Training on a single site's data using FLARE?

[litlep-nibbyt]

Is there a way to run training using only a single site's data using FLARE? I would like to compare single-site training performance vs. federated performance without rewriting the code I've written. I've tried to run the hello-pt example with only a single site, but I'm finding that the client goes into an infinite loop after performing local training.

What I did was follow the instructions here, except with poc -n 1.

2022-06-15 18:19:44,737 - Cifar10Trainer - INFO - [run=1, peer=example_project, peer_run=1, task_name=train, task_id=c7e2d834-4d0e-4061-b5d4-7702955360eb]: Epoch: 0/1, Iteration: 6000, Loss: 0.4929271195481221
2022-06-15 18:19:54,456 - Cifar10Trainer - INFO - [run=1, peer=example_project, peer_run=1, task_name=train, task_id=c7e2d834-4d0e-4061-b5d4-7702955360eb]: Epoch: 0/1, Iteration: 9000, Loss: 0.5006237920025984
2022-06-15 18:20:03,294 - Cifar10Trainer - INFO - [run=1, peer=example_project, peer_run=1, task_name=train, task_id=c7e2d834-4d0e-4061-b5d4-7702955360eb]: Epoch: 0/1, Iteration: 12000, Loss: 0.5049847854276498
2022-06-15 18:20:04,843 - ClientRunner - INFO - [run=1, peer=example_project, peer_run=1, task_name=train, task_id=c7e2d834-4d0e-4061-b5d4-7702955360eb]: finished processing task
2022-06-15 18:20:04,846 - FederatedClient - INFO - Starting to push execute result.
2022-06-15 18:20:04,847 - Communicator - INFO - Send submitUpdate to example_project server
2022-06-15 18:20:04,938 - Communicator - INFO - Received comments: example_project Received from site-1 (249630 Bytes, 1655331604 seconds). SubmitUpdate time: 0.09074807167053223 seconds
2022-06-15 18:20:04,939 - ClientRunner - INFO - [run=1, peer=example_project, peer_run=1, task_name=train, task_id=c7e2d834-4d0e-4061-b5d4-7702955360eb]: result sent to server for task: name=train, id=c7e2d834-4d0e-4061-b5d4-7702955360eb
2022-06-15 18:20:09,940 - ClientRunner - INFO - [run=1]: fetching task from server ...
2022-06-15 18:20:09,941 - FederatedClient - INFO - Starting to fetch execute task.
2022-06-15 18:20:09,945 - Communicator - INFO - Received from example_project server  (417 Bytes). getTask time: 0.0034558773040771484 seconds
2022-06-15 18:20:09,949 - FederatedClient - INFO - pull_task completed. Task name:__try_again__ Status:True
...

[YuanTingHsieh]

Hi @meijiu

Thanks for the interest.

Yes, it is possible to just use 1 site.

For the hello-pt example, you will need to change the "min_clients" from 2 to 1:

https://github.com/NVIDIA/NVFlare/blob/main/examples/hello-pt/config/config_fed_server.json#L50

After you tried this, you can also take a look at the CIFAR10 example, in there @holgerroth provide experiment setup to run both "central" training and "federated" settings in NVFlare:

https://github.com/NVIDIA/NVFlare/tree/main/examples/cifar10

[litlep-nibbyt]

Thanks @YuanTingHsieh!