tff.learning.algorithms.build_personalization_eval_computation

Builds the TFF computation for evaluating personalization strategies.

tff.learning.algorithms.build_personalization_eval_computation(
    model_fn: Callable[[], tff.learning.models.VariableModel],
    personalize_fn_dict: Mapping[str, Callable[[], tff.learning.models.VariableModel]],
    baseline_evaluate_fn: Callable[[variable.VariableModel, tf.data.Dataset], _MetricsType],
    max_num_clients: int = 100
) -> tff.Computation

The returned TFF computation broadcasts model weights from tff.SERVER to tff.CLIENTS. Each client evaluates the personalization strategies given in personalize_fn_dict. Evaluation metrics from at most max_num_clients participating clients are collected to the server.

Note: The functions in personalize_fn_dict and baseline_evaluate_fn are expected to take as input unbatched datasets, and are responsible for applying batching, if any, to the provided input datasets.

Args
model_fn	A no-arg function that returns a tff.learning.models.VariableModel. This method must not capture TensorFlow tensors or variables and use them. The model must be constructed entirely from scratch on each invocation, returning the same pre-constructed model each call will result in an error.
personalize_fn_dict	An OrderedDict that maps a string (representing a strategy name) to a no-argument function that returns a tf.function. Each tf.function represents a personalization strategy - it accepts a tff.learning.models.VariableModel (with weights already initialized to the given model weights when users invoke the returned TFF computation), an unbatched tf.data.Dataset for train, and an unbatched tf.data.Dataset for test, trains a personalized model, and returns the evaluation metrics. The evaluation metrics are represented as an OrderedDict (or a nested OrderedDict) of string metric names to scalar tf.Tensors.
baseline_evaluate_fn	A tf.function that accepts a tff.learning.models.VariableModel (with weights already initialized to the provided model weights when users invoke the returned TFF computation), and an unbatched tf.data.Dataset, evaluates the model on the dataset, and returns the evaluation metrics. The evaluation metrics are represented as an OrderedDict (or a nested OrderedDict) of string metric names to scalar tf.Tensors. This function is only used to compute the baseline metrics of the initial model.
max_num_clients	A positive int specifying the maximum number of clients to collect metrics in a round (default is 100). The clients are sampled without replacement. For each sampled client, all the personalization metrics from this client will be collected. If the number of participating clients in a round is smaller than this value, then metrics from all clients will be collected.

Returns

A federated tff.Computation with the functional type signature (<model_weights@SERVER, input@CLIENTS> -> personalization_metrics@SERVER): model_weights is a tff.learning.models.ModelWeights. Each client's input is an OrderedDict of two required keys train_data and test_data; each key is mapped to an unbatched tf.data.Dataset. personalization_metrics is an OrderedDict that maps a key 'baseline_metrics' to the evaluation metrics of the initial model (computed by baseline_evaluate_fn), and maps keys (strategy names) in personalize_fn_dict to the evaluation metrics of the corresponding personalization strategies. Note: only metrics from at most max_num_clients participating clients (sampled without replacement) are collected to the SERVER. All collected metrics are stored in a single OrderedDict (personalization_metrics shown above), where each metric is mapped to a list of scalars (each scalar comes from one client). Metric values at the same position, e.g., metric_1[i], metric_2[i]..., all come from the same client.

Raises
TypeError	If arguments are of the wrong types.
ValueError	If baseline_metrics is used as a key in personalize_fn_dict.
ValueError	If max_num_clients is not positive.