sturdy.numpy — com.sturdystats/sturdy-numpy 0.1.2

npy->dataset clj

(npy->dataset path)

Read a NumPy .npy file and return its contents as a tech.v3.dataset.

Behavior:

• 1D arrays are returned as a single column dataset (:c1).
• 2D arrays are returned as one dataset column per NumPy column.
• Row-major (C-order) data is transposed during ingestion.
• Column-major (Fortran-order) data is split without transposition.

Column names are generated as :c1, :c2, … in column order.

Dtypes:

• Signed integer and float dtypes are preserved.
• Unsigned NumPy dtypes (:u1, :u2, :u4) are represented with unsigned tech.datatype column dtypes (:uint8, :uint16, :uint32).

This function avoids intermediate persistent collections and is the recommended entry point for large arrays or downstream analytics.

npy->dataset-rowlists clj

(npy->dataset-rowlists path)

Read a 2D NumPy .npy file into a dataset with a single column :c1.

Each row in :c1 is a zero-copy view (dtype-next sub-buffer) of length ncols, representing that row in row-major order.

Notes:

• Only 2D arrays are supported.
• C-order (row-major) files are supported.
• Fortran-order files are not currently supported by this helper.
• The resulting column has elemwise dtype :object (each cell is a buffer). Some downstream systems may not accept object columns for bulk ingestion.

npy->dataset-unnested-nz clj

(npy->dataset-unnested-nz path)

Read a NumPy .npy file and return a sparse, long-form tech.v3.dataset.

This function produces an "UNNEST"-style representation with one row per non-zero entry in the array, using three columns:

• :row_no — 0-based row index (int64)
• :col_no — 0-based column index (int16)
• :val — value at (row_no, col_no), primitive-backed with dtype preserved

Behavior:

• Only non-zero values are emitted (zero entries are skipped).
• 1D arrays are treated as shape (nrows, 1).
• 2D arrays are supported.
• Both C-order (row-major) and Fortran-order (column-major) files are handled correctly without transposition.
• The physical order of rows in the result is unspecified; (row_no, col_no) uniquely identifies each entry.

Dtypes:

• Signed integer and floating-point dtypes are preserved.
• Unsigned NumPy dtypes (:u1, :u2, :u4) are represented using unsigned tech.datatype column dtypes (:uint8, :uint16, :uint32).

Performance characteristics:

• Two-pass ingestion (count non-zeros, then allocate and populate arrays).
• Avoids materializing dense columnar representations.
• No boxing in hot loops.
• Memory usage proportional to the number of non-zero entries.

This representation is especially well-suited for:

• Extremely sparse arrays
• Fast ingestion into databases such as DuckDB
• Downstream aggregation into sparse row-wise or list-based formats

If the array is dense or column-oriented access is required, prefer npy->dataset.

npy->primitive clj

(npy->primitive path)

Read a NumPy .npy file and return its contents backed by primitive arrays.

This represents the lowest-level access to .npy data provided by this namespace.

Returns a map with:

• :shape Vector of dimensions (1D or 2D).
• :dtype Keyword identifying the NumPy dtype (e.g. :f4, :i4, :u2).
• :fortran? Boolean indicating whether the on-disk layout is column-major (Fortran-order) or row-major (C-order).
• :data Flat Java primitive array containing the decoded values, laid out according to :fortran?.

No transposition, reshaping, or conversion to persistent collections is performed. Callers are responsible for interpreting the layout correctly.

This function is intended for high-performance ingestion paths and serves as the foundation for npy->dataset.

npy->vec clj

(npy->vec path)

Read a NumPy .npy file and return its contents as idiomatic Clojure data.

This is a convenience wrapper around sturdy.numpy.read/read-npy that materializes the array as persistent Clojure collections.

Returns a persistent vector (1D) or vector of vectors (2D), in row-major (C-order) layout.

Notes:

• The returned data is always in row-major order, regardless of whether the file is stored as C-order or Fortran-order.
• This representation is easy to inspect and test, but incurs allocation and boxing costs for large arrays. Prefer npy->dataset for performance- sensitive or columnar workloads.