def main(argv): if len(argv) > 1: raise app.UsageError('Expected no command-line arguments, ' 'got: {}'.format(argv)) tff.backends.native.set_local_execution_context(max_fanout=10) model_builder = functools.partial( stackoverflow_models.create_recurrent_model, vocab_size=FLAGS.vocab_size, embedding_size=FLAGS.embedding_size, latent_size=FLAGS.latent_size, num_layers=FLAGS.num_layers, shared_embedding=FLAGS.shared_embedding) loss_builder = functools.partial( tf.keras.losses.SparseCategoricalCrossentropy, from_logits=True) special_tokens = stackoverflow_dataset.get_special_tokens(FLAGS.vocab_size) pad_token = special_tokens.pad oov_tokens = special_tokens.oov eos_token = special_tokens.eos def metrics_builder(): return [ keras_metrics.MaskedCategoricalAccuracy(name='accuracy_with_oov', masked_tokens=[pad_token]), keras_metrics.MaskedCategoricalAccuracy(name='accuracy_no_oov', masked_tokens=[pad_token] + oov_tokens), # Notice BOS never appears in ground truth. keras_metrics.MaskedCategoricalAccuracy( name='accuracy_no_oov_or_eos', masked_tokens=[pad_token, eos_token] + oov_tokens), keras_metrics.NumBatchesCounter(), keras_metrics.NumTokensCounter(masked_tokens=[pad_token]), ] datasets = stackoverflow_dataset.construct_word_level_datasets( FLAGS.vocab_size, FLAGS.client_batch_size, FLAGS.client_epochs_per_round, FLAGS.sequence_length, FLAGS.max_elements_per_user, FLAGS.num_validation_examples) train_dataset, validation_dataset, test_dataset = datasets if FLAGS.uniform_weighting: def client_weight_fn(local_outputs): del local_outputs return 1.0 else: def client_weight_fn(local_outputs): return tf.cast(tf.squeeze(local_outputs['num_tokens']), tf.float32) def model_fn(): return tff.learning.from_keras_model( model_builder(), loss_builder(), input_spec=validation_dataset.element_spec, metrics=metrics_builder()) if FLAGS.noise_multiplier is not None: if not FLAGS.uniform_weighting: raise ValueError( 'Differential privacy is only implemented for uniform weighting.' ) dp_query = tff.utils.build_dp_query( clip=FLAGS.clip, noise_multiplier=FLAGS.noise_multiplier, expected_total_weight=FLAGS.clients_per_round, adaptive_clip_learning_rate=FLAGS.adaptive_clip_learning_rate, target_unclipped_quantile=FLAGS.target_unclipped_quantile, clipped_count_budget_allocation=FLAGS. clipped_count_budget_allocation, expected_clients_per_round=FLAGS.clients_per_round) weights_type = tff.learning.framework.weights_type_from_model(model_fn) aggregation_process = tff.utils.build_dp_aggregate_process( weights_type.trainable, dp_query) else: aggregation_process = None server_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags( 'server') client_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags( 'client') iterative_process = tff.learning.build_federated_averaging_process( model_fn=model_fn, server_optimizer_fn=server_optimizer_fn, client_weight_fn=client_weight_fn, client_optimizer_fn=client_optimizer_fn, aggregation_process=aggregation_process) client_datasets_fn = training_utils.build_client_datasets_fn( train_dataset, FLAGS.clients_per_round) evaluate_fn = training_utils.build_centralized_evaluate_fn( model_builder=model_builder, eval_dataset=validation_dataset, loss_builder=loss_builder, metrics_builder=metrics_builder) test_fn = training_utils.build_centralized_evaluate_fn( model_builder=model_builder, # Use both val and test for symmetry with other experiments, which # evaluate on the entire test set. eval_dataset=validation_dataset.concatenate(test_dataset), loss_builder=loss_builder, metrics_builder=metrics_builder) logging.info('Training model:') logging.info(model_builder().summary()) hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags()) training_loop_dict = utils_impl.lookup_flag_values(training_loop_flags) training_loop.run(iterative_process=iterative_process, client_datasets_fn=client_datasets_fn, validation_fn=evaluate_fn, test_fn=test_fn, hparam_dict=hparam_dict, **training_loop_dict)
def run_experiment(): """Data preprocessing and experiment execution.""" emnist_train, _ = emnist_dataset.get_federated_datasets( train_client_batch_size=FLAGS.client_batch_size, train_client_epochs_per_round=FLAGS.client_epochs_per_round, only_digits=False) _, emnist_test = emnist_dataset.get_centralized_datasets() example_dataset = emnist_train.create_tf_dataset_for_client( emnist_train.client_ids[0]) input_spec = example_dataset.element_spec client_datasets_fn = training_utils.build_client_datasets_fn( emnist_train, FLAGS.clients_per_round) evaluate_fn = training_utils.build_centralized_evaluate_fn( eval_dataset=emnist_test, model_builder=model_builder, loss_builder=loss_builder, metrics_builder=metrics_builder) validation_fn = lambda model_weights, round_num: evaluate_fn(model_weights) client_optimizer_fn = functools.partial( utils_impl.create_optimizer_from_flags, 'client') server_optimizer_fn = functools.partial( utils_impl.create_optimizer_from_flags, 'server') def tff_model_fn(): keras_model = model_builder() return tff.learning.from_keras_model(keras_model, input_spec=input_spec, loss=loss_builder(), metrics=metrics_builder()) if FLAGS.use_compression: # We create a `MeasuredProcess` for broadcast process and a # `MeasuredProcess` for aggregate process by providing the # `_broadcast_encoder_fn` and `_mean_encoder_fn` to corresponding utilities. # The fns are called once for each of the model weights created by # tff_model_fn, and return instances of appropriate encoders. encoded_broadcast_process = ( tff.learning.framework.build_encoded_broadcast_process_from_model( tff_model_fn, _broadcast_encoder_fn)) encoded_mean_process = ( tff.learning.framework.build_encoded_mean_process_from_model( tff_model_fn, _mean_encoder_fn)) else: encoded_broadcast_process = None encoded_mean_process = None iterative_process = tff.learning.build_federated_averaging_process( model_fn=tff_model_fn, client_optimizer_fn=client_optimizer_fn, server_optimizer_fn=server_optimizer_fn, aggregation_process=encoded_mean_process, broadcast_process=encoded_broadcast_process) # Log hyperparameters to CSV hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags()) results_dir = os.path.join(FLAGS.root_output_dir, 'results', FLAGS.experiment_name) utils_impl.create_directory_if_not_exists(results_dir) hparam_file = os.path.join(results_dir, 'hparams.csv') utils_impl.atomic_write_series_to_csv(hparam_dict, hparam_file) training_loop.run(iterative_process=iterative_process, client_datasets_fn=client_datasets_fn, validation_fn=validation_fn, total_rounds=FLAGS.total_rounds, experiment_name=FLAGS.experiment_name, root_output_dir=FLAGS.root_output_dir, rounds_per_eval=FLAGS.rounds_per_eval, rounds_per_checkpoint=FLAGS.rounds_per_checkpoint, rounds_per_profile=FLAGS.rounds_per_profile)
def main(argv): if len(argv) > 1: raise app.UsageError('Expected no command-line arguments, ' 'got: {}'.format(argv)) client_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags('client') server_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags('server') client_lr_callback = callbacks.create_reduce_lr_on_plateau( learning_rate=FLAGS.client_learning_rate, decay_factor=FLAGS.client_decay_factor, min_delta=FLAGS.min_delta, min_lr=FLAGS.min_lr, window_size=FLAGS.window_size, patience=FLAGS.patience) server_lr_callback = callbacks.create_reduce_lr_on_plateau( learning_rate=FLAGS.server_learning_rate, decay_factor=FLAGS.server_decay_factor, min_delta=FLAGS.min_delta, min_lr=FLAGS.min_lr, window_size=FLAGS.window_size, patience=FLAGS.patience) def iterative_process_builder( model_fn: Callable[[], tff.learning.Model], client_weight_fn: Optional[Callable[[Any], tf.Tensor]] = None, ) -> tff.templates.IterativeProcess: """Creates an iterative process using a given TFF `model_fn`. Args: model_fn: A no-arg function returning a `tff.learning.Model`. client_weight_fn: Optional function that takes the output of `model.report_local_outputs` and returns a tensor providing the weight in the federated average of model deltas. If not provided, the default is the total number of examples processed on device. Returns: A `tff.templates.IterativeProcess`. """ return adaptive_fed_avg.build_fed_avg_process( model_fn, client_lr_callback, server_lr_callback, client_optimizer_fn=client_optimizer_fn, server_optimizer_fn=server_optimizer_fn, client_weight_fn=client_weight_fn) hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags()) shared_args = utils_impl.lookup_flag_values(shared_flags) shared_args['iterative_process_builder'] = iterative_process_builder if FLAGS.task == 'cifar100': hparam_dict['cifar100_crop_size'] = FLAGS.cifar100_crop_size federated_cifar100.run_federated( **shared_args, crop_size=FLAGS.cifar100_crop_size, hparam_dict=hparam_dict) elif FLAGS.task == 'emnist_cr': federated_emnist.run_federated( **shared_args, model=FLAGS.emnist_cr_model, hparam_dict=hparam_dict) elif FLAGS.task == 'emnist_ae': federated_emnist_ae.run_federated(**shared_args, hparam_dict=hparam_dict) elif FLAGS.task == 'shakespeare': federated_shakespeare.run_federated( **shared_args, sequence_length=FLAGS.shakespeare_sequence_length, hparam_dict=hparam_dict) elif FLAGS.task == 'stackoverflow_nwp': so_nwp_flags = collections.OrderedDict() for flag_name in task_flags: if flag_name.startswith('so_nwp_'): so_nwp_flags[flag_name[7:]] = FLAGS[flag_name].value federated_stackoverflow.run_federated( **shared_args, **so_nwp_flags, hparam_dict=hparam_dict) elif FLAGS.task == 'stackoverflow_lr': so_lr_flags = collections.OrderedDict() for flag_name in task_flags: if flag_name.startswith('so_lr_'): so_lr_flags[flag_name[6:]] = FLAGS[flag_name].value federated_stackoverflow_lr.run_federated( **shared_args, **so_lr_flags, hparam_dict=hparam_dict)
def main(argv): if len(argv) > 1: raise app.UsageError('Expected no command-line arguments, ' 'got: {}'.format(argv)) emnist_train, _ = emnist_dataset.get_federated_datasets( train_client_batch_size=FLAGS.client_batch_size, train_client_epochs_per_round=FLAGS.client_epochs_per_round, only_digits=False) _, emnist_test = emnist_dataset.get_centralized_datasets() if FLAGS.model == 'cnn': model_builder = functools.partial( emnist_models.create_conv_dropout_model, only_digits=False) elif FLAGS.model == '2nn': model_builder = functools.partial( emnist_models.create_two_hidden_layer_model, only_digits=False) else: raise ValueError('Cannot handle model flag [{!s}].'.format(FLAGS.model)) loss_builder = tf.keras.losses.SparseCategoricalCrossentropy metrics_builder = lambda: [tf.keras.metrics.SparseCategoricalAccuracy()] if FLAGS.uniform_weighting: client_weighting = tff.learning.ClientWeighting.UNIFORM else: client_weighting = tff.learning.ClientWeighting.NUM_EXAMPLES def model_fn(): return tff.learning.from_keras_model( model_builder(), loss_builder(), input_spec=emnist_test.element_spec, metrics=metrics_builder()) if FLAGS.noise_multiplier is not None: if not FLAGS.uniform_weighting: raise ValueError( 'Differential privacy is only implemented for uniform weighting.') if FLAGS.noise_multiplier <= 0: raise ValueError('noise_multiplier must be positive if DP is enabled.') if FLAGS.clip is None or FLAGS.clip <= 0: raise ValueError('clip must be positive if DP is enabled.') if not FLAGS.adaptive_clip_learning_rate: aggregation_factory = tff.aggregators.DifferentiallyPrivateFactory.gaussian_fixed( noise_multiplier=FLAGS.noise_multiplier, clients_per_round=FLAGS.clients_per_round, clip=FLAGS.clip) else: if FLAGS.adaptive_clip_learning_rate <= 0: raise ValueError('adaptive_clip_learning_rate must be positive if ' 'adaptive clipping is enabled.') aggregation_factory = tff.aggregators.DifferentiallyPrivateFactory.gaussian_adaptive( noise_multiplier=FLAGS.noise_multiplier, clients_per_round=FLAGS.clients_per_round, initial_l2_norm_clip=FLAGS.clip, target_unclipped_quantile=FLAGS.target_unclipped_quantile, learning_rate=FLAGS.adaptive_clip_learning_rate) else: if FLAGS.uniform_weighting: aggregation_factory = tff.aggregators.UnweightedMeanFactory() else: aggregation_factory = tff.aggregators.MeanFactory() server_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags('server') client_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags('client') iterative_process = tff.learning.build_federated_averaging_process( model_fn=model_fn, server_optimizer_fn=server_optimizer_fn, client_weighting=client_weighting, client_optimizer_fn=client_optimizer_fn, model_update_aggregation_factory=aggregation_factory) client_datasets_fn = training_utils.build_client_datasets_fn( emnist_train, FLAGS.clients_per_round) evaluate_fn = training_utils.build_centralized_evaluate_fn( eval_dataset=emnist_test, model_builder=model_builder, loss_builder=loss_builder, metrics_builder=metrics_builder) validation_fn = lambda model_weights, round_num: evaluate_fn(model_weights) logging.info('Training model:') logging.info(model_builder().summary()) # Log hyperparameters to CSV hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags()) results_dir = os.path.join(FLAGS.root_output_dir, 'results', FLAGS.experiment_name) utils_impl.create_directory_if_not_exists(results_dir) hparam_file = os.path.join(results_dir, 'hparams.csv') utils_impl.atomic_write_series_to_csv(hparam_dict, hparam_file) training_loop.run( iterative_process=iterative_process, client_datasets_fn=client_datasets_fn, validation_fn=validation_fn, total_rounds=FLAGS.total_rounds, experiment_name=FLAGS.experiment_name, root_output_dir=FLAGS.root_output_dir, rounds_per_eval=FLAGS.rounds_per_eval, rounds_per_checkpoint=FLAGS.rounds_per_checkpoint, rounds_per_profile=FLAGS.rounds_per_profile)
def run_experiment(): """Data preprocessing and experiment execution.""" emnist_train, emnist_test = emnist_dataset.get_emnist_datasets( FLAGS.client_batch_size, FLAGS.client_epochs_per_round, only_digits=FLAGS.only_digits) example_dataset = emnist_train.create_tf_dataset_for_client( emnist_train.client_ids[0]) input_spec = example_dataset.element_spec client_datasets_fn = training_utils.build_client_datasets_fn( emnist_train, FLAGS.clients_per_round) evaluate_fn = training_utils.build_centralized_evaluate_fn( eval_dataset=emnist_test, model_builder=model_builder, loss_builder=loss_builder, metrics_builder=metrics_builder) client_optimizer_fn = functools.partial( utils_impl.create_optimizer_from_flags, 'client') server_optimizer_fn = functools.partial( utils_impl.create_optimizer_from_flags, 'server') def tff_model_fn(): keras_model = model_builder() return tff.learning.from_keras_model(keras_model, input_spec=input_spec, loss=loss_builder(), metrics=metrics_builder()) if FLAGS.use_compression: # We create a `MeasuredProcess` for broadcast process and a # `MeasuredProcess` for aggregate process by providing the # `_broadcast_encoder_fn` and `_mean_encoder_fn` to corresponding utilities. # The fns are called once for each of the model weights created by # tff_model_fn, and return instances of appropriate encoders. encoded_broadcast_process = ( tff.learning.framework.build_encoded_broadcast_process_from_model( tff_model_fn, _broadcast_encoder_fn)) encoded_mean_process = ( tff.learning.framework.build_encoded_mean_process_from_model( tff_model_fn, _mean_encoder_fn)) else: encoded_broadcast_process = None encoded_mean_process = None iterative_process = tff.learning.build_federated_averaging_process( model_fn=tff_model_fn, client_optimizer_fn=client_optimizer_fn, server_optimizer_fn=server_optimizer_fn, aggregation_process=encoded_mean_process, broadcast_process=encoded_broadcast_process) hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags()) training_loop_dict = utils_impl.lookup_flag_values(training_loop_flags) training_loop.run(iterative_process=iterative_process, client_datasets_fn=client_datasets_fn, validation_fn=evaluate_fn, hparam_dict=hparam_dict, **training_loop_dict)
def main(argv): if len(argv) > 1: raise app.UsageError('Expected no command-line arguments, ' 'got: {}'.format(argv)) emnist_train, emnist_test = emnist_dataset.get_emnist_datasets( FLAGS.client_batch_size, FLAGS.client_epochs_per_round, only_digits=False) if FLAGS.model == 'cnn': model_builder = functools.partial( emnist_models.create_conv_dropout_model, only_digits=False) elif FLAGS.model == '2nn': model_builder = functools.partial( emnist_models.create_two_hidden_layer_model, only_digits=False) else: raise ValueError('Cannot handle model flag [{!s}].'.format( FLAGS.model)) loss_builder = tf.keras.losses.SparseCategoricalCrossentropy metrics_builder = lambda: [tf.keras.metrics.SparseCategoricalAccuracy()] if FLAGS.uniform_weighting: def client_weight_fn(local_outputs): del local_outputs return 1.0 else: client_weight_fn = None # Defaults to the number of examples per client. def model_fn(): return tff.learning.from_keras_model( model_builder(), loss_builder(), input_spec=emnist_test.element_spec, metrics=metrics_builder()) if FLAGS.noise_multiplier is not None: if not FLAGS.uniform_weighting: raise ValueError( 'Differential privacy is only implemented for uniform weighting.' ) dp_query = tff.utils.build_dp_query( clip=FLAGS.clip, noise_multiplier=FLAGS.noise_multiplier, expected_total_weight=FLAGS.clients_per_round, adaptive_clip_learning_rate=FLAGS.adaptive_clip_learning_rate, target_unclipped_quantile=FLAGS.target_unclipped_quantile, clipped_count_budget_allocation=FLAGS. clipped_count_budget_allocation, expected_clients_per_round=FLAGS.clients_per_round, per_vector_clipping=FLAGS.per_vector_clipping, model=model_fn()) weights_type = tff.learning.framework.weights_type_from_model(model_fn) aggregation_process = tff.utils.build_dp_aggregate_process( weights_type.trainable, dp_query) else: aggregation_process = None server_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags( 'server') client_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags( 'client') iterative_process = tff.learning.build_federated_averaging_process( model_fn=model_fn, server_optimizer_fn=server_optimizer_fn, client_weight_fn=client_weight_fn, client_optimizer_fn=client_optimizer_fn, aggregation_process=aggregation_process) client_datasets_fn = training_utils.build_client_datasets_fn( emnist_train, FLAGS.clients_per_round) evaluate_fn = training_utils.build_evaluate_fn( eval_dataset=emnist_test, model_builder=model_builder, loss_builder=loss_builder, metrics_builder=metrics_builder) logging.info('Training model:') logging.info(model_builder().summary()) hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags()) training_loop_dict = utils_impl.lookup_flag_values(training_loop_flags) training_loop.run(iterative_process=iterative_process, client_datasets_fn=client_datasets_fn, validation_fn=evaluate_fn, hparam_dict=hparam_dict, **training_loop_dict)
def main(argv): if len(argv) > 1: raise app.UsageError('Expected no command-line arguments, ' 'got: {}'.format(argv)) tff.backends.native.set_local_execution_context(max_fanout=10) model_builder = functools.partial( stackoverflow_models.create_recurrent_model, vocab_size=FLAGS.vocab_size, embedding_size=FLAGS.embedding_size, latent_size=FLAGS.latent_size, num_layers=FLAGS.num_layers, shared_embedding=FLAGS.shared_embedding) loss_builder = functools.partial( tf.keras.losses.SparseCategoricalCrossentropy, from_logits=True) special_tokens = stackoverflow_word_prediction.get_special_tokens( FLAGS.vocab_size) pad_token = special_tokens.pad oov_tokens = special_tokens.oov eos_token = special_tokens.eos def metrics_builder(): return [ keras_metrics.MaskedCategoricalAccuracy(name='accuracy_with_oov', masked_tokens=[pad_token]), keras_metrics.MaskedCategoricalAccuracy(name='accuracy_no_oov', masked_tokens=[pad_token] + oov_tokens), # Notice BOS never appears in ground truth. keras_metrics.MaskedCategoricalAccuracy( name='accuracy_no_oov_or_eos', masked_tokens=[pad_token, eos_token] + oov_tokens), keras_metrics.NumBatchesCounter(), keras_metrics.NumTokensCounter(masked_tokens=[pad_token]), ] train_dataset, _ = stackoverflow_word_prediction.get_federated_datasets( vocab_size=FLAGS.vocab_size, train_client_batch_size=FLAGS.client_batch_size, train_client_epochs_per_round=FLAGS.client_epochs_per_round, max_sequence_length=FLAGS.sequence_length, max_elements_per_train_client=FLAGS.max_elements_per_user) _, validation_dataset, test_dataset = stackoverflow_word_prediction.get_centralized_datasets( vocab_size=FLAGS.vocab_size, max_sequence_length=FLAGS.sequence_length, num_validation_examples=FLAGS.num_validation_examples) if FLAGS.uniform_weighting: client_weighting = tff.learning.ClientWeighting.UNIFORM else: client_weighting = tff.learning.ClientWeighting.NUM_EXAMPLES def model_fn(): return tff.learning.from_keras_model( model_builder(), loss_builder(), input_spec=validation_dataset.element_spec, metrics=metrics_builder()) if FLAGS.noise_multiplier is not None: if not FLAGS.uniform_weighting: raise ValueError( 'Differential privacy is only implemented for uniform weighting.' ) if FLAGS.noise_multiplier <= 0: raise ValueError( 'noise_multiplier must be positive if DP is enabled.') if FLAGS.clip is None or FLAGS.clip <= 0: raise ValueError('clip must be positive if DP is enabled.') if not FLAGS.adaptive_clip_learning_rate: aggregation_factory = tff.aggregators.DifferentiallyPrivateFactory.gaussian_fixed( noise_multiplier=FLAGS.noise_multiplier, clients_per_round=FLAGS.clients_per_round, clip=FLAGS.clip) else: if FLAGS.adaptive_clip_learning_rate <= 0: raise ValueError( 'adaptive_clip_learning_rate must be positive if ' 'adaptive clipping is enabled.') aggregation_factory = tff.aggregators.DifferentiallyPrivateFactory.gaussian_adaptive( noise_multiplier=FLAGS.noise_multiplier, clients_per_round=FLAGS.clients_per_round, initial_l2_norm_clip=FLAGS.clip, target_unclipped_quantile=FLAGS.target_unclipped_quantile, learning_rate=FLAGS.adaptive_clip_learning_rate) else: if FLAGS.uniform_weighting: aggregation_factory = tff.aggregators.UnweightedMeanFactory() else: aggregation_factory = tff.aggregators.MeanFactory() server_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags( 'server') client_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags( 'client') iterative_process = tff.learning.build_federated_averaging_process( model_fn=model_fn, server_optimizer_fn=server_optimizer_fn, client_weighting=client_weighting, client_optimizer_fn=client_optimizer_fn, model_update_aggregation_factory=aggregation_factory) client_datasets_fn = training_utils.build_client_datasets_fn( train_dataset, FLAGS.clients_per_round) evaluate_fn = training_utils.build_centralized_evaluate_fn( model_builder=model_builder, eval_dataset=validation_dataset, loss_builder=loss_builder, metrics_builder=metrics_builder) validation_fn = lambda state, round_num: evaluate_fn(state.model) evaluate_test_fn = training_utils.build_centralized_evaluate_fn( model_builder=model_builder, # Use both val and test for symmetry with other experiments, which # evaluate on the entire test set. eval_dataset=validation_dataset.concatenate(test_dataset), loss_builder=loss_builder, metrics_builder=metrics_builder) test_fn = lambda state: evaluate_test_fn(state.model) logging.info('Training model:') logging.info(model_builder().summary()) # Log hyperparameters to CSV hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags()) results_dir = os.path.join(FLAGS.root_output_dir, 'results', FLAGS.experiment_name) utils_impl.create_directory_if_not_exists(results_dir) hparam_file = os.path.join(results_dir, 'hparams.csv') utils_impl.atomic_write_series_to_csv(hparam_dict, hparam_file) training_loop.run(iterative_process=iterative_process, client_datasets_fn=client_datasets_fn, validation_fn=validation_fn, test_fn=test_fn, total_rounds=FLAGS.total_rounds, experiment_name=FLAGS.experiment_name, root_output_dir=FLAGS.root_output_dir, rounds_per_eval=FLAGS.rounds_per_eval, rounds_per_checkpoint=FLAGS.rounds_per_checkpoint)
def run_experiment(): """Data preprocessing and experiment execution.""" emnist_train, _ = emnist_dataset.get_federated_datasets( train_client_batch_size=FLAGS.client_batch_size, train_client_epochs_per_round=FLAGS.client_epochs_per_round, only_digits=FLAGS.only_digits) _, emnist_test = emnist_dataset.get_centralized_datasets( only_digits=FLAGS.only_digits) example_dataset = emnist_train.create_tf_dataset_for_client( emnist_train.client_ids[0]) input_spec = example_dataset.element_spec # Build optimizer functions from flags client_optimizer_fn = functools.partial( utils_impl.create_optimizer_from_flags, 'client') server_optimizer_fn = functools.partial( utils_impl.create_optimizer_from_flags, 'server') def tff_model_fn(): return tff.learning.from_keras_model( keras_model=model_builder(), input_spec=input_spec, loss=tf.keras.losses.SparseCategoricalCrossentropy(), metrics=[tf.keras.metrics.SparseCategoricalAccuracy()]) if FLAGS.use_compression: # We create a `tff.templates.MeasuredProcess` for broadcast process and a # `tff.aggregators.WeightedAggregationFactory` for aggregation by providing # the `_broadcast_encoder_fn` and `_mean_encoder_fn` to corresponding # utilities. The fns are called once for each of the model weights created # by tff_model_fn, and return instances of appropriate encoders. encoded_broadcast_process = ( tff.learning.framework.build_encoded_broadcast_process_from_model( tff_model_fn, _broadcast_encoder_fn)) aggregator = tff.aggregators.MeanFactory( tff.aggregators.EncodedSumFactory(_mean_encoder_fn)) else: encoded_broadcast_process = None aggregator = None # Construct the iterative process iterative_process = tff.learning.build_federated_averaging_process( model_fn=tff_model_fn, client_optimizer_fn=client_optimizer_fn, server_optimizer_fn=server_optimizer_fn, broadcast_process=encoded_broadcast_process, model_update_aggregation_factory=aggregator) iterative_process = ( tff.simulation.compose_dataset_computation_with_iterative_process( emnist_train.dataset_computation, iterative_process)) # Create a client sampling function, mapping integer round numbers to lists # of client ids. client_selection_fn = functools.partial( tff.simulation.build_uniform_sampling_fn(emnist_train.client_ids), size=FLAGS.clients_per_round) # Create a validation function evaluate_fn = tff.learning.build_federated_evaluation(tff_model_fn) def validation_fn(state, round_num): if round_num % FLAGS.rounds_per_eval == 0: return evaluate_fn(state.model, [emnist_test]) else: return {} # Log hyperparameters to CSV hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags()) results_dir = os.path.join(FLAGS.root_output_dir, 'results', FLAGS.experiment_name) utils_impl.create_directory_if_not_exists(results_dir) hparam_file = os.path.join(results_dir, 'hparams.csv') utils_impl.atomic_write_series_to_csv(hparam_dict, hparam_file) checkpoint_manager, metrics_managers = _configure_managers() tff.simulation.run_simulation( process=iterative_process, client_selection_fn=client_selection_fn, validation_fn=validation_fn, total_rounds=FLAGS.total_rounds, file_checkpoint_manager=checkpoint_manager, metrics_managers=metrics_managers)
def main(argv): if len(argv) > 1: raise app.UsageError('Expected no command-line arguments, ' 'got: {}'.format(argv)) client_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags( 'client') server_optimizer_fn = optimizer_utils.create_optimizer_fn_from_flags( 'server') client_lr_callback = callbacks.create_reduce_lr_on_plateau( learning_rate=FLAGS.client_learning_rate, decay_factor=FLAGS.client_decay_factor, min_delta=FLAGS.min_delta, min_lr=FLAGS.min_lr, window_size=FLAGS.window_size, patience=FLAGS.patience) server_lr_callback = callbacks.create_reduce_lr_on_plateau( learning_rate=FLAGS.server_learning_rate, decay_factor=FLAGS.server_decay_factor, min_delta=FLAGS.min_delta, min_lr=FLAGS.min_lr, window_size=FLAGS.window_size, patience=FLAGS.patience) def iterative_process_builder( model_fn: Callable[[], tff.learning.Model], client_weight_fn: Optional[Callable[[Any], tf.Tensor]] = None, ) -> tff.templates.IterativeProcess: """Creates an iterative process using a given TFF `model_fn`. Args: model_fn: A no-arg function returning a `tff.learning.Model`. client_weight_fn: Optional function that takes the output of `model.report_local_outputs` and returns a tensor providing the weight in the federated average of model deltas. If not provided, the default is the total number of examples processed on device. Returns: A `tff.templates.IterativeProcess`. """ return adaptive_fed_avg.build_fed_avg_process( model_fn, client_lr_callback, server_lr_callback, client_optimizer_fn=client_optimizer_fn, server_optimizer_fn=server_optimizer_fn, client_weight_fn=client_weight_fn) hparam_dict = utils_impl.lookup_flag_values(utils_impl.get_hparam_flags()) task_spec = training_specs.TaskSpec( iterative_process_builder=iterative_process_builder, client_epochs_per_round=FLAGS.client_epochs_per_round, client_batch_size=FLAGS.client_batch_size, clients_per_round=FLAGS.clients_per_round, sampling_random_seed=FLAGS.client_datasets_random_seed) if FLAGS.task == 'cifar100': runner_spec = federated_cifar100.configure_training( task_spec, crop_size=FLAGS.cifar100_crop_size) elif FLAGS.task == 'emnist_cr': runner_spec = federated_emnist.configure_training( task_spec, model=FLAGS.emnist_cr_model) elif FLAGS.task == 'emnist_ae': runner_spec = federated_emnist_ae.configure_training(task_spec) elif FLAGS.task == 'shakespeare': runner_spec = federated_shakespeare.configure_training( task_spec, sequence_length=FLAGS.shakespeare_sequence_length) elif FLAGS.task == 'stackoverflow_nwp': runner_spec = federated_stackoverflow.configure_training( task_spec, vocab_size=FLAGS.so_nwp_vocab_size, num_oov_buckets=FLAGS.so_nwp_num_oov_buckets, sequence_length=FLAGS.so_nwp_sequence_length, max_elements_per_user=FLAGS.so_nwp_max_elements_per_user, num_validation_examples=FLAGS.so_nwp_num_validation_examples) elif FLAGS.task == 'stackoverflow_lr': runner_spec = federated_stackoverflow_lr.configure_training( task_spec, vocab_tokens_size=FLAGS.so_lr_vocab_tokens_size, vocab_tags_size=FLAGS.so_lr_vocab_tags_size, max_elements_per_user=FLAGS.so_lr_max_elements_per_user, num_validation_examples=FLAGS.so_lr_num_validation_examples) else: raise ValueError( '--task flag {} is not supported, must be one of {}.'.format( FLAGS.task, _SUPPORTED_TASKS)) training_loop.run(iterative_process=runner_spec.iterative_process, client_datasets_fn=runner_spec.client_datasets_fn, validation_fn=runner_spec.validation_fn, test_fn=runner_spec.test_fn, total_rounds=FLAGS.total_rounds, experiment_name=FLAGS.experiment_name, root_output_dir=FLAGS.root_output_dir, rounds_per_eval=FLAGS.rounds_per_eval, rounds_per_checkpoint=FLAGS.rounds_per_checkpoint, rounds_per_profile=FLAGS.rounds_per_profile, hparam_dict=hparam_dict)