Tensor Display: __repr__ and __str__

hurray.Tensor implements __repr__ and __str__ following NumPy/PyTorch display conventions.

hurray.Tensor

__repr__

For Tier 1 CPU tensors (when NumPy is installed), repr() shows the data values formatted by numpy.array2string, plus the dtype:

import hurray

t = hurray.ones([2, 3], dtype=hurray.float32)
repr(t)
# hurray.Tensor([[1. 1. 1.]
#  [1. 1. 1.]], dtype=float32)

t2 = hurray.arange(5)
repr(t2)
# hurray.Tensor([0 1 2 3 4], dtype=int64)

Large tensors are truncated automatically (NumPy threshold, default 1000 elements):

t = hurray.zeros([1000], dtype=hurray.float64)
repr(t)
# hurray.Tensor([0. 0. 0. ... 0. 0. 0.], dtype=float64)

Fallback (Tier 2 types, non-CPU devices, or NumPy not installed):

# Tier 2 — no NumPy equivalent
t = hurray.Tensor(b'\x21', hurray.dtype.int4, [2])
repr(t)
# hurray.Tensor(shape=(2,), dtype=int4, device=cpu)

__str__

str() returns the bare NumPy-style array string without the hurray.Tensor(...) wrapper — suitable for print():

t = hurray.linspace(0.0, 1.0, 5)
print(t)
# [0.   0.25 0.5  0.75 1.  ]

t2 = hurray.full([3, 3], 7.0, dtype=hurray.float32)
print(t2)
# [[7. 7. 7.]
#  [7. 7. 7.]
#  [7. 7. 7.]]

Falls back to repr() when NumPy is unavailable or for Tier 2 types.

Sparse-layout tensors

Both repr() and str() show format, shape, nnz, and dtype:

import numpy as np
import scipy.sparse as sp
import hurray

m = sp.csr_matrix(([1.0, 2.0], ([0, 1], [1, 0])), shape=(2, 2))

# Hurray's spec requires uint64 index arrays; SciPy builds int32 ones.
m.indices = m.indices.astype(np.uint64)
m.indptr = m.indptr.astype(np.uint64)

t = hurray.from_scipy(m)

repr(t)
# hurray.Tensor(layout='csr', shape=(2, 2), nnz=2, dtype=float64)

print(t)
# hurray.Tensor(layout='csr', shape=(2, 2), nnz=2, dtype=float64)

str() is identical to repr() for sparse tensors. By default the display is metadata only (SciPy-style).

Display options: metadata vs. content

Switch sparse display to a PyTorch-style content form that also shows the per-format buffer arrays. Use hurray.set_print_options to set it globally, or hurray.print_options(...) as a context manager for a scoped change (auto-reverts on exit). The default is "metadata", so existing behavior is unchanged.

import numpy as np
import scipy.sparse as sp
import hurray

m = sp.csr_matrix(([1.0, 2.0, 3.0, 4.0], ([0, 0, 1, 2], [0, 2, 1, 0])), shape=(3, 3))
m.indices = m.indices.astype(np.uint64)
m.indptr = m.indptr.astype(np.uint64)

t = hurray.from_scipy(m)

# Default — metadata only:
repr(t)
# hurray.Tensor(layout='csr', shape=(3, 3), nnz=4, dtype=float64)

# Global switch to content:
hurray.set_print_options(sparse_display="content")
repr(t)
# hurray.Tensor(layout='csr', shape=(3, 3), nnz=4, dtype=float64,
#   values=[1. 2. 3. 4.], col_indices=[0 2 1 0], row_ptr=[0 2 3 4])
hurray.get_print_options()
# {'sparse_display': 'content'}

# Or scope it to a block (reverts automatically):
hurray.set_print_options(sparse_display="metadata")
with hurray.print_options(sparse_display="content"):
    print(repr(t))   # content form
print(repr(t))       # back to metadata

The per-format arrays shown in content mode are:

Note: the sparse component accessors cover only the rank-2, SciPy-interop layouts COO, CSR, and CSC. The CSF (Compressed Sparse Fiber) layout exists in hurray-core (docs/spec/layouts/csf.md); a CSF tensor loads as a hurray.Tensor with layout == "csf", but has no component accessors yet, so it has no Python display form. Exposing rank-N CSF in the Python bindings is future work.

Content mode formats the arrays via NumPy (honoring your active numpy print options); if NumPy is not installed it falls back to the metadata string. set_print_options and print_options are backed by a contextvars.ContextVar, so the setting is isolated per asyncio task / thread context (like the strict/relaxed mode config). An invalid sparse_display value raises ValueError.

Runnable example

cd hurray-python
maturin develop
python examples/display.py