dataeval.utils.training.predict¶

dataeval.utils.training.predict(x, model, device=None, batch_size=None, preprocess_fn=None, postprocess_fn=None)¶

Make batch predictions on a model.

Parameters:¶

x : Iterable[Array]¶

An iterable of per-instance images (e.g. a numpy array batched along axis 0, a list of arrays, or any iterable yielding one image at a time). Each instance is passed through preprocess_fn individually and only then stacked into batches, so instances need not share a common shape on input – preprocess_fn is responsible for normalizing them (e.g. resizing) to a stackable shape. This is what allows variable-size images (such as detection inputs) to be supplied without pre-batching.

model : torch.nn.Module¶

PyTorch model.

device : DeviceLike or None, default None¶

The hardware device to use if specified, otherwise uses the DataEval default or torch default.

batch_size : int or None, default None¶

Batch size used during prediction. If None, uses DataEval default (1e10).

preprocess_fn : PreprocessFn or None, default None¶

Optional per-instance preprocessing applied to each image before it is stacked into a batch. Receives a single image tensor and returns the transformed tensor; use it to normalize variable-size inputs (e.g. resize) to a common, stackable shape.

postprocess_fn : PostprocessFn or None, default None¶

Optional batch-level decoding applied to each batch’s raw model output before it is moved to the CPU. Receives the model output and must return a tensor or tuple of tensors (e.g. to decode raw detection outputs into per-detection class scores).

Returns:¶

PyTorch tensor with model outputs, or tuple of tensors if model returns tuple (e.g., VAE models return (reconstruction, mu, logvar)).

Return type:¶

torch.Tensor or tuple[torch.Tensor, …]