hsafaai/vdc-linear-grid-pilot

VDC linear-grid pilot (training initialization)

This checkpoint is the 60,000-step pilot from which the two retrained estimators of the NeurIPS 2026 paper Amortized Vine Copulas for High-Dimensional Density and Information Estimation were fine-tuned: vdc-denoiser-m64-v2 (VDC-64) and vdc-denoiser-m128-v2 (VDC-128). It is published so that their fine-tuning can be repeated exactly. It is a training initialization, not an estimator for analysis; use the two fine-tuned models or vdc-denoiser-m64-v1 for inference.

• Checkpoint: vdc-linear-grid-pilot.pt, SHA-256 6e48f8c65012f4170ee9003c4aca28600abdb602ef624e47c3c3f92c0f5ace20
• Architecture: the VDC U-Net denoiser (132,953,729 parameters), exponential moving-average weights
• Training: 60,000 steps from random weights on exact cell-mass targets of corrected synthetic copulas (a bank of 200,000 target grids built from seed 42), batch 128, learning rate 2e-4 with 2,000 warm-up steps, EMA 0.9995, one H200 GPU (about 1.7 h)

Use

With vine-denoising-copula v0.3.0 or later, download this file and pass it to the fine-tuning recipes, which check its hash:

python scripts/train_linear_grid_estimator.py configs/train/linear_grid_cache.yaml
python scripts/train_linear_grid_estimator.py configs/train/linear_grid_vdc64_v2.yaml \
    initial_checkpoint=/path/to/vdc-linear-grid-pilot.pt

See the library's linear-grid estimator documentation for the recipes, their resources and verification.

Citation

@inproceedings{safaai2026vdc,
  title     = {Amortized Vine Copulas for High-Dimensional Density and Information Estimation},
  author    = {Safaai, Houman},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  note      = {arXiv:2604.20568}
}