tensor-data

Small finite f32 tensor contracts and metadata for moritzbrantner-video-analysis.

Highlights

• Checked tensor shapes with finite-value validation
• Borrowed and owned tensor views
• Lightweight metadata for interop-oriented tensor payloads
• Reshape helpers that preserve element counts

Example

use tensor_data::{F32Tensor, TensorShape};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let tensor = F32Tensor::new(TensorShape::new([1, 4, 64, 64])?, vec![0.0; 16_384])?;
    assert_eq!(tensor.shape().rank(), 4);
    assert_eq!(tensor.shape().element_count()?, 16_384);
    Ok(())
}

Behavior

TensorShape rejects empty shapes and zero-sized dimensions. Element counts are computed with checked usize multiplication, so extremely large shapes fail instead of wrapping.

F32Tensor and F32TensorView require the number of values to exactly match the shape element count. All tensor values must be finite; NaN, positive infinity, and negative infinity are rejected during construction and validation.

Reshaping changes only the shape metadata. It succeeds when the new dimensions produce the same element count and fails otherwise. Tensor values stay in their existing contiguous order.

Metadata is optional JSON metadata carried alongside tensors for transport and interop. It is not interpreted by this crate and is preserved when converting a view into an owned tensor.

Related crates

• comfyui-latents
• audio-analysis-core
• image-analysis-core