Array Display

Presentation layer for NumPy arrays — structured, 1-D, or 2-D.

ArrayDisplay wraps a NumPy array reference (no copy) and provides seven capabilities for controlling how the data is presented. Plain 1-D and 2-D arrays are reinterpreted as structured via zero-copy np.ndarray.view() under the hood.

1. Column filtering — hide columns by name, regex, or predicate
2. Dtype flattening — expand vector/matrix fields into scalar columns with custom component names
3. Column grouping — bundle related columns under a shared label for two-level header rendering
4. Enum mapping — lazily map integer ID columns to human-readable names
5. Row index — designate a row identifier (column name, group name, or external ndarray)
6. Array slicing — create lightweight index arrays for bounded access
7. Adapters — render as text table or pandas DataFrame on bounded slices

Installation

pip install vcti-array-display>=3.1.0

With pandas support (optional):

pip install vcti-array-display[pandas]>=3.1.0

---

Input Shapes

ArrayDisplay accepts three input shapes. Structured arrays are used as-is; plain 1-D and 2-D arrays are reinterpreted as structured via np.ndarray.view() (zero-copy for C-contiguous input).

import numpy as np
from vcti.arraydisplay import ArrayDisplay

# Structured array — fields already named
dt = np.dtype([('id', 'i4'), ('value', 'f8')])
arr = np.array([(1, 10.0), (2, 20.0)], dtype=dt)
view = ArrayDisplay(arr)                # view.view_columns == ['id', 'value']

# Plain 1-D — single column, default label "value"
view = ArrayDisplay(np.array([1.5, 2.0, 3.5]))
# view.view_columns == ['value']
view = ArrayDisplay(np.array([1.5, 2.0, 3.5]), field_names=['temperature'])

# Plain 2-D — auto-named col_0, col_1, ...
arr2 = np.random.rand(1000, 3)
view = ArrayDisplay(arr2)               # ['col_0', 'col_1', 'col_2']
view = ArrayDisplay(arr2, field_names=['x', 'y', 'z'])           # explicit
view = ArrayDisplay(arr2, field_name_prefix='comp')              # ['comp_0', ...]

Non-contiguous 2-D arrays raise ValueError — call np.ascontiguousarray() first if you accept the copy cost. ndim > 2 is rejected explicitly; reshape first.

---

1. Column Filtering

Hide unwanted columns using any combination of exact names, regex patterns, and callable predicates.

import re
from vcti.arraydisplay import ArrayDisplay, FILLER_COLUMNS, LENGTH_COLUMNS, VOID_COLUMNS

dt = np.dtype([
    ('id', 'i4'), ('f0', 'V4'), ('label', 'U20'),
    ('label_len', 'i4'), ('value', 'f8'), ('f3', 'V8'),
])
arr = np.zeros(10, dtype=dt)

# Pre-built patterns for C++ interop noise
view = ArrayDisplay(arr, exclude_columns=[FILLER_COLUMNS, LENGTH_COLUMNS])
view.view_columns  # ['id', 'label', 'value']

# Or exclude by dtype — catches all void padding regardless of name
view = ArrayDisplay(arr, exclude_columns=[VOID_COLUMNS, LENGTH_COLUMNS])

# Mix exact names, regex, and callables
view = ArrayDisplay(arr, exclude_columns=[
    FILLER_COLUMNS,                        # regex: ^f\d+$
    LENGTH_COLUMNS,                        # regex: _len$
    "debug_flag",                          # exact name
    re.compile(r"^tmp_"),                  # custom regex
    lambda name, dtype: dtype.kind == 'V', # by dtype
])

No magic defaults — ArrayDisplay(arr) shows all columns. Pre-built patterns FILLER_COLUMNS, LENGTH_COLUMNS, and VOID_COLUMNS are opt-in. Callables receive (name, dtype) for filtering by type, shape, or both.

2. Dtype Flattening

Expand vector and matrix fields into individual scalar columns, with optional user-defined component names.

dt = np.dtype([('id', 'i4'), ('position', 'f8', (3,))])
arr = np.zeros(5, dtype=dt)

# Default numeric suffixes (flatten_dtype=True is the default)
view = ArrayDisplay(arr)
view.view_columns  # ['id', 'position_0', 'position_1', 'position_2']

# Custom component names
view = ArrayDisplay(arr, component_names={'position': ['x', 'y', 'z']})
view.view_columns  # ['id', 'position_x', 'position_y', 'position_z']

# Opt out of flattening to preserve the original dtype
view = ArrayDisplay(arr, flatten_dtype=False)
view.view_columns  # ['id', 'position']  (single column, not useful for display)

# Component grouping is tracked for consumers (e.g., header spanning)
view.field_components
# {'position': ['position_x', 'position_y', 'position_z']}

# Rename after construction
view.set_component_names('position', ['lat', 'lon', 'alt'])

Note: Default flattened names (e.g., position_0) are generated by vcti-nputils and may truncate long field names. Use component_names to ensure predictable, readable column names regardless of the defaults.

3. Column Grouping

Bundle existing columns under a shared label. Rendered as a level-1 header in to_table() spanning the group's member columns. Purely visual — does not by itself affect to_dataframe(). Multiple groups may coexist.

dt = np.dtype([
    ('element_idx', 'i4'), ('face_num', 'i4'),
    ('stress', 'f8'),
])
arr = np.array([(1, 0, 34.2), (1, 1, 45.3), (2, 0, 67.8)], dtype=dt)

view = ArrayDisplay(arr)
view.add_column_group('loc', ['element_idx', 'face_num'])

print(view.to_table())
#          loc           |
# element_idx | face_num | stress
# ------------+----------+-------
#           1 |        0 |   34.2
#           1 |        1 |   45.3
#           2 |        0 |   67.8

Group names must not collide with dtype field names or existing group names. remove_column_group(name) removes a group; if it happens to be the current row index, the index is cleared as well.

4. Enum Mapping

Map integer ID columns to human-readable names. The mapping is stored as a recipe and resolved lazily — only when presenting a bounded slice.

dt = np.dtype([('id', 'i4'), ('element_type', 'i4'), ('value', 'f8')])
arr = np.array([(1, 1, 10.5), (2, 2, 20.3), (3, 1, 15.0)], dtype=dt)
view = ArrayDisplay(arr)

view.add_name_columns({
    'element_type_name': ('element_type', {1: 'QUAD', 2: 'HEX'}),
})
view.view_columns  # ['id', 'element_type_name', 'value']

# The mapping is NOT materialized yet — it's a recipe:
view.name_columns
# {'element_type_name': ('element_type', {1: 'QUAD', 2: 'HEX'})}

5. Row Index

Designate what each row represents. Becomes the pandas Index / MultiIndex in to_dataframe() and renders left-most in to_table(). One row index is active at a time.

set_index(source) accepts three forms of source:

Single column (dtype field)

dt = np.dtype([('node_id', 'i4'), ('value', 'f8')])
arr = np.array([(1, 10.0), (2, 20.0), (3, 30.0)], dtype=dt)

view = ArrayDisplay(arr, index='node_id')
# to_table: "node_id" left-most, single-level header
# to_dataframe: pandas Index named "node_id"

Column group name (multi-column via a group)

view = ArrayDisplay(arr)
view.add_column_group('face_id', ['element_idx', 'face_num'])
view.set_index('face_id')

# to_table: "face_id" level-1 over (element_idx, face_num), then data
# to_dataframe: pandas MultiIndex on (element_idx, face_num)

External ndarray (shared across multiple ArrayDisplays)

face_id = np.array(
    [(1, 0), (1, 1), (2, 0)],
    dtype=[('element_idx', 'i4'), ('face_num', 'i4')],
)
stress_arr = np.array([(34.2, 12.1), (45.3, 8.7), (67.8, 22.4)],
                      dtype=[('xx', 'f8'), ('yy', 'f8')])
strain_arr = np.array([(0.01, 0.02), (0.03, 0.01), (0.02, 0.01)],
                      dtype=[('xx', 'f8'), ('yy', 'f8')])

# One identifier, two displays — no duplication of index data
stress = ArrayDisplay(stress_arr, index=face_id)
strain = ArrayDisplay(strain_arr, index=face_id)

Notes:

• Raw values are used — enum name mappings (add_name_columns) don't affect the pandas index; this keeps .loc lookups stable
• A level-1 header appears only when the index is a group name (the group's label is shown); single-column dtype-field index and external ndarray index produce single-level output
• Plain 1-D external index uses "id" as the column label

6. Array Slicing

Create lightweight index arrays for bounded access. The index array is tiny regardless of the underlying array size.

idx = view.to_slice(head=20)                       # first 20 rows
idx = view.to_slice(tail=10)                       # last 10 rows
idx = view.to_slice(head=10, tail=5)               # first 10 + last 5
idx = view.to_slice(mask=arr['value'] > 50)        # boolean filter
idx = view.to_slice(indices=np.array([0, 100, 999]))  # explicit

view.array[idx]                                    # sliced data

7. Adapters

Text table (pure numpy — no extra dependencies)

print(view.to_table(head=10, tail=5))
# id | element_type_name | value
# ---+-------------------+------
#  1 | QUAD              |  10.5
#  2 | HEX               |  20.3
# ...
# [1000 rows x 3 columns]

pandas DataFrame (optional dependency)

df = view.to_dataframe()                   # full array
df = view.to_dataframe(head=100)           # first 100 rows
df = view.to_dataframe(resolve_names=False)  # raw ID columns

Enum names are materialized using pd.Categorical for memory efficiency.

Jupyter notebooks

ArrayDisplay provides _repr_html_() — displays a bounded row window with enum names resolved automatically.

---

Performance

Designed for large CAE arrays (millions of rows, GBs of data):

• No array copy — wraps a reference to the original numpy array
• Lazy enum resolution — resolved only on the displayed slice
• Index arrays — to_slice() returns kilobytes regardless of array size
• Bounded presentation — to_table() and to_dataframe(head=N) never touch the full array
• pd.Categorical — ~100x less memory than string columns for enum values

---

API Summary

Cost column:

• Cheap — metadata only; no iteration over rows.
• Moderate — O(displayed rows × cols) or O(N) one-off; safe for bounded slices.
• Heavy — O(N × C) with allocation; avoid on large arrays unless you need it.

Method	Pillar	Cost	Description
set_view_columns(...)	Filtering	Cheap	Configure visible columns
include_view_columns(cols)	Filtering	Cheap	Add columns to the view
exclude_view_columns(cols)	Filtering	Cheap	Remove columns from the view
replace_view_columns(mapping)	Filtering	Cheap	Rename columns in the view
set_array(arr, ...)	Flattening	Cheap	Set array; vector fields auto-flatten via zero-copy view
set_component_names(field, names)	Flattening	Cheap	Rename flattened components
add_column_group(name, cols)	Grouping	Cheap	Register a visual column group
remove_column_group(name)	Grouping	Cheap	Remove a group
add_name_columns(mappings)	Enum mapping	Cheap	Register lazy enum recipes
set_index(source)	Row index	Cheap	Set row identifier — column name, group name, or ndarray
clear_index()	Row index	Cheap	Remove the row identifier
to_slice(head=H, tail=T)	Slicing	Cheap	O(H + T) index array
to_slice(mask=M)	Slicing	Moderate	O(N) boolean scan
to_table(head=H, tail=T)	Adapter	Moderate	O((H + T) × cols)
to_dataframe(head=H)	Adapter	Moderate	O(H × cols)
to_dataframe() (no bounds)	Adapter	Heavy	O(N × C) pandas materialization
copy()	Misc	Cheap	Shared array, config deep-copied

See docs/performance.md for full complexity analysis.

Note on to_dataframe(): Without head/tail/indices bounds, this method copies the full array into pandas — pandas cannot share memory across heterogeneous dtypes, so expect ~2× memory usage during conversion. Use the bounded form view.to_dataframe(head=100) for inspection and analysis; reserve the unbounded form for genuine full-array export (parquet / CSV / etc).

Examples

See examples/full_pipeline.py for a complete end-to-end script demonstrating all seven pillars.

Dependencies

• numpy (>=1.24) — required
• vcti-nputils (>=1.0.0) — required
• pandas (>=2.0) — optional, for to_dataframe() and Jupyter display