theairlabcmu/wai-webdataset-metadata

Transfer metadata for WAI shard datasets

This folder is the WebDataset machine's dataset_metadata_dir. It contains the original split lists, packed complete covisibility graphs, and a portable wai-storage.json manifest. Shard catalogs go in scene_catalogs/ after they are built on the machine hosting the shards.

Layout and existing Memor configuration

root_data_dir/
  ase/                  -> /host/shards/ase/
  blendedmvs/           -> /host/shards/blendedmvs/
  scannetppv2/          -> /host/shards/scannetppv2/
  ...other dataset folders or symlinks...
  dataset_metadata/     <- transfer this folder here
    wai-storage.json
    train/*_scene_list_train.npy
    val/*_scene_list_val.npy
    test/*_scene_list_test.npy
    covisibility/*.covisibility.tar
    scene_catalogs/*.metadata.sqlite   <- build on host
    examples/                         <- sample/provenance only

Keep using the three existing machine settings:

root_data_dir: /data/wai
dataset_metadata_dir: ${.root_data_dir}/dataset_metadata
dataset_cache_dir: /local/streaming_metadata_cache

No new config fields or changes to dataset_streaming presets are required. Memor detects wai-storage.json in dataset_metadata_dir and resolves the graph and catalog paths by the dataset's existing scene-list prefix. Dataset roots still come from the existing root: ${machine.root_data_dir}/<dataset> entries. Train and validation use the same automatic detection.

Ordinary metadata directories without that manifest keep individual-file loading. storage_format: files is an optional explicit override. The existing covisibility_root_data_dir setting is bypassed when a packed graph is selected. Use the current writable dataset_cache_dir; node-local SSD is preferable when available. Cache files are regenerated and are not part of this transfer.

If the transfer folder occupies the already configured dataset_metadata_dir, no config edit is needed. Otherwise change only that existing path (and root_data_dir if the dataset location changes). Leave the model/checkpoint and experiment paths in the machine profile as they are.

Use the updated wai-core and Memor checkouts on the host. Install wai-core into the training environment with python -m pip install -e /path/to/wai-core. Copy this directory using a resumable tool such as rsync; the graph archives are ordinary files. There is no need to wrap the 196 GiB of existing tars in another tar. Local hard links used during assembly avoid a second data copy; normal transfer creates independent destination files.

One-time preparation on the shard host

Create dataset symlinks at root_data_dir/<dataset> to directories containing the corresponding tar shards. Then build each full scene catalog, for example:

python /path/to/wai-core/tools/prepare_shard_storage.py metadata \
  --root /data/wai/scannetppv2 \
  --output /data/wai/dataset_metadata/scene_catalogs/scannetppv2.metadata.sqlite \
  --covisibility-path /data/wai/dataset_metadata/covisibility/scannetppv2.covisibility.tar \
  --frame-order file_path

For other datasets, use their original whole-scene JSON to preserve graph order:

python /path/to/wai-core/tools/prepare_shard_storage.py metadata \
  --root /data/wai/ase \
  --output /data/wai/dataset_metadata/scene_catalogs/ase.metadata.sqlite \
  --covisibility-path /data/wai/dataset_metadata/covisibility/ase.covisibility.tar \
  --scene-metadata-root /original/wai/ase

Original metadata means <dataset>/<scene>/scene_meta.json, in the order used to create the complete graph. The ScanNet++ file_path rule is specific to its producer; do not assume it applies to all datasets. The catalog builder verifies embedded shard matrices against the complete graph when present. Supply --shard-pattern if shard filenames differ from *.shard-*.tar.

The ScanNet++ catalog under examples/ covers only scene 6cc2231b9c, not the full dataset. It is deliberately outside scene_catalogs/ so training cannot mistake it for a complete host catalog. Shard catalogs are portable across paths but must be rebuilt after shards are repacked or the scene coverage changes.

If a catalog is absent, the existing fallback scans shards once per parent dataset instance. It still requires a known frame ordering. The manifest records the ScanNet++ ordering rule; other split datasets require a prebuilt catalog or an original scene_metadata_root entry in the manifest. Training never repacks the covisibility graphs or reruns transfer-folder assembly.

Assemble on the source machine (already performed for this transfer)

The repeatable source-side command is:

python tools/prepare_shard_storage.py bundle \
  --archive-dir /media/local/wai-shard-storage \
  --split-metadata-root /home/inf/wai/Memor/data/mapanything_dataset_metadata \
  --output /media/local/wai-transfer-metadata

The command requires a new output folder and preserves all source files. Only split lists for the packed datasets are included; their bytes and order remain unchanged. Existing sample catalogs and the source packing report are included under examples/ for reference.