Learn
Explore
Toggle theme
Sign in
Create account
Back to Courses
Progress
Loading...
Module 3
BEV
1
Training Bird’s-Eye View (BEV) Models for Robotics at Scale with Ray on Anyscale
2
Cell 1: Define NuScenes storage paths and validate your runtime
3
Cell 2: Download large datasets reliably with resume support
4
Cell 3: Download the NuScenes mini dataset archive
5
Cell 4: Safely extract the NuScenes dataset into cluster storage
6
Cell 5: Initialize the NuScenes dataset and validate your installation
7
Cell 6: Inspect dataset scale and enumerate available scenes
8
Cell 7: Inspect a single scene sample and its sensor layout
9
Cell 8: Visualize camera annotations for a single timestep
10
Cell 9: Visualize lidar data in the ego-centric top-down frame
11
Cell 10: Project lidar points into the camera image plane
12
Cell 11: Visualize lidar intensity projected into the camera image
13
Cell 12: Render all camera views for a single timestep
14
Cell 13: Create a lightweight subset manifest for scalable training
15
Cell 14: Import core libraries for Ray Data + Ray Train BEV training
16
Cell 15: Define persistent storage paths and select your subset manifest
17
Cell 16: Define your BEV grid, camera inputs, and label space
18
Cell 17: Build a lightweight training manifest from NuScenes sample tokens
19
Cell 18: Split the manifest into training and validation sets
20
Cell 19: Preprocess NuScenes samples into fixed-shape BEV training tensors
21
Cell 20: Build Ray Data datasets for distributed preprocessing
22
Cell 21: Run a one-batch sanity check on your Ray Data pipeline
23
Cell 22: Define a minimal camera-only BEV Transformer and validate shapes
24
Cell 23: Define the Ray Train worker loop with DDP, mixed precision, and checkpoint resume
25
Cell 24: Launch distributed training with TorchTrainer and persist results
26
Conclusion
Feedback
Back to modules
Module progress
0%
