google-research/timesfm · error · ValueError

Unsupported array shape: {x.shape}

Error message

Unsupported array shape: {x.shape}

What it means

_to_padded_jax_array pads each covariate array to powers of two, but only handles 1-D and 2-D arrays. Passing an array with ndim >= 3 (or other unsupported shape) raises this ValueError.

Source

Thrown at src/timesfm/utils/xreg_lib.py:57

def _repeat(elements: Iterable[Any], counts: Iterable[int]) -> np.ndarray:
  return np.array(
    list(itertools.chain.from_iterable(map(itertools.repeat, elements, counts)))
  )


def _to_padded_jax_array(x: np.ndarray) -> jax.Array:
  if x.ndim == 1:
    (i,) = x.shape
    di = 2 ** math.ceil(math.log2(i)) - i
    return jnp.pad(x, ((0, di),), mode="constant", constant_values=0.0)
  elif x.ndim == 2:
    i, j = x.shape
    di = 2 ** math.ceil(math.log2(i)) - i
    dj = 2 ** math.ceil(math.log2(j)) - j
    return jnp.pad(x, ((0, di), (0, dj)), mode="constant", constant_values=0.0)
  else:
    raise ValueError(f"Unsupported array shape: {x.shape}")


# Per time series normalization: forward.
def normalize(batch):
  stats = [(np.mean(x), np.where((w := np.std(x)) > _TOL, w, 1.0)) for x in batch]
  new_batch = [(x - stat[0]) / stat[1] for x, stat in zip(batch, stats)]
  return new_batch, stats


# Per time series normalization: inverse.
def renormalize(batch, stats):
  return [x * stat[1] + stat[0] for x, stat in zip(batch, stats)]


class BatchedInContextXRegBase:
  """Helper class for in-context regression covariate formatting.

  Attributes:

View on GitHub (pinned to 331c6d33cb)

Solutions

1. Reshape the covariate array to 1-D or 2-D (per-series 1-D arrays of shape (horizon,) or 2-D (n, horizon)).
2. Check x.ndim/x.shape with numpy before passing; squeeze or drop the extra axis.
3. Split multi-feature covariates into separate named covariates in train_dynamic_numerical_covariates/test_dynamic_numerical_covariates.

Example fix

// before
np.array(cov).shape  # (5, 10, 3)
// after
np.array(cov).reshape(5, 30).shape  # or split into 3 named covariates of shape (5, 10)

import numpy as np
cov = np.asarray(cov_array)
if cov.ndim > 2:
    raise ValueError(f"Covariate array must be 1-D or 2-D, got shape {cov.shape}")

def is_padded_compatible(x) -> bool:
    import numpy as np
    a = np.asarray(x)
    return a.ndim in (1, 2)

try:
    covs.create_covariate_matrix()
except ValueError as e:
    if "Unsupported array shape" in str(e):
        cov_array = np.asarray(cov_array).reshape(len(series), horizon)
        covs.create_covariate_matrix()
    else:
        raise

When it happens

Trigger: Calling xreg fit() (which calls _to_padded_jax_array) with dynamic numerical/categorical covariate arrays that have 3+ dimensions, e.g. shape (n_series, horizon, extra_dim) or mis-shaped lists converted to 3-D numpy arrays.

Common situations: Supplying covariates as nested lists with an unintended extra dimension (e.g. np.array of ragged/extra-nested lists), or passing multi-feature covariates where the API expects one array per series per covariate.

Related errors

• Output dims must be a multiple of 4: {config.output_dims} %
• Memory dimension ({self.qkv_features}) must be divisible by
• Layer norm: {config.attention_norm} not supported.
• Layer norm: {config.feedforward_norm} not supported.
• Activation: {config.ff_activation} not supported.

AI-assisted analysis of google-research/timesfm@331c6d33cb (2026-08-29). Data as JSON: /api/errors/d18f21ca9b71f8b7. Report an issue: GitHub.