ach0309/classification-cl-rubik-assets

Classification continual-learning assets (RUBIK Pi 3)

Datasets, student/teacher weights and training recipes used by the classification continual-learning (CL) experiments on the RUBIK Pi 3 (QCS6490). The runtime code is not in this repo; it is published separately on GitHub.

License and attribution

The images are object crops cut from BDD100K driving images (ann.json keeps the original image name and bounding box in bdd100k-info). They remain subject to the BDD100K terms of use. Check those terms before using or redistributing the images. Student weights were fine-tuned on Cityscapes crops, whose terms also apply to derived use.

Layout

The folder layout mirrors /home/ubuntu/cl-assets on the RUBIK Pi 3, where the runtime looks for it.

dataset/cl-scenarios/test_250721/scenario-type_{1,2,3}-clear/
    ann.json        32,400 annotations (category 0-8, source image, bbox)
    images.tar      32,400 JPEG crops (extracts to images/)
dataset/provenance/ per-window class mix and validation of the three scenarios
weight/cityscapes-fine-tuned/9-labels/last/{vit-s,vit-b}/
    last-<model>-32-seed_{0,1,2}-...pth             full fine-tuning
    last-<model>-32-lora_r16-seed_{0,1,2}-...pth    LoRA r=16 (base + adapter in one file)
teacher/vit-teacher-checkpoints/restored/
    ViT-L_16_...safetensors, ViT-H_14_...safetensors, *.metadata.json
    calibration / validation / preprocessing JSON for the NPU teacher path
recipe/moe-cl-config/   training configs (Hydra) used for the students
RECIPE.md               training recipes (students, LoRA, teachers)
MANIFEST.json           role, size and SHA-256 of every file
SHA256SUMS              checksums, verify with `sha256sum -c SHA256SUMS`

Classes: 0 car, 1 traffic sign, 2 traffic light, 3 person, 4 truck, 5 bus, 6 bike, 7 motor, 8 train. Each scenario is 9 windows of 3,600 frames (30 FPS x 120 s). Scenario 1 has only 11 train crops.

Download and restore on the RUBIK Pi 3

Set HF_HUB_DISABLE_XET=1 for every download. On the RUBIK Pi 3 the default (Xet) transfer failed with a CDN connection error (CAS Client Error ... error sending request for url (https://...cdn.hf.co/...)), while the plain HTTP transfer worked: 1.4 GB took about 7 minutes, with a few automatic retries after read time-outs. Running the same command again resumes an interrupted download. Other networks may not need this.

Python (this is the method that was tested for a full-size download):

pip install -U huggingface_hub
HF_HUB_DISABLE_XET=1 python3 - <<'EOF'
from huggingface_hub import snapshot_download
snapshot_download('ach0309/classification-cl-rubik-assets', repo_type='dataset', local_dir='/home/ubuntu/cl-assets')
EOF
cd /home/ubuntu/cl-assets
sha256sum -c SHA256SUMS
for s in dataset/cl-scenarios/test_250721/scenario-type_*-clear; do tar -xf $s/images.tar -C $s; done

The whole repository is about 6.9 GB. To download only what you need, pass allow_patterns. This subset (scenario 1, the ViT-S seed-0 student and the ViT-L teacher, about 1.4 GB) was downloaded this way and used for a smoke run:

snapshot_download('ach0309/classification-cl-rubik-assets', repo_type='dataset', local_dir='/home/ubuntu/cl-assets',
  allow_patterns=['SHA256SUMS', 'MANIFEST.json',
    'dataset/cl-scenarios/test_250721/scenario-type_1-clear/*',
    'weight/cityscapes-fine-tuned/9-labels/last/vit-s/last-vit-s-32-seed_0-*',
    'teacher/vit-teacher-checkpoints/restored/ViT-L_16_rank1_epoch30_clean100.safetensors',
    'teacher/vit-teacher-checkpoints/restored/calibration_candidate_cl18.json',
    'teacher/vit-teacher-checkpoints/restored/validation_npu_a16_v4.json'])

With a subset, sha256sum -c SHA256SUMS reports every file you did not download as missing; use sha256sum -c --ignore-missing SHA256SUMS, and extract only the images.tar files that you have.

Command line (hf): the same download can be done as shown below. This form was tested only with two small files (SHA256SUMS and one JSON), not with the large files.

HF_HUB_DISABLE_XET=1 hf download ach0309/classification-cl-rubik-assets --repo-type dataset \
  --local-dir /home/ubuntu/cl-assets --include "SHA256SUMS" \
  --include "dataset/cl-scenarios/test_250721/scenario-type_1-clear/*"

Not included

• Runtime code, cl-runtime Python dependencies, QNN libraries.
• Cityscapes / BDD100K / ImageNet originals and the classification_cl_T1_T3_20261105 teacher training split.
• The original .pt teacher checkpoints (the .safetensors files are exact conversions) and the ImageNet-only ViT-L/16 cost-test teacher.
• Run outputs (cl-runs), including the adapter files the runtime exports.