TFF simulations with accelerators

This tutorial will describe how to setup TFF simulations with accelerators. We focus on single-machine (multi-)GPU for now and will update this tutorial with multi-machine and TPU settings.

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

Before we begin

First, let us make sure the notebook is connected to a backend that has the relevant components compiled.

pip install --quiet --upgrade tensorflow-federated
pip install -U tensorboard_plugin_profile
%load_ext tensorboard
import collections
import time

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

Check if TF can detect physical GPUs and create a virtual multi-GPU environment for TFF GPU simulations. The two virtual GPUs will have limited memory to demonstrate how to configure TFF runtime.

gpu_devices = tf.config.list_physical_devices('GPU')
if not gpu_devices:
  raise ValueError('Cannot detect physical GPU device in TF')
# TODO: b/277213652 - Remove this call, as it doesn't work with C++ executor
tf.config.set_logical_device_configuration(
    gpu_devices[0], 
    [tf.config.LogicalDeviceConfiguration(memory_limit=1024),
     tf.config.LogicalDeviceConfiguration(memory_limit=1024)])
tf.config.list_logical_devices()
[LogicalDevice(name='/device:CPU:0', device_type='CPU'),
 LogicalDevice(name='/device:GPU:0', device_type='GPU'),
 LogicalDevice(name='/device:GPU:1', device_type='GPU')]

Run the following "Hello World" example to make sure the TFF environment is correctly setup. If it doesn't work, please refer to the Installation guide for instructions.

@tff.federated_computation
def hello_world():
  return 'Hello, World!'

hello_world()
b'Hello, World!'

EMNIST experimental setup

In this tutorial, we train an EMNIST image classifier with Federated Averaging algorithm. Let us start by loading the MNIST example from the TFF website.

emnist_train, emnist_test = tff.simulation.datasets.emnist.load_data(only_digits=True)

We define a function preprocessing the EMNIST example following the simple_fedavg example. Note that the argument client_epochs_per_round controls the number of local epochs on clients in federated learning.

def preprocess_emnist_dataset(client_epochs_per_round, batch_size, test_batch_size):

  def element_fn(element):
    return collections.OrderedDict(
        x=tf.expand_dims(element['pixels'], -1), y=element['label'])

  def preprocess_train_dataset(dataset):
    # Use buffer_size same as the maximum client dataset size,
    # 418 for Federated EMNIST
    return dataset.map(element_fn).shuffle(buffer_size=418).repeat(
        count=client_epochs_per_round).batch(batch_size, drop_remainder=False)

  def preprocess_test_dataset(dataset):
    return dataset.map(element_fn).batch(test_batch_size, drop_remainder=False)

  train_set = emnist_train.preprocess(preprocess_train_dataset)
  test_set = preprocess_test_dataset(
      emnist_test.create_tf_dataset_from_all_clients())
  return train_set, test_set

We use a VGG-like model, i.e., each block has two 3x3 convolutions and number of filters are doubled when the feature maps are subsampled.

def _conv_3x3(input_tensor, filters, strides):
  """2D Convolutional layer with kernel size 3x3."""

  x = tf.keras.layers.Conv2D(
      filters=filters,
      strides=strides,
      kernel_size=3,
      padding='same',
      kernel_initializer='he_normal',
      use_bias=False,
  )(input_tensor)
  return x


def _basic_block(input_tensor, filters, strides):
  """A block of two 3x3 conv layers."""

  x = input_tensor
  x = _conv_3x3(x, filters, strides)
  x = tf.keras.layers.Activation('relu')(x)

  x = _conv_3x3(x, filters, 1)
  x = tf.keras.layers.Activation('relu')(x)
  return x


def _vgg_block(input_tensor, size, filters, strides):
  """A stack of basic blocks."""
  x = _basic_block(input_tensor, filters, strides=strides)
  for _ in range(size - 1):
      x = _basic_block(x, filters, strides=1)
  return x


def create_cnn(num_blocks, conv_width_multiplier=1, num_classes=10):
  """Create a VGG-like CNN model. 

  The CNN has (6*num_blocks + 2) layers.
  """
  input_shape = (28, 28, 1)  # channels_last
  img_input = tf.keras.layers.Input(shape=input_shape)
  x = img_input
  x = tf.image.per_image_standardization(x)

  x = _conv_3x3(x, 16 * conv_width_multiplier, 1)
  x = _vgg_block(x, size=num_blocks, filters=16 * conv_width_multiplier, strides=1)
  x = _vgg_block(x, size=num_blocks, filters=32 * conv_width_multiplier, strides=2)
  x = _vgg_block(x, size=num_blocks, filters=64 * conv_width_multiplier, strides=2)

  x = tf.keras.layers.GlobalAveragePooling2D()(x)
  x = tf.keras.layers.Dense(num_classes)(x)

  model = tf.keras.models.Model(
      img_input,
      x,
      name='cnn-{}-{}'.format(6 * num_blocks + 2, conv_width_multiplier))
  return model

Now let us define the training loop for EMNIST. Note that use_experimental_simulation_loop=True in tff.learning.algorithms.build_weighted_fed_avg is suggested for performant TFF simulation, and required to take advantage of multi-GPUs on a single machine. See simple_fedavg example for how to define customized federated learning algorithm that has high performance on GPUs, one of the key features is to explicitly use for ... iter(dataset) for training loops.

def keras_evaluate(model, test_data, metric):
  metric.reset_states()
  for batch in test_data:
    preds = model(batch['x'], training=False)
    metric.update_state(y_true=batch['y'], y_pred=preds)
  return metric.result()


def run_federated_training(client_epochs_per_round, 
                           train_batch_size, 
                           test_batch_size, 
                           cnn_num_blocks, 
                           conv_width_multiplier,
                           server_learning_rate, 
                           client_learning_rate, 
                           total_rounds, 
                           clients_per_round, 
                           rounds_per_eval,
                           logdir='logdir'):

  train_data, test_data = preprocess_emnist_dataset(
      client_epochs_per_round, train_batch_size, test_batch_size)
  data_spec = test_data.element_spec

  def _model_fn():
    keras_model = create_cnn(cnn_num_blocks, conv_width_multiplier)
    loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
    return tff.learning.models.from_keras_model(
        keras_model, input_spec=data_spec, loss=loss)

  def _server_optimizer_fn():
    return tf.keras.optimizers.SGD(learning_rate=server_learning_rate)

  def _client_optimizer_fn():
    return tf.keras.optimizers.SGD(learning_rate=client_learning_rate)

  learning_process = tff.learning.algorithms.build_weighted_fed_avg(
      model_fn=_model_fn, 
      server_optimizer_fn=_server_optimizer_fn, 
      client_optimizer_fn=_client_optimizer_fn, 
      use_experimental_simulation_loop=True)

  metric = tf.keras.metrics.SparseCategoricalAccuracy(name='test_accuracy')
  eval_model = create_cnn(cnn_num_blocks, conv_width_multiplier)
  logging.info(eval_model.summary())

  server_state = learning_process.initialize()
  start_time = time.time()
  for round_num in range(total_rounds):
    sampled_clients = np.random.choice(
        train_data.client_ids,
        size=clients_per_round,
        replace=False)
    sampled_train_data = [
        train_data.create_tf_dataset_for_client(client)
        for client in sampled_clients
    ]
    if round_num == total_rounds-1:
      with tf.profiler.experimental.Profile(logdir):
        result = learning_process.next(
            server_state, sampled_train_data)
    else:
      result = learning_process.next(
            server_state, sampled_train_data)
    server_state = result.state
    train_metrics = result.metrics['client_work']['train']
    print(f'Round {round_num} training loss: {train_metrics["loss"]}, '
     f'time: {(time.time()-start_time)/(round_num+1.)} secs')
    if round_num % rounds_per_eval == 0 or round_num == total_rounds-1:
      model_weights = learning_process.get_model_weights(server_state)
      model_weights.assign_weights_to(eval_model)
      accuracy = keras_evaluate(eval_model, test_data, metric)
      print(f'Round {round_num} validation accuracy: {accuracy * 100.0}')

Single GPU execution

The default runtime of TFF is the same as TF: when GPUs are provided, the first GPU will be chosen for execution. We run the previously defined federated training for several rounds with a relatively small model. The last round of execution is profiled with tf.profiler and visualized by tensorboard. The profiling verified the first GPU is used.

run_federated_training(
    client_epochs_per_round=1, 
    train_batch_size=16, 
    test_batch_size=128, 
    cnn_num_blocks=2, 
    conv_width_multiplier=4,
    server_learning_rate=1.0, 
    client_learning_rate=0.01,
    total_rounds=10,
    clients_per_round=16, 
    rounds_per_eval=2,
    )
Model: "cnn-14-4"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 28, 28, 1)]       0         
_________________________________________________________________
tf.image.per_image_standardi (None, 28, 28, 1)         0         
_________________________________________________________________
conv2d (Conv2D)              (None, 28, 28, 64)        576       
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation (Activation)      (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_1 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_2 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_3 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 14, 14, 128)       73728     
_________________________________________________________________
activation_4 (Activation)    (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_5 (Activation)    (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_6 (Activation)    (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_8 (Conv2D)            (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_7 (Activation)    (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_9 (Conv2D)            (None, 7, 7, 256)         294912    
_________________________________________________________________
activation_8 (Activation)    (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_10 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_9 (Activation)    (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_11 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_10 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_12 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_11 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
global_average_pooling2d (Gl (None, 256)               0         
_________________________________________________________________
dense (Dense)                (None, 10)                2570      
=================================================================
Total params: 2,731,082
Trainable params: 2,731,082
Non-trainable params: 0
_________________________________________________________________
Round 0 training loss: 2.4688243865966797, time: 13.382015466690063 secs
Round 0 validation accuracy: 15.240497589111328
Round 1 training loss: 2.3217368125915527, time: 9.311999917030334 secs
Round 2 training loss: 2.3100595474243164, time: 6.972411632537842 secs
Round 2 validation accuracy: 11.226489067077637
Round 3 training loss: 2.303222417831421, time: 6.467299699783325 secs
Round 4 training loss: 2.2976326942443848, time: 5.526083135604859 secs
Round 4 validation accuracy: 11.224040031433105
Round 5 training loss: 2.2919719219207764, time: 5.468692660331726 secs
Round 6 training loss: 2.2911534309387207, time: 4.935825347900391 secs
Round 6 validation accuracy: 11.833855628967285
Round 7 training loss: 2.2871201038360596, time: 4.918408691883087 secs
Round 8 training loss: 2.2818832397460938, time: 4.602836343977186 secs
Round 8 validation accuracy: 11.385677337646484
Round 9 training loss: 2.2790346145629883, time: 4.99558527469635 secs
Round 9 validation accuracy: 11.226489067077637
%tensorboard --logdir=logdir --port=0

Larger model and OOM

Let us run a larger model on CPU with less federated rounds.

run_federated_training(
    client_epochs_per_round=1, 
    train_batch_size=16, 
    test_batch_size=128, 
    cnn_num_blocks=4, 
    conv_width_multiplier=4,
    server_learning_rate=1.0, 
    client_learning_rate=0.01,
    total_rounds=5,
    clients_per_round=16, 
    rounds_per_eval=2,
    )
Model: "cnn-26-4"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_4 (InputLayer)         [(None, 28, 28, 1)]       0         
_________________________________________________________________
tf.image.per_image_standardi (None, 28, 28, 1)         0         
_________________________________________________________________
conv2d_39 (Conv2D)           (None, 28, 28, 64)        576       
_________________________________________________________________
conv2d_40 (Conv2D)           (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_36 (Activation)   (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_41 (Conv2D)           (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_37 (Activation)   (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_42 (Conv2D)           (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_38 (Activation)   (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_43 (Conv2D)           (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_39 (Activation)   (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_44 (Conv2D)           (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_40 (Activation)   (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_45 (Conv2D)           (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_41 (Activation)   (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_46 (Conv2D)           (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_42 (Activation)   (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_47 (Conv2D)           (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_43 (Activation)   (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_48 (Conv2D)           (None, 14, 14, 128)       73728     
_________________________________________________________________
activation_44 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_49 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_45 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_50 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_46 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_51 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_47 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_52 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_48 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_53 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_49 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_54 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_50 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_55 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_51 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_56 (Conv2D)           (None, 7, 7, 256)         294912    
_________________________________________________________________
activation_52 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_57 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_53 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_58 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_54 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_59 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_55 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_60 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_56 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_61 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_57 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_62 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_58 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_63 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_59 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
global_average_pooling2d_3 ( (None, 256)               0         
_________________________________________________________________
dense_3 (Dense)              (None, 10)                2570      
=================================================================
Total params: 5,827,658
Trainable params: 5,827,658
Non-trainable params: 0
_________________________________________________________________
Round 0 training loss: 2.437223434448242, time: 24.121686458587646 secs
Round 0 validation accuracy: 9.024785041809082
Round 1 training loss: 2.3081459999084473, time: 19.48685622215271 secs
Round 2 training loss: 2.305305242538452, time: 15.73950457572937 secs
Round 2 validation accuracy: 9.791339874267578
Round 3 training loss: 2.303149700164795, time: 15.194068729877472 secs
Round 4 training loss: 2.3026506900787354, time: 14.036769819259643 secs
Round 4 validation accuracy: 12.193867683410645

This model might hit an out of memory issue on a single GPU. The migration from large scale CPU experiments to GPU simulation can be constrained by memory usage as GPUs often have limited memeories. There are several parameters can be tuned in TFF runtime to mitigate OOM issue

# Control concurrency by `max_concurrent_computation_calls`.
tff.backends.native.set_sync_local_cpp_execution_context(
    max_concurrent_computation_calls=16/2)

run_federated_training(
    client_epochs_per_round=1, 
    train_batch_size=16, 
    test_batch_size=128, 
    cnn_num_blocks=4, 
    conv_width_multiplier=4,
    server_learning_rate=1.0, 
    client_learning_rate=0.01,
    total_rounds=5,
    clients_per_round=16, 
    rounds_per_eval=2,
    )
Model: "cnn-26-4"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 28, 28, 1)]       0         
_________________________________________________________________
tf.image.per_image_standardi (None, 28, 28, 1)         0         
_________________________________________________________________
conv2d (Conv2D)              (None, 28, 28, 64)        576       
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation (Activation)      (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_1 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_2 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_3 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_4 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_5 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_6 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_8 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_7 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_9 (Conv2D)            (None, 14, 14, 128)       73728     
_________________________________________________________________
activation_8 (Activation)    (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_10 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_9 (Activation)    (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_11 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_10 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_12 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_11 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_13 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_12 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_14 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_13 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_15 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_14 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_16 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_15 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_17 (Conv2D)           (None, 7, 7, 256)         294912    
_________________________________________________________________
activation_16 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_18 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_17 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_19 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_18 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_20 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_19 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_21 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_20 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_22 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_21 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_23 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_22 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_24 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_23 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
global_average_pooling2d (Gl (None, 256)               0         
_________________________________________________________________
dense (Dense)                (None, 10)                2570      
=================================================================
Total params: 5,827,658
Trainable params: 5,827,658
Non-trainable params: 0
_________________________________________________________________
Round 0 training loss: 2.4990053176879883, time: 11.922378778457642 secs
Round 0 validation accuracy: 11.224040031433105
Round 1 training loss: 2.307560920715332, time: 9.916815996170044 secs
Round 2 training loss: 2.3032877445220947, time: 7.68927804629008 secs
Round 2 validation accuracy: 11.224040031433105
Round 3 training loss: 2.302366256713867, time: 7.681552231311798 secs
Round 4 training loss: 2.301671028137207, time: 7.613566827774048 secs
Round 4 validation accuracy: 11.224040031433105

Optimize performance

Techniques in TF that could achieve better performance can generally be used in TFF, e.g., mixed precision training and XLA. The speedup (on GPUs like V100) and memory saving of mixed precision can often be significant, which could be examined by tf.profiler.

# Mixed precision training. 
tff.backends.native.set_sync_local_cpp_execution_context()
policy = tf.keras.mixed_precision.experimental.Policy('mixed_float16')
tf.keras.mixed_precision.experimental.set_policy(policy)

run_federated_training(
    client_epochs_per_round=1, 
    train_batch_size=16, 
    test_batch_size=128, 
    cnn_num_blocks=4, 
    conv_width_multiplier=4,
    server_learning_rate=1.0, 
    client_learning_rate=0.01,
    total_rounds=5,
    clients_per_round=16, 
    rounds_per_eval=2,
    logdir='mixed'
    )
Model: "cnn-26-4"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 28, 28, 1)]       0         
_________________________________________________________________
tf.image.per_image_standardi (None, 28, 28, 1)         0         
_________________________________________________________________
conv2d (Conv2D)              (None, 28, 28, 64)        576       
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation (Activation)      (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_1 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_2 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_3 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_4 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_5 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_6 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_8 (Conv2D)            (None, 28, 28, 64)        36864     
_________________________________________________________________
activation_7 (Activation)    (None, 28, 28, 64)        0         
_________________________________________________________________
conv2d_9 (Conv2D)            (None, 14, 14, 128)       73728     
_________________________________________________________________
activation_8 (Activation)    (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_10 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_9 (Activation)    (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_11 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_10 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_12 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_11 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_13 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_12 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_14 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_13 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_15 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_14 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_16 (Conv2D)           (None, 14, 14, 128)       147456    
_________________________________________________________________
activation_15 (Activation)   (None, 14, 14, 128)       0         
_________________________________________________________________
conv2d_17 (Conv2D)           (None, 7, 7, 256)         294912    
_________________________________________________________________
activation_16 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_18 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_17 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_19 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_18 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_20 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_19 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_21 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_20 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_22 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_21 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_23 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_22 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
conv2d_24 (Conv2D)           (None, 7, 7, 256)         589824    
_________________________________________________________________
activation_23 (Activation)   (None, 7, 7, 256)         0         
_________________________________________________________________
global_average_pooling2d (Gl (None, 256)               0         
_________________________________________________________________
dense (Dense)                (None, 10)                2570      
=================================================================
Total params: 5,827,658
Trainable params: 5,827,658
Non-trainable params: 0
_________________________________________________________________
Round 0 training loss: 2.4187185764312744, time: 18.763780117034912 secs
Round 0 validation accuracy: 9.977468490600586
Round 1 training loss: 2.305102825164795, time: 13.712820529937744 secs
Round 2 training loss: 2.304737091064453, time: 9.993690172831217 secs
Round 2 validation accuracy: 11.779976844787598
Round 3 training loss: 2.2996833324432373, time: 9.29404467344284 secs
Round 4 training loss: 2.299349308013916, time: 9.195427560806275 secs
Round 4 validation accuracy: 11.224040031433105
%tensorboard --logdir=mixed --port=0