Output NumPy File

2 VersionsOutput

Writes each upstream Tensor<Float> to a .npy file under output_path, subsampling with every_n and capping retention with max_files. The .npy header carries the tensor shape unless flatten collapses it to 1-D.

How it fits

Typical backends

Tensor checkpointing. Read tensors from disk, run a numeric transform, and write the results back as fresh .npy frames.

Inference capture. Stand in a model stage that emits a float tensor per frame and persist a rolling window of those tensors for offline analysis.

Caveats

Hard constraintThe writer accepts only Tensor<Float>; float32 storage is fixed and no other element type is supported.
FallbackSetting every_n to zero turns the component into a no-op — inputs are consumed but nothing is written.
Parameter interactionWhen flatten is on the tensor is stored as a flat 1-D array; off, the declared multi-dimensional shape is preserved in the file header.
State lifetimeA positive max_files enables a rolling window that deletes the oldest files in output_path, so unrelated .npy files dropped there while the component runs can be removed.
State lifetimeThe frame index is a process-local counter that resets on restart, so reusing output_path overwrites earlier files unless filename_prefix also changes.

Versions

6d98499bdefault
Automated release
8/1/2026