Error fitting RVine (`fastkde.pdf`)

[EvoLandEco]

Thank you for the library. I encountered an error when following the examples, trying to fit to a rvine model:

import torch
import pandas as pd
import numpy as np
from sklearn.preprocessing import QuantileTransformer
from torchvinecopulib.vinecop import VineCop  

# Load
real = pd.read_csv("example_csv.csv")
COND_COLS = ["X", "Y"] + [f"Covariates_{i+1}" for i in range(34)] # X, Y, Covariates_1, ..., Covariates_34
TARGET_COLS = [f"A{i+1}" for i in range(22)]  # A1, A2, ..., A22
ALL_COLS    = COND_COLS + TARGET_COLS
real        = real[ALL_COLS]

# Transform
qt   = QuantileTransformer(output_distribution="uniform")
u_re = qt.fit_transform(real)                      # shape (n, D)

# Fit
vine = VineCop( num_dim=u_re.shape[1],
                is_cop_scale=True, 
                num_step_grid=128)
vine.fit( torch.tensor(u_re, dtype=torch.float64, device=DEVICE),
          mtd_vine="rvine",
          mtd_bidep="kendall_tau",
          first_tree_vertex=(0, 1) )  

Error message:

Traceback (most recent call last):
  File "/env/lib/python3.13/site-packages/IPython/core/interactiveshell.py", line 3672, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
    ~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "<ipython-input-27-49f79aab11e3>", line 1, in <module>
    vine.fit( torch.tensor(u_re, dtype=torch.float64, device=DEVICE),
    ~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
              mtd_vine="rvine",
              ^^^^^^^^^^^^^^^^^
              mtd_bidep="kendall_tau",
              ^^^^^^^^^^^^^^^^^^^^^^^^
              first_tree_vertex=(0, 1) )
              ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/env/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/env/lib/python3.13/site-packages/torchvinecopulib/vinecop/__init__.py", line 587, in fit
    bcp.fit(
    ~~~~~~~^
        obs=obs_bcp,
        ^^^^^^^^^^^^
    ...<3 lines>...
        mtd_tll=mtd_tll,
        ^^^^^^^^^^^^^^^^
    )
    ^
  File "/env/lib/python3.13/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/env/lib/python3.13/site-packages/torchvinecopulib/bicop/__init__.py", line 199, in fit
    ).values
      ^^^^^^
AttributeError: 'tuple' object has no attribute 'values'

Could you please shed some light on the source of the error?

[EvoLandEco]

At this line in the repo:

torchvinecopulib/torchvinecopulib/bicop/__init__.py

Lines 193 to 200 in a671595

	elif mtd_kde == "fastKDE":
	pdf_grid = torch.from_numpy(
	fkpdf(
	obs[:, 0].cpu(),
	obs[:, 1].cpu(),
	num_points=self.num_step_grid * 2 + 1,
	).values
	).to(device=device, dtype=dtype)

By manully checking fkpdf (pdf from fastkde), it in fact returns a tuple with attrs 0 and 1, .values seems to be non-applicable here:

from fastkde import pdf as fkpdf
pdf_grid = torch.from_numpy(
    fkpdf(
        u_re[:, 0],
        u_re[:, 1],
        num_points=128 * 2 + 1,
    ).values)

  File "<ipython-input-36-0e4080430025>", line 7, in <module>
    ).values)
      ^^^^^^
AttributeError: 'tuple' object has no attribute 'values'

I am currently using fastkde == 2.1.5, it appears to be the latest version

[tvatter]

@TY-Cheng maybe it would be better to use tell as default for nonparametric since it gives better copula pdf estimates and it's often faster...

EDIT: ah my bad it's actually the case, please discard my comment.

[TY-Cheng]

Hi @EvoLandEco thanks for reporting!
Would suggest using is_cop_scale=True during init (like you did) and mtd_kde='tll' inside .fit() for now, which skips fastkde and use the tll non-parametric bicop (see Nagler, Thomas, Schellhase, Christian and Czado, Claudia. "Nonparametric estimation of simplified vine copula models: comparison of methods" Dependence Modeling, vol. 5, no. 1, 2017, pp. 99-120. https://doi.org/10.1515/demo-2017-0007). This is a quick fix like @tvatter mentioned.

Or can specify fastkde==2.1.3 in your environment.

I will find time to update along fastkde==2.1.5 side, lock its version in .toml, and set default to tll.

[EvoLandEco]

Thanks for the prompt reply. I already tried to use mtd_kde='tll, but more errors would occur in other pipelines. I have reversed fastkde to 2.1.3, I think this would be the best choice for now