Federated GBDT Example

This example shows how to use fedlearner to train and test two-party federated GBDT models on the MNIST dataset.

Data

To run the the example, first prepare the data with:

python make_data.py --verify-example-ids=True

This will generate files in data.

The data files are CSV files with out header. Each row contains float features (and labels) for a data example. The first N columns are features for both leader and follower. In train and test mode, the last column of leader's data file contain labels (0 or 1). In eval mode there are no labels in the data.

Training

You need to start two processes, one for leader and one for follower, for each training task.

Run the following command to start leader:

python -m fedlearner.model.tree.trainer leader \
    --local-addr=localhost:50051 \
    --peer-addr=localhost:50052 \
    --data-path=data/leader_train.csv \
    --export-model-path=leader_model.proto \
    --verify-example-ids=true

Run the following command in another terminal session to start follower:

python -m fedlearner.model.tree.trainer follower \
    --local-addr=localhost:50052 \
    --peer-addr=localhost:50051 \
    --data-path=data/follower_train.csv \
    --export-model-path=follower_model.proto \
    --verify-example-ids=true

There are many modeling parameters you can change with additional commandline arguments. Please refer to help for more info:

python -m fedlearner.model.tree.trainer

Evaluation

Now you can load the previously trained model test its accuracy on test data. This is similar to training except you need to set mode to test:

Run the following command to start leader:

python -m fedlearner.model.tree.trainer leader \
    --local-addr=localhost:50051 \
    --peer-addr=localhost:50052 \
    --mode=test \
    --data-path=data/leader_test.csv \
    --load-model-path=leader_model.proto \
    --verify-example-ids=true

Run the following command in another terminal session to start follower:

python -m fedlearner.model.tree.trainer follower \
    --local-addr=localhost:50052 \
    --peer-addr=localhost:50051 \
    --mode=test \
    --data-path=data/follower_test.csv \
    --load-model-path=follower_model.proto \
    --verify-example-ids=true

Performance