
  flat_tokens = tf.sparse.reduce_sum(
      batch.tokens, axis=0, output_is_sparse=True)
  flat_tokens = tf.cast(flat_tokens, dtype=TOKEN_COUNT_DTYPE)
  return tf.sparse.add(token_counts, flat_tokens)

# Simple tests
# Create the initial zero token counts using empty tensors.
initial_token_counts = tf.SparseTensor(
    indices=tf.zeros(shape=(0, 1), dtype=TOKEN_DTYPE),
    values=tf.zeros(shape=(0,), dtype=TOKEN_COUNT_DTYPE),
    dense_shape=(WORD_VOCAB_SIZE,))

client_token_counts = batched_dataset1.reduce(initial_token_counts,
                                              token_count_fn)
tokens = tf.reshape(client_token_counts.indices, (-1,)).numpy()
print('tokens:', tokens)
np.testing.assert_array_equal(tokens, [0, 1, 4, 8])
# The count is the number of *examples* in which the token/word
# occurs, not the total number of occurences, since we still featurize
# multiple occurences in the same example as a "1".
counts = client_token_counts.values.numpy()
print('counts:', counts)
np.testing.assert_array_equal(counts, [2, 3, 1, 1])

tokens: [0 1 4 8]
counts: [2 3 1 1]

我們將選取模型參數，這些參數對應於裝置上最常出現的 MAX_TOKENS_SELECTED_PER_CLIENT 個符記。如果裝置上出現的符記少於此數量，我們會填補清單以啟用 federated_select 的使用。

請注意，其他策略可能更好，例如，隨機選取符記 (可能根據其出現機率)。這可確保模型的所有切片 (用戶端擁有資料的切片) 都有一些更新的機會。

@tf.function
def keys_for_client(client_dataset, max_tokens_per_client):
  """Computes a set of max_tokens_per_client keys."""
  initial_token_counts = tf.SparseTensor(
      indices=tf.zeros((0, 1), dtype=TOKEN_DTYPE),
      values=tf.zeros((0,), dtype=TOKEN_COUNT_DTYPE),
      dense_shape=(WORD_VOCAB_SIZE,))
  client_token_counts = client_dataset.reduce(initial_token_counts,
                                              token_count_fn)
  # Find the most-frequently occuring tokens
  tokens = tf.reshape(client_token_counts.indices, shape=(-1,))
  counts = client_token_counts.values
  perm = tf.argsort(counts, direction='DESCENDING')
  tokens = tf.gather(tokens, perm)
  counts = tf.gather(counts, perm)
  num_raw_tokens = tf.shape(tokens)[0]
  actual_num_tokens = tf.minimum(max_tokens_per_client, num_raw_tokens)
  selected_tokens = tokens[:actual_num_tokens]
  paddings = [[0, max_tokens_per_client - tf.shape(selected_tokens)[0]]]
  padded_tokens = tf.pad(selected_tokens, paddings=paddings)
  # Make sure the type is statically determined
  padded_tokens = tf.reshape(padded_tokens, shape=(max_tokens_per_client,))

  # We will pass these tokens as keys into `federated_select`, which
  # requires SELECT_KEY_DTYPE=np.int32 keys.
  padded_tokens = tf.cast(padded_tokens, dtype=SELECT_KEY_DTYPE)
  return padded_tokens, actual_num_tokens

# Simple test

# Case 1: actual_num_tokens > max_tokens_per_client
selected_tokens, actual_num_tokens = keys_for_client(batched_dataset1, 3)
assert tf.size(selected_tokens) == 3
assert actual_num_tokens == 3

# Case 2: actual_num_tokens < max_tokens_per_client
selected_tokens, actual_num_tokens = keys_for_client(batched_dataset1, 10)
assert tf.size(selected_tokens) == 10
assert actual_num_tokens == 4

將整體符記對應至本機符記

上述選取為我們提供了範圍 [0, actual_num_tokens) 中的密集符記集，我們將用於裝置上模型。然而，我們讀取的資料集具有來自更大的整體詞彙表範圍 [0, WORD_VOCAB_SIZE) 的符記。

因此，我們需要將整體符記對應至其對應的本機符記。本機符記 ID 僅由先前步驟中計算的 selected_tokens 張量中的索引給定。

@tf.function
def map_to_local_token_ids(client_data, client_keys):
  global_to_local = tf.lookup.StaticHashTable(
      # Note int32 -> int64 maps are not supported
      tf.lookup.KeyValueTensorInitializer(
          keys=tf.cast(client_keys, dtype=TOKEN_DTYPE),
          # Note we need to use tf.shape, not the static 
          # shape client_keys.shape[0]
          values=tf.range(0, limit=tf.shape(client_keys)[0],
                          dtype=TOKEN_DTYPE)),
      # We use -1 for tokens that were not selected, which can occur for clients
      # with more than MAX_TOKENS_SELECTED_PER_CLIENT distinct tokens.
      # We will simply remove these invalid indices from the batch below.
      default_value=-1)

  def to_local_ids(sparse_tokens):
    indices_t = tf.transpose(sparse_tokens.indices)
    batch_indices = indices_t[0]  # First column
    tokens = indices_t[1]  # Second column
    tokens = tf.map_fn(
        lambda global_token_id: global_to_local.lookup(global_token_id), tokens)
    # Remove tokens that aren't actually available (looked up as -1):
    available_tokens = tokens >= 0
    tokens = tokens[available_tokens]
    batch_indices = batch_indices[available_tokens]

    updated_indices = tf.transpose(
        tf.concat([[batch_indices], [tokens]], axis=0))
    st = tf.sparse.SparseTensor(
        updated_indices,
        tf.ones(tf.size(tokens), dtype=FEATURE_DTYPE),
        # Each client has at most MAX_TOKENS_SELECTED_PER_CLIENT distinct tokens.
        dense_shape=[sparse_tokens.dense_shape[0], MAX_TOKENS_SELECTED_PER_CLIENT])
    st = tf.sparse.reorder(st)
    return st

  return client_data.map(lambda b: BatchType(to_local_ids(b.tokens), b.tags))

# Simple test
client_keys, actual_num_tokens = keys_for_client(
    batched_dataset3, MAX_TOKENS_SELECTED_PER_CLIENT)
client_keys = client_keys[:actual_num_tokens]

d = map_to_local_token_ids(batched_dataset3, client_keys)
batch  = next(iter(d))
all_tokens = tf.gather(batch.tokens.indices, indices=1, axis=1)
# Confirm we have local indices in the range [0, MAX):
assert tf.math.reduce_max(all_tokens) < MAX_TOKENS_SELECTED_PER_CLIENT
assert tf.math.reduce_max(all_tokens) >= 0

在每個用戶端上訓練本機 (子) 模型

請注意，federated_select 會以與選取鍵相同的順序，將選取的切片傳回為 tf.data.Dataset。因此，我們首先定義公用程式函式，以取得此類 Dataset，並將其轉換為單一密集張量，該張量可用作用戶端模型的模型權重。

@tf.function
def slices_dataset_to_tensor(slices_dataset):
  """Convert a dataset of slices to a tensor."""
  # Use batching to gather all of the slices into a single tensor.
  d = slices_dataset.batch(MAX_TOKENS_SELECTED_PER_CLIENT,
                           drop_remainder=False)
  iter_d = iter(d)
  tensor = next(iter_d)
  # Make sure we have consumed everything
  opt = iter_d.get_next_as_optional()
  tf.Assert(tf.logical_not(opt.has_value()), data=[''], name='CHECK_EMPTY')
  return tensor

# Simple test
weights = np.random.random(
    size=(MAX_TOKENS_SELECTED_PER_CLIENT, TAG_VOCAB_SIZE)).astype(np.float32)
model_slices_as_dataset = tf.data.Dataset.from_tensor_slices(weights)
weights2 = slices_dataset_to_tensor(model_slices_as_dataset)
np.testing.assert_array_equal(weights, weights2)

我們現在擁有定義簡單本機訓練迴圈所需的所有元件，該迴圈將在每個用戶端上執行。

@tf.function
def client_train_fn(model, client_optimizer,
                    model_slices_as_dataset, client_data,
                    client_keys, actual_num_tokens):

  initial_model_weights = slices_dataset_to_tensor(model_slices_as_dataset)
  assert len(model.trainable_variables) == 1
  model.trainable_variables[0].assign(initial_model_weights)

  # Only keep the "real" (unpadded) keys.
  client_keys = client_keys[:actual_num_tokens]

  client_data = map_to_local_token_ids(client_data, client_keys)

  loss_fn = tf.keras.losses.BinaryCrossentropy()
  for features, labels in client_data:
    with tf.GradientTape() as tape:
      predictions = model(features)
      loss = loss_fn(labels, predictions)
    grads = tape.gradient(loss, model.trainable_variables)
    client_optimizer.apply_gradients(zip(grads, model.trainable_variables))

  model_weights_delta = model.trainable_weights[0] - initial_model_weights
  model_weights_delta = tf.slice(model_weights_delta, begin=[0, 0], 
                           size=[actual_num_tokens, -1])
  return client_keys, model_weights_delta

# Simple test
# Note if you execute this cell a second time, you need to also re-execute
# the preceeding cell to avoid "tf.function-decorated function tried to 
# create variables on non-first call" errors.
on_device_model = create_logistic_model(MAX_TOKENS_SELECTED_PER_CLIENT,
                                        TAG_VOCAB_SIZE)
client_optimizer = tf.keras.optimizers.SGD(learning_rate=0.001)
client_keys, actual_num_tokens = keys_for_client(
    batched_dataset2, MAX_TOKENS_SELECTED_PER_CLIENT)

model_slices_as_dataset = tf.data.Dataset.from_tensor_slices(
    np.zeros((MAX_TOKENS_SELECTED_PER_CLIENT, TAG_VOCAB_SIZE),
             dtype=np.float32))

keys, delta = client_train_fn(
    on_device_model,
    client_optimizer,
    model_slices_as_dataset,
    client_data=batched_dataset3,
    client_keys=client_keys,
    actual_num_tokens=actual_num_tokens)

print(delta)

聚合 IndexedSlices

我們使用 tff.federated_aggregate 為 IndexedSlices 建構聯邦式稀疏總和。此簡單實作具有 dense_shape 在先前以靜態方式知道的限制。另請注意，此總和僅為半稀疏，因為用戶端 -> 伺服器通訊是稀疏的，但伺服器在 accumulate 和 merge 中維護總和的密集表示法，並輸出此密集表示法。

def federated_indexed_slices_sum(slice_indices, slice_values, dense_shape):
  """
  Sums IndexedSlices@CLIENTS to a dense @SERVER Tensor.

  Intermediate aggregation is performed by converting to a dense representation,
  which may not be suitable for all applications.

  Args:
    slice_indices: An IndexedSlices.indices tensor @CLIENTS.
    slice_values: An IndexedSlices.values tensor @CLIENTS.
    dense_shape: A statically known dense shape.

  Returns:
    A dense tensor placed @SERVER representing the sum of the client's
    IndexedSclies.
  """
  slices_dtype = slice_values.type_signature.member.dtype
  zero = tff.tensorflow.computation(
      lambda: tf.zeros(dense_shape, dtype=slices_dtype))()

  @tf.function
  def accumulate_slices(dense, client_value):
    indices, slices = client_value
    # There is no built-in way to add `IndexedSlices`, but 
    # tf.convert_to_tensor is a quick way to convert to a dense representation
    # so we can add them.
    return dense + tf.convert_to_tensor(
        tf.IndexedSlices(slices, indices, dense_shape))


  return tff.federated_aggregate(
      (slice_indices, slice_values),
      zero=zero,
      accumulate=tff.tensorflow.computation(accumulate_slices),
      merge=tff.tensorflow.computation(lambda d1, d2: tf.add(d1, d2, name='merge')),
      report=tff.tensorflow.computation(lambda d: d))

建構最小 federated_computation 作為測試

dense_shape = (6, 2)
indices_type = tff.TensorType(np.int64, (None,))
values_type = tff.TensorType(np.float32, (None, 2))
client_slice_type = tff.FederatedType(
    (indices_type, values_type), tff.CLIENTS)

@tff.federated_computation(client_slice_type)
def test_sum_indexed_slices(indices_values_at_client):
  indices, values = indices_values_at_client
  return federated_indexed_slices_sum(indices, values, dense_shape)

print(test_sum_indexed_slices.type_signature)

({<int64[?],float32[?,2]>}@CLIENTS -> float32[6,2]@SERVER)

x = tf.IndexedSlices(
    values=np.array([[2., 2.1], [0., 0.1], [1., 1.1], [5., 5.1]],
                    dtype=np.float32),
    indices=[2, 0, 1, 5],
    dense_shape=dense_shape)
y = tf.IndexedSlices(
    values=np.array([[0., 0.3], [3.1, 3.2]], dtype=np.float32),
    indices=[1, 3],
    dense_shape=dense_shape)

# Sum one.
result = test_sum_indexed_slices([(x.indices, x.values)])
np.testing.assert_array_equal(tf.convert_to_tensor(x), result)

# Sum two.
expected = [[0., 0.1], [1., 1.4], [2., 2.1], [3.1, 3.2], [0., 0.], [5., 5.1]]
result = test_sum_indexed_slices([(x.indices, x.values), (y.indices, y.values)])
np.testing.assert_array_almost_equal(expected, result)

將所有項目放入 federated_computation 中

我們現在使用 TFF 將元件繫結在一起，形成 tff.federated_computation。

DENSE_MODEL_SHAPE = (WORD_VOCAB_SIZE, TAG_VOCAB_SIZE)
client_data_type = tff.SequenceType(batched_dataset1.element_spec)
model_type = tff.TensorType(np.float32, shape=DENSE_MODEL_SHAPE)

我們使用基於聯邦式平均的基本伺服器訓練函式，以 1.0 的伺服器學習率套用更新。重要的是，我們將更新 (delta) 套用至模型，而不是僅僅平均用戶端提供的模型，否則，如果模型的指定切片在給定回合中未由任何用戶端訓練，則其係數可能會歸零。

@tff.tensorflow.computation
def server_update(current_model_weights, update_sum, num_clients):
  average_update = update_sum / num_clients
  return current_model_weights + average_update

我們需要更多 tff.tensorflow.computation 元件

# Function to select slices from the model weights in federated_select:
select_fn = tff.tensorflow.computation(
    lambda model_weights, index: tf.gather(model_weights, index))


# We need to wrap `client_train_fn` as a `tff.tensorflow.computation`, making
# sure we do any operations that might construct `tf.Variable`s outside
# of the `tf.function` we are wrapping.
@tff.tensorflow.computation
def client_train_fn_tff(model_slices_as_dataset, client_data, client_keys,
                        actual_num_tokens):
  # Note this is amaller than the global model, using
  # MAX_TOKENS_SELECTED_PER_CLIENT which is much smaller than WORD_VOCAB_SIZE.
  # We would like a model of size `actual_num_tokens`, but we
  # can't build the model dynamically, so we will slice off the padded
  # weights at the end.
  client_model = create_logistic_model(MAX_TOKENS_SELECTED_PER_CLIENT,
                                       TAG_VOCAB_SIZE)
  client_optimizer = tf.keras.optimizers.SGD(learning_rate=0.1)
  return client_train_fn(client_model, client_optimizer,
                         model_slices_as_dataset, client_data, client_keys,
                         actual_num_tokens)

@tff.tensorflow.computation
def keys_for_client_tff(client_data):
  return keys_for_client(client_data, MAX_TOKENS_SELECTED_PER_CLIENT)

我們現在已準備好將所有部分放在一起！

@tff.federated_computation(
    tff.FederatedType(model_type, tff.SERVER), tff.FederatedType(client_data_type, tff.CLIENTS))
def sparse_model_update(server_model, client_data):
  max_tokens = tff.federated_value(MAX_TOKENS_SELECTED_PER_CLIENT, tff.SERVER)
  keys_at_clients, actual_num_tokens = tff.federated_map(
      keys_for_client_tff, client_data)

  model_slices = tff.federated_select(keys_at_clients, max_tokens, server_model,
                                      select_fn)

  update_keys, update_slices = tff.federated_map(
      client_train_fn_tff,
      (model_slices, client_data, keys_at_clients, actual_num_tokens))

  dense_update_sum = federated_indexed_slices_sum(update_keys, update_slices,
                                                  DENSE_MODEL_SHAPE)
  num_clients = tff.federated_sum(tff.federated_value(1.0, tff.CLIENTS))

  updated_server_model = tff.federated_map(
      server_update, (server_model, dense_update_sum, num_clients))

  return updated_server_model


print(sparse_model_update.type_signature)

(<server_model=float32[13,4]@SERVER,client_data={<tokens=<indices=int64[?,2],values=int32[?],dense_shape=int64[2]>,tags=float32[?,4]>*}@CLIENTS> -> float32[13,4]@SERVER)

讓我們訓練模型！

現在我們有了訓練函式，讓我們試用看看。

server_model = create_logistic_model(WORD_VOCAB_SIZE, TAG_VOCAB_SIZE)
server_model.compile(  # Compile to make evaluation easy.
    optimizer=tf.keras.optimizers.Adagrad(learning_rate=0.0),  # Unused
    loss=tf.keras.losses.BinaryCrossentropy(),
    metrics=[ 
      tf.keras.metrics.Precision(name='precision'),
      tf.keras.metrics.AUC(name='auc'),
      tf.keras.metrics.Recall(top_k=2, name='recall_at_2'),
  ])

def evaluate(model, dataset, name):
  metrics = model.evaluate(dataset, verbose=0)
  metrics_str = ', '.join([f'{k}={v:.2f}' for k, v in 
                          (zip(server_model.metrics_names, metrics))])
  print(f'{name}: {metrics_str}')

print('Before training')
evaluate(server_model, batched_dataset1, 'Client 1')
evaluate(server_model, batched_dataset2, 'Client 2')
evaluate(server_model, batched_dataset3, 'Client 3')

model_weights = server_model.trainable_weights[0]

client_datasets = [batched_dataset1, batched_dataset2, batched_dataset3]
for _ in range(10):  # Run 10 rounds of FedAvg
  # We train on 1, 2, or 3 clients per round, selecting
  # randomly.
  cohort_size = np.random.randint(1, 4)
  clients = np.random.choice([0, 1, 2], cohort_size, replace=False)
  print('Training on clients', clients)
  model_weights = sparse_model_update(
      model_weights, [client_datasets[i] for i in clients])
server_model.set_weights([model_weights])

print('After training')
evaluate(server_model, batched_dataset1, 'Client 1')
evaluate(server_model, batched_dataset2, 'Client 2')
evaluate(server_model, batched_dataset3, 'Client 3')

Before training
Client 1: loss=0.69, precision=0.00, auc=0.50, recall_at_2=0.60
Client 2: loss=0.69, precision=0.00, auc=0.50, recall_at_2=0.50
Client 3: loss=0.69, precision=0.00, auc=0.50, recall_at_2=0.40
Training on clients [0 1]
Training on clients [0 2 1]
Training on clients [2 0]
Training on clients [1 0 2]
Training on clients [2]
Training on clients [2 0]
Training on clients [1 2 0]
Training on clients [0]
Training on clients [2]
Training on clients [1 2]
After training
Client 1: loss=0.67, precision=0.80, auc=0.91, recall_at_2=0.80
Client 2: loss=0.68, precision=0.67, auc=0.96, recall_at_2=1.00
Client 3: loss=0.65, precision=1.00, auc=0.93, recall_at_2=0.80