Skip to content

Multi-dataset training

Scenario: You want to train on RELLIS-3D, GOOSE-3D, and SemanticKITTI simultaneously. Each dataset has its own sensor, so intensity statistics differ. RELLIS is also 3× smaller than the others.

APIs: concat(), repeat(), per-channel transform(), sample-level transform().

Dataset-specific normalisation before concat

Apply normalisation per source before merging so that each sensor's intensity distribution is centred the same way. ConcatDataset takes the key intersection — rename GOOSE's trav_label to trav_gt first so all three sources expose the same schema.

from pathlib import Path
from torch.utils.data import DataLoader
from apairo import Rellis3DDataset, Goose3DDataset, SemanticKittiDataset

def normalize_intensity(mean, std):
    def _fn(pts):
        pts = pts.copy()
        pts[:, 3] = (pts[:, 3] - mean) / (std + 1e-6)
        return pts
    return _fn

def rename_key(src, dst):
    def _fn(sample):
        sample.data[dst] = sample.data.pop(src)
        return sample
    return _fn

ds_rellis = (
    Rellis3DDataset("/data/RELLIS", keys=["lidar", "trav_gt"])
    .split("train")
    .transform("lidar", normalize_intensity(mean=0.28, std=0.14))
)

ds_goose = (
    Goose3DDataset("/data/GOOSE_3D/train", keys=["lidar", "trav_label"])
    .transform("lidar", normalize_intensity(mean=0.52, std=0.21))
    .transform(rename_key("trav_label", "trav_gt"))
)

ds_kitti = (
    SemanticKittiDataset("/data/semantic_kitti", keys=["lidar", "trav_gt"], split="train")
    .transform("lidar", normalize_intensity(mean=0.41, std=0.18))
)

Upsampling with repeat()

RELLIS has ~3× fewer frames than the other two. repeat(3) makes it contribute proportionally. With stochastic augmentation, each repetition produces different samples.

ds_train = ds_rellis.repeat(3).concat(ds_goose, ds_kitti)

print(f"Total frames : {len(ds_train)}")
print(f"Keys         : {ds_train.keys}")  # ["lidar", "trav_gt"]

concat() takes the key intersection — all three sources expose ["lidar", "trav_gt"] after renaming.

DataLoader

loader = DataLoader(ds_train, batch_size=8, shuffle=True, num_workers=4)

ds_train is a standard PyTorch dataset; no adapter needed.

Prerequisite

Run the per-dataset traversability preprocessors first so that trav_gt (RELLIS / SemanticKITTI) and trav_label (GOOSE) exist on disk. See Preprocessing.