FedAvg Early Stopping with ModelController (#2648)

* add early stopping with model controller

* add save_flmodel

* save model improvements

* add script and example

* add notebook

examples/hello-world/hello-fedavg/README.md

@@ -0,0 +1,28 @@
+# Hello FedAvg
+
+In this example we highlight the flexibility of the ModelController API, and show how to write a Federated Averaging workflow with early stopping, model selection, and saving and loading. Follow along in the [hello-fedavg.ipynb](hello-fedavg.ipynb) notebook for more details.
+
+### 1. Setup
+
+```
+pip install nvflare~=2.5.0rc torch torchvision tensorboard
+```
+
+### 2. PTFedAvgEarlyStopping using ModelController API
+
+The ModelController API enables the option to easily customize a workflow with Python code.
+
+- FedAvg: We subclass the BaseFedAvg class to leverage the predefined aggregation functions.
+- Early Stopping: We add a `stop_condition` argument (eg. `"accuracy >= 80"`) and end the workflow early if the corresponding global model metric meets the condition.
+- Model Selection: As and alternative to using a `IntimeModelSelector` componenet for model selection, we instead compare the metrics of the models in the workflow to select the best model each round.
+- Saving/Loading: Rather than configuring a persistor such as `PTFileModelPersistor` component, we choose to utilize PyTorch's save and load functions and save the metadata of the FLModel separately.
+
+### 3. Run the script
+
+Use the Job API to define and run the example with the simulator:
+
+```
+python3 pt_fedavg_early_stopping_script.py
+```
+
+View the results in the job workspace: `/tmp/nvflare/jobs/workdir`.