spatialencoder_full

Full SpatialEncoder training dataset prepared on 2026-05-25.

The original relative file layout under the local data root is preserved. Manifests and preparation stats are stored under metadata/spatialencoder_full_20260525/.

Training items:

• CA-1M: 2966
• hyperism: 560
• ADT: 64
• Manifest entries: 391760
• Logical size excluding directory entries: 2044.28 GiB
• Directory entries: 64

Goal

This dataset is intended to become the BOX_DATA_PATH / BOX_DATA_VAL_PATH input tree used by SpatialEncoder training.

After preparation, the training code should see:

${BOX_DATA_PATH}/
├── CA-1M/
│   ├── train/
│   │   └── ca1m-train-<video_id>.tar
│   ├── val/
│   │   └── ca1m-val-<video_id>.tar
│   └── val-unzip/
├── hyperism/
│   └── hyperism/
├── aria_digital_twin/
│   └── ADT/
├── pickle/
│   └── CA-1M/
│       └── *train*.pkl
├── BoxFromMotion/
│   └── dataset/
│       ├── CA-1M.json
│       ├── hyperism.json
│       └── ADT.json
├── json_wo_pose/
└── val-json/

Use:

export BOX_DATA_PATH=/path/to/spatialencoder_full
export BOX_DATA_VAL_PATH=/path/to/spatialencoder_full/BoxFromMotion/dataset

1. Download

Install the Hugging Face CLI if needed:

pip install -U "huggingface_hub[cli]"

Download the dataset while preserving repository paths:

DATA_ROOT=/mnt/nvme6/jieneng/data/spatialencoder_full
mkdir -p "$DATA_ROOT"

huggingface-cli download qicq1c/spatialencoder_full \
  --repo-type dataset \
  --local-dir "$DATA_ROOT" \
  --local-dir-use-symlinks False

The full dataset is about 2 TiB, so make sure the target filesystem has enough space before starting.

2. Expand Packed Add-Ons If Present

If the download contains archive files such as pickle.zip, hyperism-train-json.zip, or hyperism-val-json.zip, unzip them at the data root:

cd "$DATA_ROOT"

for z in pickle.zip hyperism-train-json.zip hyperism-val-json.zip; do
  if [ -f "$z" ]; then
    unzip -o "$z" -d "$DATA_ROOT"
  fi
done

If the download contains hyperism_required_shards/*.tar, expand those shards into the Hyperism frame directory:

mkdir -p "$DATA_ROOT/hyperism/hyperism/unzip"

if [ -d "$DATA_ROOT/hyperism_required_shards" ]; then
  for shard in "$DATA_ROOT"/hyperism_required_shards/*.tar; do
    tar -xf "$shard" -C "$DATA_ROOT/hyperism/hyperism/unzip"
  done
fi

Skip this step for files that are already expanded in the final tree.

3. Check Required Files

Run these checks before training:

export BOX_DATA_PATH="$DATA_ROOT"
export BOX_DATA_VAL_PATH="$DATA_ROOT/BoxFromMotion/dataset"

test -d "$BOX_DATA_PATH/CA-1M/train"
test -d "$BOX_DATA_PATH/CA-1M/val"
test -d "$BOX_DATA_PATH/pickle/CA-1M"
test -d "$BOX_DATA_PATH/hyperism/hyperism"
test -d "$BOX_DATA_PATH/aria_digital_twin/ADT"

test -f "$BOX_DATA_VAL_PATH/CA-1M.json"
test -f "$BOX_DATA_VAL_PATH/hyperism.json"
test -f "$BOX_DATA_VAL_PATH/ADT.json"

find "$BOX_DATA_PATH/CA-1M/train" -name 'ca1m-train-*.tar' | wc -l
find "$BOX_DATA_PATH/pickle/CA-1M" -name '*train*.pkl' | wc -l

Expected minimum result:

• CA-1M/train contains many ca1m-train-*.tar files.
• pickle/CA-1M contains CA-1M iterable training metadata.
• BoxFromMotion/dataset/{CA-1M,hyperism,ADT}.json exist.
• Hyperism and ADT frame paths referenced by the json files exist under BOX_DATA_PATH.

4. Set Training Environment

From the SpatialEncoder code checkout:

cd /path/to/SpatialEncoder

export BOX_DATA_PATH=/path/to/spatialencoder_full
export BOX_DATA_VAL_PATH=/path/to/spatialencoder_full/BoxFromMotion/dataset
export BOX_WEIGHTS_PATH=/path/to/training_output
export SAM3_CHECKPOINT=$BOX_WEIGHTS_PATH/sam3.1_multiplex.pt

export PYTORCH_ALLOC_CONF=expandable_segments:True
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1
export OPENBLAS_NUM_THREADS=1
export NUMEXPR_NUM_THREADS=1
export NCCL_DEBUG=WARN
export TORCH_NCCL_BLOCKING_WAIT=1

Download the SAM 3.1 checkpoint separately into BOX_WEIGHTS_PATH:

wget -P "$BOX_WEIGHTS_PATH" \
  --header="Authorization: Bearer YOUR_HF_TOKEN" \
  https://huggingface.co/facebook/sam3.1/resolve/main/sam3.1_multiplex.pt

5. Dataset Smoke Tests

The merged training config samples datasets according to:

trainer.data.train.dataset.weights = [CA-1M, hyperism, ADT]

Before starting a long run, verify each dataset can print loss:

# CA-1M only
trainer.data.train.dataset.weights='[1,0,0]'

# Hyperism only
trainer.data.train.dataset.weights='[0,1,0]'

# ADT only
trainer.data.train.dataset.weights='[0,0,1]'

Example 8-GPU smoke command:

RUN_NAME=SpatialEncoder_smoke_$(date +%Y%m%d_%H%M%S)

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 env -u LD_LIBRARY_PATH python sam3/train/train.py \
  -c configs/depth/train_merged_iterable_da3_best_memory_extras_lowmem_actckpt_fa3.yaml \
  --use-cluster 0 \
  --num-gpus 8 \
  paths.experiment_log_dir="$BOX_WEIGHTS_PATH/Exps/$RUN_NAME" \
  trainer.model.use_fa3=true \
  trainer.distributed.gradient_as_bucket_view=false \
  trainer.data.train.dataset.weights='[0,1,0]' \
  trainer.logging.log_freq=1 \
  trainer.logging.log_scalar_frequency=1

It is ready if the log reaches lines like:

Train Epoch: [0][   0/...] ... Losses/train_all_loss: ...

The first batch can be slow because workers are filling caches. Later steps should have near-zero Data Time.

6. Mixed Training

Once all three single-dataset smoke tests print loss, launch the mixed run:

RUN_NAME=SpatialEncoder_mixed_8gpu_$(date +%Y%m%d_%H%M%S)

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 env -u LD_LIBRARY_PATH python sam3/train/train.py \
  -c configs/depth/train_merged_iterable_da3_best_memory_extras_lowmem_actckpt_fa3.yaml \
  --use-cluster 0 \
  --num-gpus 8 \
  paths.experiment_log_dir="$BOX_WEIGHTS_PATH/Exps/$RUN_NAME" \
  trainer.model.use_fa3=true \
  trainer.distributed.gradient_as_bucket_view=false \
  trainer.data.train.dataset.weights='[0.4,0.2,0.4]' \
  trainer.logging.log_freq=1 \
  trainer.logging.log_scalar_frequency=1

For quick debugging on slow storage, temporarily add:

scratch.num_train_workers=2

For the default full setting, omit that override; the config uses scratch.num_train_workers=16.

Troubleshooting

• If training appears stuck before the first loss, check whether dataloader workers are still starting. With num_train_workers=16, the first batch can take around 1-2 minutes on large mixed data.
• If only one dataset fails, rerun with the corresponding one-hot weight to isolate missing files.
• If use_fa3=true fails at import or CUDA runtime, retry with trainer.model.use_fa3=false to separate data issues from FA3 compatibility issues.
• If a run is interrupted, kill the whole process group and confirm GPUs are free with:

nvidia-smi --query-compute-apps=pid,process_name,used_memory --format=csv,noheader,nounits