.. # Copyright (C) 2020-2023 Intel Corporation .. # SPDX-License-Identifier: Apache-2.0 .. _overriding_agg_fn: ***************************** Override Aggregation Function ***************************** With the aggregator-based workflow, you can use custom aggregation functions for each task via Python\*\ API or command line interface. Python API ========== 1. Create an implementation of :class:`openfl.interface.aggregation_functions.core.AggregationFunction`. 2. In the ``override_config`` keyword argument of the :func:`openfl.native.run_experiment` native function, pass the implementation as a ``tasks.{task_name}.aggregation_type`` parameter. .. note:: See `Federated PyTorch MNIST Tutorial `_ for an example of the custom aggregation function. Command Line Interface ====================== Predefined Aggregation Functions -------------------------------- Choose from the following predefined aggregation functions: - ``openfl.interface.aggregation_functions.WeightedAverage`` (default) - ``openfl.interface.aggregation_functions.Median`` - ``openfl.interface.aggregation_functions.GeometricMedian`` - ``openfl.interface.aggregation_functions.AdagradAdaptiveAggregation`` - ``openfl.interface.aggregation_functions.AdamAdaptiveAggregation`` - ``openfl.interface.aggregation_functions.YogiAdaptiveAggregation`` .. _adaptive_aggregation_functions: Adaptive Aggregation Functions ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ .. note:: To create adaptive aggregation functions, the user must specify parameters for the aggregation optimizer (``NumPyAdagrad``, ``NumPyAdam`` or ``NumPyYogi``) that will aggregate the global model. Theese parameters parameters are passed via **keywords**. Also, user must pass one of the arguments: ``params`` - model parameters (a dictionary with named model parameters in the form of numpy arrays), or pass ``model_interface`` - an instance of the `ModelInterface `_ class. If user pass both ``params`` and ``model_interface``, then the optimizer parameters are initialized via ``params``, ignoring ``model_interface`` argument. See the `AdagradAdaptiveAggregation `_ definitions for details. `Adaptive federated optimization `_ original paper. ``AdagradAdaptiveAggregation`` usage example: .. code-block:: python from openfl.interface.interactive_api.experiment import TaskInterface, ModelInterface from openfl.interface.aggregation_functions import AdagradAdaptiveAggregation TI = TaskInterface() MI = ModelInterface(model=model, optimizer=optimizer, framework_plugin=framework_adapter) ... # Creating aggregation function agg_fn = AdagradAdaptiveAggregation(model_interface=MI, learning_rate=0.4) # Define training task @TI.register_fl_task(model='model', data_loader='train_loader', \ device='device', optimizer='optimizer') @TI.set_aggregation_function(agg_fn) def train(...): ... You can define your own numpy based optimizer, which will be used for global model aggreagation: .. code-block:: python from openfl.utilities.optimizers.numpy.base_optimizer import Optimizer class MyOpt(Optimizer): """My optimizer implementation.""" def __init__( self, *, params: Optional[Dict[str, np.ndarray]] = None, model_interface=None, learning_rate: float = 0.001, param1: Any = None, param2: Any = None ) -> None: """Initialize. Args: params: Parameters to be stored for optimization. model_interface: Model interface instance to provide parameters. learning_rate: Tuning parameter that determines the step size at each iteration. param1: My own defined parameter. param2: My own defined parameter. """ super().__init__() pass # Your code here! def step(self, gradients: Dict[str, np.ndarray]) -> None: """ Perform a single step for parameter update. Implement your own optimizer weights update rule. Args: gradients: Partial derivatives with respect to optimized parameters. """ pass # Your code here! ... from openfl.interface.aggregation_functions import WeightedAverage from openfl.interface.aggregation_functions.core import AdaptiveAggregation # Creating your implemented optimizer instance based on numpy: my_own_optimizer = MyOpt(model_interface=MI, learning_rate=0.01) # Creating aggregation function agg_fn = AdaptiveAggregation(optimizer=my_own_optimizer, agg_func=WeightedAverage()) # WeightedAverage() is used for aggregating # parameters that are not inside the given optimizer. # Define training task @TI.register_fl_task(model='model', data_loader='train_loader', \ device='device', optimizer='optimizer') @TI.set_aggregation_function(agg_fn) def train(...): ... .. note:: If you do not understand how to write your own numpy based optimizer, please see the `NumPyAdagrad `_ and `AdaptiveAggregation `_ definitions for details. Custom Aggregation Functions ---------------------------- OpenFL provides interfaces to support your own custom aggregation functions. You can also create your own implementation of :class:`openfl.interface.aggregation_functions.core.AggregationFunction`. See `example `_ for details. 1. Define the behavior of the aggregation. 2. Include the implementation in the **plan.yaml** file in the **plan** directory of your workspace. 3. In the **tasks** section, pick a task for which you want to change the aggregation and insert ``aggregation_type`` section with a single ``template`` key that defines a module path to your class. The following is an example of a **plan.yaml** with a modified aggregation function: .. code-block:: yaml # ... # other top-level sections # ... tasks: aggregated_model_validation: function: validate kwargs: apply: global metrics: - acc defaults: plan/defaults/tasks_torch.yaml locally_tuned_model_validation: function: validate kwargs: apply: local metrics: - acc settings: {} train: function: train_batches aggregation_type: template: openfl.interface.aggregation_functions.Median kwargs: metrics: - loss Interactive API ================ You can override aggregation function that will be used for the task this function corresponds to. In order to do this, call the ``set_aggregation_function`` decorator method of ``TaskInterface`` and pass ``AggregationFunction`` subclass instance as a parameter. For example, you can try: .. code-block:: python from openfl.interface.aggregation_functions import Median TI = TaskInterface() agg_fn = Median() @TI.register_fl_task(model='model', data_loader='train_loader', \ device='device', optimizer='optimizer') @TI.set_aggregation_function(agg_fn) .. warning:: All tasks with the same type of aggregation use the same class instance. If ``AggregationFunction`` implementation has its own state, then this state will be shared across tasks. ``AggregationFunction`` requires a single ``call`` function. This function receives tensors for a single parameter from multiple collaborators with additional metadata (see definition of :meth:`openfl.interface.aggregation_functions.core.AggregationFunction.call`) and returns a single tensor that represents the result of aggregation. .. note:: See the `definition `_ of :class:`openfl.interface.aggregation_functions.core.AggregationFunction.call` for details. Example of a Custom Aggregation Function ======================================== This is an example of a custom tensor clipping aggregation function that multiplies all local tensors by 0.3 and averages them according to weights equal to data parts to produce the resulting global tensor. .. code-block:: python from openfl.interface.aggregation_functions import AggregationFunction import numpy as np class ClippedAveraging(AggregationFunction): def __init__(self, ratio): self.ratio = ratio def call(self, local_tensors, db_iterator, tensor_name, fl_round, *__): """Aggregate tensors. Args: local_tensors(list[openfl.utilities.LocalTensor]): List of local tensors to aggregate. db_iterator: iterator over history of all tensors. Columns: - 'tensor_name': name of the tensor. Examples for `torch.nn.Module`s: 'conv1.weight', 'fc2.bias'. - 'round': 0-based number of round corresponding to this tensor. - 'tags': tuple of tensor tags. Tags that can appear: - 'model' indicates that the tensor is a model parameter. - 'trained' indicates that tensor is a part of a training result. These tensors are passed to the aggregator node after local learning. - 'aggregated' indicates that tensor is a result of aggregation. These tensors are sent to collaborators for the next round. - 'delta' indicates that value is a difference between rounds for a specific tensor. also one of the tags is a collaborator name if it corresponds to a result of a local task. - 'nparray': value of the tensor. tensor_name: name of the tensor fl_round: round number tags: tuple of tags for this tensor Returns: np.ndarray: aggregated tensor """ clipped_tensors = [] previous_tensor_value = None for record in db_iterator: if ( record['round'] == (fl_round - 1) and record['tensor_name'] == tensor_name and 'aggregated' in record['tags'] and 'delta' not in record['tags'] ): previous_tensor_value = record['nparray'] weights = [] for local_tensor in local_tensors: prev_tensor = previous_tensor_value if previous_tensor_value is not None else local_tensor.tensor delta = local_tensor.tensor - prev_tensor new_tensor = prev_tensor + delta * self.ratio clipped_tensors.append(new_tensor) weights.append(local_tensor.weight) return np.average(clipped_tensors, weights=weights, axis=0) A full implementation can be found at `Federated_Pytorch_MNIST_custom_aggregation_Tutorial.ipynb `_ Example of a Privileged Aggregation Function ============================================ Most of the time the AggregationFunction interface is sufficient to implement custom methods, but in certain scenarios users may want to store additional information inside the TensorDB Dataframe beyond the aggregated tensor. The :class:`openfl.interface.aggregation_functions.experimental.PrivilegedAggregationFunction` interface is provided for this use, and gives the user direct access to aggregator's TensorDB dataframe (notice the `tensor_db` param in the call function replaces the `db_iterator` from the standard AggregationFunction interface). As the name suggests, this interface is called privileged because with great power comes great responsibility, and modifying the TensorDB dataframe directly can lead to unexpected behavior and experiment failures if entries are arbitrarily deleted. .. code-block:: python from openfl.interface.aggregation_functions.experimental import PrivilegedAggregationFunction import numpy as np import pandas as pd class PrioritizeLeastImproved(PrivilegedAggregationFunction): """ Give collaborator with the least improvement in validation accuracy more influence over future weights """ def call(self, local_tensors, tensor_db, tensor_name, fl_round, tags): """Aggregate tensors. Args: local_tensors(list[openfl.utilities.LocalTensor]): List of local tensors to aggregate. tensor_db: Aggregator's TensorDB [writable]. Columns: - 'tensor_name': name of the tensor. Examples for `torch.nn.Module`s: 'conv1.weight', 'fc2.bias'. - 'round': 0-based number of round corresponding to this tensor. - 'tags': tuple of tensor tags. Tags that can appear: - 'model' indicates that the tensor is a model parameter. - 'trained' indicates that tensor is a part of a training result. These tensors are passed to the aggregator node after local learning. - 'aggregated' indicates that tensor is a result of aggregation. These tensors are sent to collaborators for the next round. - 'delta' indicates that value is a difference between rounds for a specific tensor. also one of the tags is a collaborator name if it corresponds to a result of a local task. - 'nparray': value of the tensor. tensor_name: name of the tensor fl_round: round number tags: tuple of tags for this tensor Returns: np.ndarray: aggregated tensor """ from openfl.utilities import change_tags tensors, weights, collaborators = zip(*[(x.tensor, x.weight, x.col_name) for idx,x in enumerate(local_tensors)]) tensors, weights, collaborators = np.array(tensors), np.array(weights), collaborators if fl_round > 0: metric_tags = ('metric','validate_agg') collaborator_accuracy = {} previous_col_accuracy = {} change_in_accuracy = {} for col in collaborators: col_metric_tag = change_tags(metric_tags,add_field=col) collaborator_accuracy[col] = float(tensor_db[(tensor_db['tensor_name'] == 'acc') & (tensor_db['round'] == fl_round) & (tensor_db['tags'] == col_metric_tag)]['nparray']) previous_col_accuracy[col] = float(tensor_db[(tensor_db['tensor_name'] == 'acc') & (tensor_db['round'] == fl_round - 1) & (tensor_db['tags'] == col_metric_tag)]['nparray']) change_in_accuracy[col] = collaborator_accuracy[col] - previous_col_accuracy[col] least_improved_collaborator = min(change_in_accuracy,key=change_in_accuracy.get) # Dont add least improved collaborator more than once if len(tensor_db[(tensor_db['tags'] == ('least_improved',)) & (tensor_db['round'] == fl_round)]) == 0: tensor_db.loc[tensor_db.shape[0]] = \ ['_','_',fl_round,True,('least_improved',),np.array(least_improved_collaborator)] fx.logger.info(f'Least improved collaborator = {least_improved_collaborator}') fx.logger.info(f"Least improved = {tensor_db[(tensor_db['tags'] == ('least_improved',)) & (tensor_db['nparray'] == np.array(least_improved_collaborator))]}") fx.logger.info(f'Collaborator accuracy = {collaborator_accuracy}') fx.logger.info(f'Change in accuracy {change_in_accuracy}') least_improved_weight_factor = 0.1 * len(tensor_db[(tensor_db['tags'] == ('least_improved',)) & (tensor_db['nparray'] == np.array(least_improved_collaborator))]) weights[collaborators.index(least_improved_collaborator)] += least_improved_weight_factor weights = weights / np.sum(weights) return np.average(tensors, weights=weights, axis=0) A full implementation can be found at `Federated_Pytorch_MNIST_custom_aggregation_Tutorial.ipynb `_