drift-sdk 0.1.4

Tiny Python SDK for the Drift dataset API

Drift

Drift is an open-source Python SDK for working with real-world driving data and training end-to-end driving models.

The goal is to make embodied driving research easier to start. Instead of spending days wiring up dataset parsers, route metadata, and segment downloads, you can install one package, load a dataset, and begin building a training pipeline from real driving data.

Today, Drift includes a complete real-world driving dataset with:

• 15.08 hours of human driving
• 933 synchronized driving segments
• 68 routes
• 74.9 GiB of synchronized camera, vehicle telemetry, and navigation data

Drift exposes datasets, routes, and segments as simple Python objects while hiding the storage details underneath. It is best understood as a clean access layer over driving data and artifacts, rather than a fully prepared ML dataset.

Quickstart

For a one-click start, open the example notebook in Google Colab:

• drift_train_driving.ipynb

It is a simple end-to-end quickstart for loading the dataset, downloading segments, training a driving policy, and exporting a model. The notebook is set up to work well on an NVIDIA A100.

Drift aims to do for end-to-end driving data what Hugging Face Datasets did for NLP datasets: provide a simple, reproducible interface that makes experimentation more accessible.

The intended SDK flow is:

1. Load dataset metadata.
2. List routes.
3. Inspect route segment references.
4. Resolve each segment.
5. Download the specific files your pipeline needs.

Installation

pip install drift-sdk

import drift

API endpoint

The hosted API used by the examples below is:

https://drift-api-production-1d47.up.railway.app

You do not need to run the backend locally to try the examples.

Core objects

The SDK exposes three data objects plus one client:

• drift.client.DriftClient
• drift.dataset.Dataset
• drift.route.Route
• drift.segment.Segment

The top-level convenience helpers are:

• drift.load(dataset_id, base_url=None) -> Dataset
• drift.route(route_id, base_url=None) -> Route
• drift.segment(segment_id, base_url=None) -> Segment

Client methods

DriftClient is the full method-level API:

from drift.client import DriftClient

client = DriftClient(base_url="https://drift-api-production-1d47.up.railway.app")

dataset = client.load_dataset("jordy-v0")
routes = client.list_routes("jordy-v0")
route = client.get_route(routes[0].route_id)
segment = client.get_segment(route.segments[0]["segment_id"])

client.close()

Available methods:

• load_dataset(dataset_id) returns a Dataset
• list_routes(dataset_id) returns list[Route]
• get_route(route_id) returns a Route
• get_segment(segment_id) returns a Segment
• close() closes the underlying HTTP client

Dataset methods

Use drift.load(...) or DriftClient.load_dataset(...) when you need dataset summary metadata.

import drift

dataset = drift.load("jordy-v0")
print(dataset.id)
print(dataset.routes_count)
print(dataset.segments_count)

Dataset fields:

• id
• routes_count
• segments_count

Route methods

Use drift.route(...) or DriftClient.get_route(...) when you need the segment references for a route.

import drift

client = drift.client.DriftClient()
routes = client.list_routes("jordy-v0")

route = client.get_route(routes[0].route_id)
print(route.route_id)
print(route.num_segments)
print(route.dataset)
print(route.segments[:3])

client.close()

Route fields:

• route_id
• num_segments
• created_at
• dataset
• segments

Each route.segments item is currently a small reference dictionary such as:

{"segment_id": "...", "segment_index": 0}

Segment methods

Use drift.segment(...) or DriftClient.get_segment(...) when you need the downloadable files for one segment.

import drift

segment = drift.segment("<segment-id>")
print(segment.segment_id)
print(segment.route_id)
print(segment.segment_index)
print(segment.files)
print(segment.download_url)

Segment fields:

• segment_id
• route_id
• segment_index
• files
• download_url

segment.files is a dictionary of alias to presigned URL. Current aliases may include:

• qcamera
• qlog
• rlog
• ecamera
• fcamera
• navigation

Downloading files

Segment.download(destination=None, chunk_size=8192) downloads the segment's primary file. Today that is download_url if present, otherwise files["fcamera"].

import os
import drift

os.makedirs("downloads", exist_ok=True)

segment = drift.segment("<segment-id>")
path = segment.download("downloads/segment.bin")
print(path)

Important limitation:

• Segment.download(...) downloads one file per call.
• It does not download every file listed in segment.files.
• If your pipeline needs navigation, qlog, or another alias specifically, fetch the URL from segment.files and download it yourself.

End-to-end metadata workflow

This is the real SDK usage pattern that a training pipeline should start from.

from pathlib import Path
import drift

BASE_URL = "https://drift-api-production-1d47.up.railway.app"
DATASET_ID = "jordy-v0"
CACHE_DIR = Path("training-cache")
CACHE_DIR.mkdir(parents=True, exist_ok=True)

client = drift.client.DriftClient(base_url=BASE_URL)

dataset = client.load_dataset(DATASET_ID)
print(f"dataset={dataset.id} routes={dataset.routes_count} segments={dataset.segments_count}")

routes = client.list_routes(DATASET_ID)
segment_ids = [
    segment_ref["segment_id"]
    for route in routes
    for segment_ref in route.segments
]

print(f"resolved {len(segment_ids)} segment references")

for segment_id in segment_ids[:8]:
    segment = client.get_segment(segment_id)
    target = CACHE_DIR / f"{segment.segment_id}.bin"
    segment.download(str(target))
    print(f"downloaded {target}")

client.close()

Training-pipeline pattern

The SDK does not expose download_dataset(...), and it does not materialize a ready-made parquet or CSV training table. A trainer should explicitly:

1. Query dataset and route metadata.
2. Expand route segment references into segment IDs.
3. Download the needed segment files.
4. Parse those raw artifacts into model-ready tensors or tables.

This is a simplified version of the same control flow used by train_driving.py:

from pathlib import Path
import drift

BASE_URL = "https://drift-api-production-1d47.up.railway.app"
DATASET_ID = "jordy-v0"
CACHE_DIR = Path(".cache/drift-data")
CACHE_DIR.mkdir(parents=True, exist_ok=True)

client = drift.client.DriftClient(base_url=BASE_URL)
routes = client.list_routes(DATASET_ID)

segment_ids = []
for route in routes:
    for segment_ref in route.segments:
        segment_ids.append(segment_ref["segment_id"])

for segment_id in segment_ids[:32]:
    segment = client.get_segment(segment_id)

    if "navigation" in segment.files:
        print("navigation url:", segment.files["navigation"])

    target = CACHE_DIR / f"{segment.segment_id}.bin"
    segment.download(str(target))

client.close()

If your trainer expects .parquet, .csv, .jsonl, or image tensors, you still need a conversion step after download.

PyTorch example

drift.torch.DriftDataset is a metadata dataset. It does not decode frames or controls for you. It resolves segment metadata lazily and can optionally download one primary file per sample.

from torch.utils.data import DataLoader
from drift.torch import DriftDataset

dataset = DriftDataset(
    dataset_id="jordy-v0",
    base_url="https://drift-api-production-1d47.up.railway.app",
    download_dir=".cache/drift-torch",
)

print("segments:", len(dataset))

sample = dataset[0]
print(sample["segment_id"])
print(sample["files"])
print(sample.get("download_path"))

loader = DataLoader(dataset, batch_size=4, shuffle=False)
batch = next(iter(loader))
print(batch["segment_id"])

The payload returned by each item is shaped for a downstream parser:

• segment_id
• route_id
• segment_index
• files
• download_url
• download_path when download_dir is set

TensorFlow / Keras example

drift.tensorflow.DriftDataset is also metadata-first. Iterate over it directly or convert it into a tf.data.Dataset if you want a Keras-compatible input stream.

import tensorflow as tf
from drift.tensorflow import DriftDataset

segments = DriftDataset(
    dataset_id="jordy-v0",
    base_url="https://drift-api-production-1d47.up.railway.app",
)

for item in segments:
    print(item["segment_id"], item["files"])
    break

keras_dataset = segments.to_keras_dataset()
keras_dataset = keras_dataset.take(4)

for item in keras_dataset:
    tf.print(item["segment_id"], item["download_url"])

This is useful when your actual model input pipeline lives in a later mapping stage that downloads or decodes the referenced artifacts.

Common mistake

This is wrong:

client.load_dataset("jordy-v0")

That only returns summary metadata. It does not download data.

This is the correct conceptual flow:

dataset = client.load_dataset("jordy-v0")
routes = client.list_routes(dataset.id)
route = client.get_route(routes[0].route_id)
segment = client.get_segment(route.segments[0]["segment_id"])
segment.download("downloads/example.bin")

Release notes

0.1.1

• polished package metadata and documentation
• added release workflow and repository scaffolding
• improved SDK examples for downloads and framework integrations