Local model seems non-local

[choi328328]

Hi. My name is Byungjin Choi, and I am a Medical informatician in Korea.
Thank you for constructing such an excellent framework.
However, there are some things I can't understand in the hello-pt-tb example.

I think the nvflare works like below:

Scatter_and_Gather flow start:
for each round:
1) Get the global model(weight) from the server
2) Local train
3) Save local model
4) Send weight to the server
Scatter_and_Gather Done

CrosssiteEval flow start :
In each client’s validation data:
Valiate Global model
Validate each client’s model
Report results

If I understand correctly, the 'local model' is not trained only with local data.
Instead, it is a fine-tuned (or local adaptation) model of the global model(n-1 round).

Wouldn't it be more appropriate to compare the FL global model with the model trained only on local data?

[holgerroth]

You are correct in your observation. The local model is a "personalized" or "fine-tuned" version of the previous global model. Depending on the implementation of the Learner, we could also do some model selection based on the local validation scores. The CIFAR10Learner is doing that for example here.

By setting the num_rounds=1 in config_fed_server.json, you can basically simulate each client only training on their local data (that's the current default in the hello-pt-tb example. Also increase epochs in config_fed_client.json to do a complete local training.

[choi328328]

Thank you for your kind reply! My question is resolved.