dataeval.models.build_model_input¶

dataeval.models.build_model_input(images, spec, *, height=None, width=None)¶

Build the opinionated image input tensor for a CV model.

Each input image is normalized, converted to the color layout declared by spec, resized to the target (height, width), and stacked into a single batched FP32 tensor. Integer pixels are assumed 8-bit [0, 255] and scaled to [0, 1]; float images are assumed already normalized and passed through unchanged.

Parameters:¶

images : Sequence[ArrayLike]¶

Images in CHW (channels, height, width) layout. Each may have a different spatial size; all are resized to the resolved target size.

spec : ModelIOSpec¶

Model input/output contract, typically from read_model_metadata(). Supplies the target color layout and the default height/width.

height : int or None, default None¶

Target height override. Required when spec.height is -1 (variable); otherwise overrides the spec value when set.

width : int or None, default None¶

Target width override. Required when spec.width is -1 (variable); otherwise overrides the spec value when set.

Returns:¶

Batched tensor of shape (B, C, H, W) with values in [0, 1], where B is len(images) and C is 3 for "RGB" or 1 for "GRAYSCALE".

Return type:¶

NDArray[np.float32]

Raises:¶

ValueError – If the resolved height or width is variable (-1) with no override, if an image is not 3-dimensional (CHW), or if its channel count cannot be converted to the target layout.

See also

ModelIOSpec

The input/output contract consumed here.

Examples

>>> import numpy as np
>>> spec = ModelIOSpec(
...     task="IMAGE_CLASSIFICATION",
...     channels="RGB",
...     height=8,
...     width=8,
...     batch_size=-1,
...     n_classes=4,
... )
>>> images = [np.zeros((3, 16, 16), dtype=np.uint8), np.full((3, 12, 20), 255, dtype=np.uint8)]
>>> tensor = build_model_input(images, spec)
>>> tensor.shape
(2, 3, 8, 8)
>>> tensor.dtype
dtype('float32')