nano_dino

Monorepo scaffold for training an ICON-style operator-learning model with a DINOv2-style training framework.

What is included

• third_party/dinov2/: upstream DINOv2 repo as a git submodule
• third_party/icon_tutorial/: upstream ICON tutorial repo as a git submodule
• src/icondino/: local glue code (datasets, model wrappers, train/eval entrypoints)
• configs/: local experiment configs (paths, model, train defaults)
• scripts/: local run scripts and PBS templates
• data/, checkpoints/, outputs/: local symlinks to scratch-backed runtime directories (created by setup script)

Current status

This repo is initialized as a working scaffold. Data integration is intentionally deferred (per project plan).

Quick start

1. Initialize submodules (already tracked in this repo)
   • bash scripts/setup/init_submodules.sh
2. Link runtime storage to scratch (data/, checkpoints/)
   • bash scripts/setup/link_runtime_to_scratch.sh
3. Prepare uv environment (idempotent/minimal by default)
   • bash scripts/setup/install_uv_env.sh
4. Dry-run local train entrypoint (no real data required)
   • bash scripts/local/train.sh

Design notes

• The project keeps upstream code in third_party/ and concentrates custom logic in src/icondino/.
• The training meta-architecture in src/icondino/models/meta_arch.py exposes DINOv2-like hooks (forward_backward, update_teacher, get_params_groups) so it can later be wired into DINOv2 training infrastructure with minimal changes.
• ICON tutorial currently uses WENO-generated HDF5 data for 1D conservation-law tasks; see docs/registry/icon_tutorial_inventory.md.
• Docs are organized as guide/specs/worklog/registry; see docs/README.md.