hsafaai/vdc-denoiser-m128-v2

VDC-128 linear-grid denoiser (exploratory retraining)

vdc-denoiser-m128-v2 is the 128 x 128 linear-grid pair-copula estimator called VDC-128 in the appendix "Exploratory retraining on corrected data" of the NeurIPS 2026 paper Amortized Vine Copulas for High-Dimensional Density and Information Estimation. It is an exploratory model trained after the paper's review. The paper's main results use vdc-denoiser-m64-v1.

• Checkpoint: vdc-denoiser-m128-v2.pt, SHA-256 f0a753d50b3cfb089e6d7dec999cfe741c6342c188454ec83183a9b3514d4653
• Architecture: original VDC U-Net denoiser, 132,953,729 parameters, float32 weights
• Training: 8,000 likelihood fine-tuning steps from a 60,000-step pilot trained from random weights (pilot SHA-256 6e48f8c65012f4170ee9003c4aca28600abdb602ef624e47c3c3f92c0f5ace20); inputs use latent uniforms with probability 0.75 and rank pseudo-observations otherwise
• Companion model: vdc-denoiser-m64-v2

Use

Requires vine-denoising-copula v0.3.0 or later, the first release with linear-grid support. The library applies the evaluated inference settings:

from vdc import VineCopulaModel, estimate_pair_copula_from_samples, load_pretrained_model

bundle = load_pretrained_model("vdc-denoiser-m128-v2", device="cuda")
copula = estimate_pair_copula_from_samples(bundle, uv)   # uv: (n, 2) pseudo-observations
copula.pdf(u, v); copula.h_u_given_v(u, v); copula.hinv_u_given_v(q, v)

vine = VineCopulaModel.from_pretrained(bundle, vine_type="dvine")
vine.fit(U_train, diffusion_model=bundle.model, verbose=False)
vine.logpdf(U_test)

The network reads a 128 x 128 histogram of pseudo-observations on uniform cells (counts divided by n times the cell area), log n and the probit coordinates of the cell centers, at t = 0. Its output (clamped to [-15, 15] and exponentiated) gives node values of a bilinear density, projected to uniform margins with integrated basis weights (30 iterations; 15 for S&P100). Each pair copula is re-projected in float64 (IPFP with Newton completion); its PDF, conditional CDFs and their inverses are exact for that density. Evaluated with float32, TF32 disabled, network batches of 16 edges, outer batches of 64 edges and conditional clipping 1e-12.

Evaluation (paper appendix; one H100, four CPU threads)

Bivariate likelihood on the nine cases of the paper's Table 1, five seeds, n = 2,000, 20,000 test points per fit: mean held-out KL 0.00311 +/- 0.00213 nats versus 0.00611 for the original model (49.1% lower; lower on 41/45 fits). Parametric selection remains more accurate on these in-library families (0.00036).

UCI vines (MAF splits, NLL bits/dim, lower is better; PIT-KS over all retained test rows):

Dataset	Train	Original	TLL	vdc-denoiser-m128-v2	PIT-KS
POWER	5k	-0.1413	-0.1822	-0.1639	0.0149
GAS	5k	-1.4773	-1.7050	-1.5957	0.0368
HEPMASS	5k	-0.2372	-0.2009	-0.2643	0.0161
MINIBOONE	5k	-0.9689	-0.7444	-0.9661	0.0232
POWER	50k	-0.1495	-0.2169	-0.1924	0.0103
GAS	50k	-1.4926	-1.8113	-1.6800	0.0245
HEPMASS	50k	-0.2468	-0.2653	-0.2936	0.0138
MINIBOONE	29,556	-0.9784	-0.8716	-0.9856	0.0202

Chronological S&P100 (339 held-out days): NLL -0.41876 bits/dim versus -0.41849 for the original model and -0.29276 for TLL; paired difference from the original -0.00027 (40-day block bootstrap 95% interval [-0.00345, 0.00323]). Fit time 81.0 s; warmed bivariate batches 18.87 ms per edge.

Limitations

• Exploratory; it does not replace the paper's main accuracy or speed claims.
• Model development also used copula families from the paper's out-of-distribution suite, so there is no clean out-of-distribution evaluation.
• The random S&P100 split overlaps model-development data (272 of 339 test days) and is not a held-out evaluation; only the chronological split is.
• One pilot and one fine-tuning seed. A replicate pilot fine-tuned with the same recipe was worse than the original on chronological S&P100 for VDC-128.
• Floating-point results are reproducible bitwise only on the recorded device type and software (PyTorch 2.9.1, NumPy 2.2.6, SciPy 1.15.3, pyvinecopulib 0.7.5).
• Continuous margins and the simplifying assumption are required.

Training and files

• vdc-denoiser-m128-v2.pt: weights and embedded configuration (model_state_dict, config, step).
• manifest.json: library manifest (hash, evaluated inference settings, provenance).
• training/: the pilot, data and fine-tuning recipes, runnable with scripts/train_linear_grid_estimator.py from the library repository.
• evaluation.json: the appendix results above.
• SHA256SUMS: hashes of all files.

Citation

Safaai, H. Amortized Vine Copulas for High-Dimensional Density and Information Estimation. Advances in Neural Information Processing Systems (NeurIPS), 2026.