nvflare.app_opt.xgboost.histogram_based_v2.csv_data_loader module

class CSVDataLoader(folder: str)[source]

Bases: XGBDataLoader

Reads CSV dataset and returns XGB data matrix with automatic client-specific loading.

This data loader automatically handles site-specific data loading. Even though you pass the same folder path to all clients, each client will load its own data based on its client_id (which is injected by the framework at runtime).

Expected folder structure:

{folder}/ ├── site-1/ │ ├── train.csv │ └── valid.csv ├── site-2/ │ ├── train.csv │ └── valid.csv └── site-3/

├── train.csv └── valid.csv

For horizontal mode (row split):

  • Each site’s CSV contains all features + labels

  • Each site has different rows (samples)

For vertical mode (column split):

  • site-1 (rank 0) contains subset of features + labels

  • Other sites contain different features, no labels

  • All sites have the same rows (samples)

Parameters:

folder – Base folder path containing client-specific subdirectories. Each client will automatically load from {folder}/{client_id}/

Example

# In your job script - same data loader for all clients
for i in range(1, 4):
    dataloader = CSVDataLoader(folder="/tmp/data/horizontal")
    recipe.add_to_client(f"site-{i}", dataloader)

# At runtime:
# site-1 loads: /tmp/data/horizontal/site-1/train.csv
# site-2 loads: /tmp/data/horizontal/site-2/train.csv
# site-3 loads: /tmp/data/horizontal/site-3/train.csv

Note

In vertical mode, the label owner is always rank 0 (typically site-1).

load_data()[source]

Loads data for xgboost.

Returns:

A tuple of train_data, validation_data