[Q&A] Using pre-trained initial model instead of training from scratch

[umgpy]

Python version (python3 -V)

3.9.12

NVFlare version (python3 -m pip list | grep "nvflare")

2.3.3

NVFlare branch (if running examples, please use the branch that corresponds to the NVFlare version, git branch)

No response

Operating system

Ubuntu 20.04

Have you successfully run any of the following examples?

• hello-numpy-sag with simulator
• hello-pt with simulator
• hello-numpy-sag with POC
• hello-pt with POC

Please describe your question

I am following the nvflare-monai Prostate 3D example and I am trying to have an initial per-trained model, instead of training from scratch. Does one have to define that in the SupervisedLearner class?

[YuanTingHsieh]

@holgerroth or @ZiyueXu77 can you help on this when you got time, thanks!

[holgerroth]

For FedAvg, you can use the server to initialize the state of the global model before training. You can add an argument to the PTFileModelPersistor, namely source_ckpt_file_full_name which will load your pretrained checkpoint before sending the weights out to the clients for local training.