FedProx Update Branch (#175)

Co-authored-by: Ethan Hugh <ethanhugh@Ethans-MacBook-Air.local>

Author: EthanHugh

docs/Examples.md

@@ -16,5 +16,5 @@ The git repository includes a number of examples that showcased how third-party
 |[FedAtt](https://arxiv.org/abs/1812.07108) | An implementation of the FedAtt algorithm — Ji et al., "[Learning Private Neural Language Modeling with Attentive Aggregation](https://arxiv.org/abs/1812.07108)," in the Proceedings of the 2019 International Joint Conference on Neural Networks (IJCNN 2019). To run this example, change directory to `examples/fedatt`, and run `python fedatt.py -c <configuration file>`. | Yes |
 |[FedNova](https://proceedings.neurips.cc/paper/2020/hash/564127c03caab942e503ee6f810f54fd-Abstract.html) | Change directory to `examples/fednova` and run `python fednova.py -c <configuration file>`. | Yes |
 |[FedSarah](https://arxiv.org/pdf/1703.00102.pdf)                             | Change directory to `examples/fedsarah` and run `python fedsarah.py -c <configuration file>`. | Yes |
-|[FedProx](https://arxiv.org/pdf/1812.06127.pdf)                              | Set `optimizer` of `trainer` to `FedProx` in your configuration file. | Yes |
+|[FedProx](https://arxiv.org/pdf/1812.06127.pdf)                              | Change directory to `examples/fedprox` and run `python fedprox.py -c <configuration file>`. | Yes |
 |[SCAFFOLD](https://arxiv.org/pdf/1910.06378.pdf)                             | Change directory to `examples/scaffold` and run `python scaffold.py -c <configuration file>`. | Not yet |

examples/fedprox/fedprox_trainer.py

@@ -25,6 +25,7 @@ def flatten_weights_from_model(model):
 
 class FedProxLocalObjective:
     """ Representing the local objective of FedProx clients. """
+
     def __init__(self, model):
         self.model = model
         self.init_global_weights = flatten_weights_from_model(model)
@@ -44,29 +45,6 @@ def compute_objective(self, outputs, labels):
 
 class Trainer(basic.Trainer):
     """ The federated learning trainer for the FedProx client. """
-    def train_process(self, config, trainset, sampler, cut_layer=None):
-        """The main training loop in FedProx framework. """
-
-        # For FedProx, the server will accept partial solutions from straggling clients
-        # after waiting them for a certain amount of time. To re-create this scenario in
-        # an experiment, a proportion of the selected clients will train for a smaller
-        # number of epochs to simulate the stragglers that return with partial solutions,
-        # as mentioned in Section 5.2
-        if hasattr(Config().clients, 'straggler_simulation') and Config(
-        ).clients.straggler_simulation:
-            np.random.seed(self.client_id)
-            # Determine whether this selected client is a straggler
-            strag_prop = Config().clients.straggler_percentage / 100
-            is_straggler = np.random.choice([True, False],
-                                            p=[strag_prop, 1 - strag_prop])
-            if is_straggler:
-                # Choose the epoch uniformly as mentioned in Section 5.2 of the paper
-                global_epochs = Config().trainer.epochs
-                config['epochs'] = np.random.choice(np.arange(
-                    1, global_epochs))
-
-        super(Trainer, self).train_process(config, trainset, sampler,
-                                           cut_layer)
 
     def loss_criterion(self, model):
         """ Return the loss criterion for FedProx clients. """