Input NumPy File

2 VersionsInput

Loads NumPy array files from the bound directory and emits each as a Tensor<Float>, either whole or sliced into batches along a chosen axis at a paced rate. Use it to replay saved arrays into a tensor pipeline.

How it fits

Typical backends

Replay saved arrays and report each tensor's contents as JSON.

Stream arrays and log each emitted batch for inspection.

Caveats

I/O contractIn mode batch, slices never span file boundaries and a trailing partial chunk is dropped rather than zero-padded, so the last batch of a file may be missing.
Parameter interactionbatch_size of zero in batch mode falls back to whole-file emission, and batch_axis selects the slicing axis with negative values counting from the end; an axis out of range fails the call.
FallbackAn unrecognised mode value falls back to whole-file emission; only plain array files are read, while archives and pickled arrays are skipped.
Latencyfps paces emission and a value of zero removes the limit; a stalled consumer does not let the component burst to catch up, since the pacing clock resets to the present when it resumes.
State lifetimeThe file list is scanned at startup or when the bound slot changes, so files added at runtime are not picked up until the slot is touched. mode, fps, batch_size, batch_axis, and the slot are hot-swappable while cast_to_float32 binds once at startup.

Versions

097f822edefault
Automated release
6/15/2026