In-process debugging

In-process execution is the fastest way to validate a reader, inspect row shape, or test a transform. It does not use the local launcher worker supervisor.

import refiner as mdr

pipeline = mdr.read_lerobot("hf://datasets/macrodata/aloha_static_battery_ep005_009")
row = pipeline.take(1)[0]

print(row.episode_id)
print(row.num_frames)
print(row.to_frame_table().names)

Use this before writing transforms. It answers the most important question: "What keys and semantic properties does my row actually expose?"

for row in pipeline.iter_rows():
    print(row.episode_id, row.num_frames)
    break

iter_rows() streams rows from the source through all transforms. It is useful for long datasets because it does not materialize the whole pipeline.

rows = pipeline.take(5)

Prefer take() over materialize() unless the dataset is intentionally tiny. materialize() computes every output row into memory.

def show_shape(row):
    print(row.episode_id, row.num_frames)
    return row

pipeline.map(show_shape).take(3)

Keep debug transforms small and remove them before launching real jobs.

Sinks do not write through in-process inspection methods like take() or iter_rows(). Inspect the pipeline before attaching the sink:

source = mdr.read_lerobot("hf://datasets/macrodata/aloha_static_battery_ep005_009")
trimmed = source.map(mdr.robotics.motion_trim(threshold=0.001))

print(trimmed.take(1)[0].num_frames)

Attach the writer when you are ready to launch the pipeline:

pipeline = trimmed.write_lerobot("hf://buckets/acme-robotics/aloha_trimmed")

• Episode Rows
• Row Transforms
• Motion Trimming