From 1d0ae2dd52c75bcefd3704ea461fc5ad0f3bac22 Mon Sep 17 00:00:00 2001 From: Erik Welch Date: Fri, 31 Jul 2026 14:34:59 -0500 Subject: [PATCH 1/3] Render rich reprs without pandas, byte-identical to pandas output The rich repr and _repr_html_ previously built a pandas DataFrame and let pandas render it. Rendering is now done in-house: a lightweight numpy-only frame, a ported text formatter (cell formatting with precision and scientific switchover, column widths, max_rows and max_columns truncation, terminal fit, multiline wrap), and an html builder reproducing the notebook table. With pandas absent, users now get the same rich repr they previously needed pandas for; with pandas present, output is byte-identical, and pandas display options that could affect the old output are still honored (max_rows, min_rows, max_columns, width, expand_frame_repr, precision, max_colwidth, chop_threshold, colheader_justify, float_format, html.border, html.use_mathjax). Known non-reproduced corners, all obscure: CJK east-asian width handling; pprint_nest_depth and max_seq_items on huge UDT cells; notebook_repr_html=False, where pandas' _repr_html_ returns None and the old code embedded the literal string "None" in its place, so the new table is a fix rather than a regression; and max_colwidth of 3 or less, where pandas narrows the column labels too and leaves them ragged while the hand renderer keeps them padded to a uniform width. That last case needs a deliberate max_colwidth <= 3 (the default is 50) together with column labels of differing digit counts, and it affects the text repr only, since the html path strips cell padding. The no-pandas tests now assert the pandas-absent output equals the pandas-present output instead of pinning a header-only fallback. Verified against the previous pandas render with a differential that swaps which formatting module graphblas.core resolves to and compares repr() and _repr_html_() end to end: 12376 checks over shapes, density, mask and transpose views, and every dtype including two UDTs; 82320 checks over float and complex cell formatting (precision 0 through 17, the scientific switchover boundaries, chop_threshold, float_format); 1024 checks over expression and infix-expression reprs in both autocompute modes; and 49140 checks over pandas display options crossed with six terminal sizes. The only divergences are the two described above. Rendering is also faster, since pandas' formatter machinery is no longer built per call: repr of a 4x4 Matrix ~850us -> ~90us, of a 10-element Vector ~1.5ms -> ~124us, of a 1e6-element Vector ~9.6ms -> ~1.1ms, and _repr_html_ of a 4x4 Matrix ~373us -> ~53us. These are best-of-5 with both paths interleaved in one process; the ratios (7x to 12x) held across repeated runs on a shared machine, the absolute figures less so. --- graphblas/core/formatting.py | 784 ++++++++++++++++++++++++++--- graphblas/tests/test_formatting.py | 176 +------ 2 files changed, 738 insertions(+), 222 deletions(-) diff --git a/graphblas/core/formatting.py b/graphblas/core/formatting.py index 5fe9b6972..cd9380c0c 100644 --- a/graphblas/core/formatting.py +++ b/graphblas/core/formatting.py @@ -1,4 +1,11 @@ -# This file imports pandas, so it should only be imported when formatting +# The rich repr and _repr_html_ are hand-rendered here to reproduce pandas' +# DataFrame text/HTML output byte-for-byte without importing pandas. When pandas +# is installed we still read its display.* options so a user's option_context is +# honored; when it is absent we fall back to pandas' documented defaults. +import math +import re +import shutil + import numpy as np from .. import backend, config, monoid, unary @@ -14,6 +21,35 @@ except ImportError: # pragma: no cover (import) has_pandas = False +# pandas display.* defaults, used verbatim when pandas is not installed so the +# hand renderer produces the same output it would with a freshly imported pandas. +_DISPLAY_DEFAULTS = { + "max_rows": 60, + "min_rows": 10, + "max_columns": 0, + "width": 80, + "expand_frame_repr": True, + "precision": 6, + "max_colwidth": 50, + "chop_threshold": None, + "colheader_justify": "right", + "float_format": None, + "html.border": 1, + "html.use_mathjax": True, +} + + +def _display_option(name): + """Read a pandas display. option, or fall back to its default. + + Reading from pandas keeps a user's ``pd.option_context`` honored while the + rendering logic itself stays pandas-free. + """ + if has_pandas: + return pd.get_option(f"display.{name}") + return _DISPLAY_DEFAULTS[name] + + # This was written by a complete novice at CSS. # If you can help make it better, please do! CSS_STYLE = """ @@ -219,7 +255,7 @@ def _update_vector_array(arr, vector, columns, column_offset, *, mask=None): def _get_max_columns(): - max_columns = pd.options.display.max_columns + max_columns = _display_option("max_columns") if max_columns == 0: # We are probably in a terminal and pandas will automatically size the data correctly. # In this case, let's get a sufficiently large amount of data to show and defer to pandas. @@ -241,13 +277,647 @@ def _get_chunk(length, min_length, max_length): return chunk, chunk_groups +class _Column: + """One rendered column: a label, its raw cell values, and how to format them. + + ``kind`` selects the pandas array formatter this column would have used: + "object" (GenericArrayFormatter), "float" (FloatArrayFormatter, also complex), + or "int" (IntArrayFormatter). ``numeric`` mirrors pandas ``is_numeric_dtype`` + and controls whether the column header gets a leading space. + """ + + __slots__ = ("label", "values", "kind", "numeric") + + def __init__(self, label, values, kind, numeric): + self.label = label + self.values = values + self.kind = kind + self.numeric = numeric + + +class _GBFrame: + """A minimal object/typed-column table, standing in for a pandas DataFrame. + + ``col_name`` is the columns' index name (pandas ``df.columns.name``); when set + it appears in the top-left corner cell, as vector reprs rely on. + """ + + __slots__ = ("columns", "index", "col_name") + + def __init__(self, columns, index, col_name=None): + self.columns = columns + self.index = index + self.col_name = col_name + + @property + def ncols(self): + return len(self.columns) + + @property + def nrows(self): + return len(self.index) + + def slice_cols(self, idx): + return _GBFrame([self.columns[i] for i in idx], self.index, self.col_name) + + def slice_rows(self, idx): + columns = [ + _Column(c.label, [c.values[i] for i in idx], c.kind, c.numeric) for c in self.columns + ] + return _GBFrame(columns, [self.index[i] for i in idx], self.col_name) + + +def _isna_cell(x): + # Gaps in the dense grid are float NaN; present values (including displayed + # "nan"/"inf") are never a bare float NaN, so this only flags the gaps. + return x is None or (isinstance(x, float) and math.isnan(x)) + + +def _count_present(arr): + return sum(1 for x in arr.flat if not _isna_cell(x)) + + +def _dtype_kind_numeric(dtype): + kind = dtype.kind + if kind in "fc": + return "float", True + if kind in "iu": + return "int", True + if kind == "b": + # bool renders via the generic formatter, but is_numeric_dtype(bool) is True + return "object", True + return "object", False + + +def _make_dense_frame(arr, columns, index): + nrows = len(index) + out = [] + for j, label in enumerate(columns): + vals = [("" if _isna_cell(arr[i, j]) else arr[i, j]) for i in range(nrows)] + out.append(_Column(label, vals, "object", False)) + return _GBFrame(out, list(index)) + + +def _make_coo_frame(label_arrays, add_dots): + n = len(label_arrays[0][1]) + index = list(range(n)) + out = [] + for label, values in label_arrays: + values = np.asarray(values) + if add_dots: + out.append(_Column(label, [*values.tolist(), "..."], "object", False)) + else: + kind, numeric = _dtype_kind_numeric(values.dtype) + out.append(_Column(label, values.tolist(), kind, numeric)) + if add_dots: + index.append("...") + return _GBFrame(out, index) + + +# --- cell formatting (reproduces pandas array formatters for object/int/float) --- + +_NUMBER_RE = re.compile(r"^\s*[\+-]?[0-9]+\.[0-9]*$") + + +def _is_float_scalar(v): + # Matches pandas.lib.is_float: python/numpy floats, but not bool/int/complex. + return isinstance(v, (float, np.floating)) + + +def _pprint(v): + # Reproduces pandas printing.pprint_thing for our cell types (escape_chars for + # tab/cr/nl, quote_strings=False): scalars -> str, sequences recurse. + if isinstance(v, (list, tuple)): + body = ", ".join(_pprint(e) for e in v) + if isinstance(v, tuple) and len(v) == 1: + body += "," + return f"[{body}]" if isinstance(v, list) else f"({body})" + s = str(v) + return s.replace("\t", r"\t").replace("\r", r"\r").replace("\n", r"\n") + + +def _trim_zeros_single_float(s): + s = s.rstrip("0") + if s.endswith("."): + s += "0" + return s + + +def _trim_zeros_float(str_floats): + trimmed = list(str_floats) + + def is_number_with_decimal(x): + return _NUMBER_RE.match(x) is not None + + def should_trim(values): + numbers = [x for x in values if is_number_with_decimal(x)] + return len(numbers) > 0 and all(x.endswith("0") for x in numbers) + + while should_trim(trimmed): + trimmed = [x[:-1] if is_number_with_decimal(x) else x for x in trimmed] + return [x + "0" if is_number_with_decimal(x) and x.endswith(".") else x for x in trimmed] + + +def _trim_zeros_complex(str_complexes): + real_part, imag_part = [], [] + for x in str_complexes: + trimmed = re.split(r"(?{padded_length}}" + "j" + for real_pt, imag_pt in zip(padded_parts[:n], padded_parts[n:], strict=True) + ] + + +def _value_formatter(fmt_str, threshold): + def base(v): + return fmt_str.format(value=v) + + if threshold is None: + return base + + def formatter(v): + return base(v) if abs(v) > threshold else base(0.0) + + return formatter + + +def _format_reals_with_na(values, formatter, na_rep): + return [na_rep if (v != v) else formatter(v) for v in values] + + +def _format_complex_with_na(values, formatter, na_rep): + out = [] + for val in values: + re_v, im_v = val.real, val.imag + re_na, im_na = re_v != re_v, im_v != im_v + if not re_na and not im_na: + out.append(formatter(val)) + elif not re_na: + out.append(f"{formatter(re_v)}+{na_rep}j") + elif not im_na: + imag_formatted = formatter(im_v).strip() + if imag_formatted.startswith("-"): + out.append(f"{na_rep}{imag_formatted}j") + else: + out.append(f"{na_rep}+{imag_formatted}j") + else: + out.append(f"{na_rep}+{na_rep}j") + return out + + +def _format_float_column(values, digits): + # Reproduces FloatArrayFormatter for fixed_width, leading_space=True, na_rep="NaN". + arr = np.asarray(values) + is_complex = np.iscomplexobj(arr) + na_rep = "NaN" + if (float_format := _display_option("float_format")) is not None: + # A user display.float_format callable makes FloatArrayFormatter drop + # fixed_width: each value (real or complex) is just float_format(value), + # with no trailing-zero trim and no scientific switchover. Iterate the + # numpy array (not .tolist()) so the callable receives numpy scalars, as + # pandas does; e.g. a "%f"-style callable then casts complex the same way. + return [na_rep if (v != v) else float_format(v) for v in arr] + seq = arr.tolist() + threshold = _display_option("chop_threshold") + + def format_with(fmt_str): + formatter = _value_formatter(fmt_str, threshold) + if is_complex: + return _trim_zeros_complex(_format_complex_with_na(seq, formatter, na_rep)) + return _trim_zeros_float(_format_reals_with_na(seq, formatter, na_rep)) + + result = format_with(f"{{value: .{digits:d}f}}") + too_long = bool(result) and max(len(x) for x in result) > digits + 6 + abs_vals = np.abs(arr) + has_large = bool((abs_vals > 1e6).any()) + has_small = bool(((abs_vals < 10.0 ** (-digits)) & (abs_vals > 0)).any()) + if has_small or (too_long and has_large): + result = format_with(f"{{value: .{digits:d}e}}") + return list(result) + + +def _justify(strings, width, mode="right"): + if mode == "left": + return [x.ljust(width) for x in strings] + if mode == "center": + return [x.center(width) for x in strings] + return [x.rjust(width) for x in strings] + + +def _make_fixed_width(strings, justify="right", minimum=None): + if not strings: + return list(strings) + max_len = max(len(x) for x in strings) + if minimum is not None: + max_len = max(minimum, max_len) + conf_max = _display_option("max_colwidth") + if conf_max is not None and max_len > conf_max: + max_len = conf_max + + def just(x): + if conf_max is not None and conf_max > 3 and len(x) > max_len: + x = x[: max_len - 3] + "..." + return x + + return _justify([just(x) for x in strings], max_len, justify) + + +def _format_labels(labels): + # Reproduce pandas Index._format_flat(include_name=False) for our label types: + # integer labels are padded to a uniform width (left-justified, with a sign + # column when any are negative); string labels are left as-is. + if labels and all(isinstance(x, (int, np.integer)) and not isinstance(x, bool) for x in labels): + pattern = "{: d}" if any(x < 0 for x in labels) else "{:d}" + strs = [pattern.format(x) for x in labels] + width = max(len(s) for s in strs) + return [s.ljust(width) for s in strs] + return [str(x) for x in labels] + + +def _adjoin(space, lists): + # Port of pandas printing.adjoin (ascii len/ljust); glues columns with `space`. + lengths = [max(map(len, x)) + space for x in lists[:-1]] + lengths.append(max(map(len, lists[-1]))) + max_len = max(map(len, lists)) + padded = [] + for i, lst in enumerate(lists): + nl = [x.ljust(lengths[i]) for x in lst] + nl = [" " * lengths[i]] * (max_len - len(lst)) + nl + padded.append(nl) + return "\n".join("".join(parts) for parts in zip(*padded, strict=True)) + + +def _binify(cols, line_width): + adjoin_width = 1 + bins = [] + curr_width = 0 + i_last = len(cols) - 1 + for i, w in enumerate(cols): + w_adjoined = w + adjoin_width + curr_width += w_adjoined + if i_last == i: + wrap = curr_width + 1 > line_width and i > 0 + else: + wrap = curr_width + 2 > line_width and i > 0 + if wrap: + bins.append(i) + curr_width = w_adjoined + bins.append(len(cols)) + return bins + + +def _console_width(): + # pandas repr sets the wrap width from console.get_console_size(); reuse it + # when present so the wrap decision is identical. Without pandas (or if that + # private module moved) fall back to display.width. + if has_pandas: + try: + from pandas.io.formats.console import get_console_size + + return get_console_size()[0] + except Exception: # pragma: no cover (defensive across pandas versions) + pass + return _display_option("width") + + +class _TextFormatter: + """Reproduces pandas DataFrameFormatter + StringFormatter for text repr.""" + + def __init__(self, frame, max_rows, min_rows, max_cols): + self.frame = frame + self.max_rows = max_rows + self.min_rows = min_rows + self.max_cols = max_cols + self.justify = _display_option("colheader_justify") + self.tr_frame = frame + self.tr_col_num = None + self.tr_row_num = None + self.max_cols_fitted = self._calc_max_cols_fitted() + self.max_rows_fitted = self._calc_max_rows_fitted() + self.truncate() + + def _is_in_terminal(self): + return self.max_cols == 0 or self.max_rows == 0 + + def _calc_max_cols_fitted(self): + if not self._is_in_terminal(): + return self.max_cols + width = shutil.get_terminal_size()[0] + if self.max_cols == 0 and self.frame.ncols > width: + return width + return self.max_cols + + def _calc_max_rows_fitted(self): + if self._is_in_terminal() and self.max_rows == 0: + # rows available for data: terminal height minus dots + prompt + header + return shutil.get_terminal_size()[1] - 3 + max_rows = self.max_rows + if max_rows and self.frame.nrows > max_rows and self.min_rows: + max_rows = min(self.min_rows, max_rows) + return max_rows + + @property + def is_truncated_horizontally(self): + return bool(self.max_cols_fitted and self.frame.ncols > self.max_cols_fitted) + + @property + def is_truncated_vertically(self): + return bool(self.max_rows_fitted and self.frame.nrows > self.max_rows_fitted) + + @property + def is_truncated(self): + return self.is_truncated_horizontally or self.is_truncated_vertically + + def truncate(self): + if self.is_truncated_horizontally: + self._truncate_horizontally() + if self.is_truncated_vertically: + self._truncate_vertically() + + def _truncate_horizontally(self): + col_num = self.max_cols_fitted // 2 + if col_num >= 1: + _len = self.tr_frame.ncols + self.tr_frame = self.tr_frame.slice_cols( + [*range(col_num), *range(_len - col_num, _len)] + ) + else: + col_num = self.max_cols + self.tr_frame = self.tr_frame.slice_cols(list(range(col_num))) + self.tr_col_num = col_num + + def _truncate_vertically(self): + row_num = self.max_rows_fitted // 2 + if row_num >= 1: + _len = self.tr_frame.nrows + self.tr_frame = self.tr_frame.slice_rows( + [*range(row_num), *range(_len - row_num, _len)] + ) + else: + row_num = self.max_rows + self.tr_frame = self.tr_frame.slice_rows(list(range(row_num))) + self.tr_row_num = row_num + + def _format_col_raw(self, col): + if col.kind == "int": + return [f"{x: d}" for x in col.values] + if col.kind == "float": + return _format_float_column(col.values, _display_option("precision")) + precision = _display_option("precision") + float_format = _display_option("float_format") + out = [] + for v in col.values: + # A float NaN is excluded from pandas' float-format branch (it uses + # is_float(v) & notna(v)) and rendered as the na_rep "NaN" instead. + if _is_float_scalar(v) and not math.isnan(v): + if float_format is not None: + # A user display.float_format callable replaces the default + # precision render (and adds no sign-space of its own). + out.append(float_format(v)) + else: + out.append(_trim_zeros_single_float(f"{v: .{precision}f}")) + elif v is None: + out.append(" None") + elif _is_float_scalar(v): + out.append(" NaN") + else: + out.append(f" {_pprint(v)}") + return out + + def _get_body_strcols(self): + # Column labels are formatted together (integer labels padded to a uniform + # width) the way pandas Index._format_flat does, not per column. + labels = _format_labels([col.label for col in self.tr_frame.columns]) + strcols = [] + for col, label in zip(self.tr_frame.columns, labels, strict=True): + header = f" {label}" if col.numeric else label + header_colwidth = len(header) + # pandas fixes width twice: format_array right-justifies to the cell + # content width, then the body pass re-justifies with colheader_justify + # (which only matters when the header is wider, or when it is "left"). + fmt_values = _make_fixed_width(self._format_col_raw(col), "right") + fmt_values = _make_fixed_width(fmt_values, self.justify, minimum=header_colwidth) + max_len = max(max((len(x) for x in fmt_values), default=0), header_colwidth) + cheader = _justify([header], max_len, self.justify) + strcols.append(cheader + fmt_values) + return strcols + + def _get_index_strcol(self): + idx = _make_fixed_width([str(x) for x in self.tr_frame.index], justify="left") + corner = "" if self.frame.col_name is None else str(self.frame.col_name) + return [corner, *idx] + + def get_strcols(self): + strcols = self._get_body_strcols() + strcols.insert(0, self._get_index_strcol()) + return strcols + + @property + def _adjusted_tr_col_num(self): + return self.tr_col_num + 1 # index column is always shown + + def _insert_dot_separators(self, strcols): + index_length = len(self._get_index_strcol()) + if self.is_truncated_horizontally: + strcols.insert(self._adjusted_tr_col_num, [" ..."] * index_length) + if self.is_truncated_vertically: + self._insert_dots_vertical(strcols, index_length) + return strcols + + def _insert_dots_vertical(self, strcols, index_length): + n_header_rows = index_length - self.tr_frame.nrows + row_num = self.tr_row_num + for ix, col in enumerate(strcols): + cwidth = len(col[row_num]) + is_dot_col = self.is_truncated_horizontally and ix == self._adjusted_tr_col_num + dots = "..." if (cwidth > 3 or is_dot_col) else ".." + if ix == 0: + dot_mode = "left" + elif is_dot_col: + cwidth = 4 + dot_mode = "right" + else: + dot_mode = "right" + col.insert(row_num + n_header_rows, _justify([dots], cwidth, dot_mode)[0]) + + def _get_strcols(self): + strcols = self.get_strcols() + if self.is_truncated: + strcols = self._insert_dot_separators(strcols) + return strcols + + def _fit_to_terminal(self, strcols): + lines = _adjoin(1, strcols).split("\n") + max_len = max(len(x) for x in lines) + width = shutil.get_terminal_size()[0] + adj_dif = max_len - width + 1 # +1 to avoid too-wide repr (pandas GH #17023) + col_lens = [max((len(x) for x in col), default=0) for col in strcols] + n_cols = len(col_lens) + while adj_dif > 0 and n_cols > 1: + mid = round(n_cols / 2) + adj_dif -= col_lens.pop(mid) + 1 + n_cols = len(col_lens) + max_cols_fitted = max(n_cols - 1, 2) # minus index column; show at least two + self.max_cols_fitted = max_cols_fitted + self.truncate() + return _adjoin(1, self._get_strcols()) + + def _join_multiline(self, strcols, line_width): + adjoin_width = 1 + strcols = list(strcols) + idx = strcols.pop(0) + line_width -= max(len(x) for x in idx) + adjoin_width + col_widths = [max((len(x) for x in col), default=0) for col in strcols] + col_bins = _binify(col_widths, line_width) + nbins = len(col_bins) + blocks = [] + start = 0 + for i, end in enumerate(col_bins): + row = strcols[start:end] + row.insert(0, idx) + if nbins > 1: + nrows = len(row[-1]) + if end <= len(strcols) and i < nbins - 1: + row.append([" \\", *[" "] * (nrows - 1)]) + else: + row.append([" "] * nrows) + blocks.append(_adjoin(adjoin_width, row)) + start = end + return "\n\n".join(blocks) + + def to_string(self, line_width): + strcols = self._get_strcols() + if line_width is None: + return _adjoin(1, strcols) + if self.max_cols > 0: + return self._join_multiline(strcols, line_width) + return self._fit_to_terminal(strcols) + + +def _render_text(frame): + max_cols = _display_option("max_columns") + line_width = _console_width() if _display_option("expand_frame_repr") else None + fmt = _TextFormatter(frame, _display_option("max_rows"), _display_option("min_rows"), max_cols) + return fmt.to_string(line_width) + + +# The scoped style block pandas' NotebookFormatter emits ahead of the table. +_HTML_STYLE = ( + "" +) + + +def _html_escape(s): + return s.replace("&", "&").replace("<", "<").replace(">", ">") + + +class _HtmlBuilder: + """Reproduces pandas NotebookFormatter (DataFrame._repr_html_) markup.""" + + indent_delta = 2 + + def __init__(self, fmt): + self.fmt = fmt + self.tr = fmt.tr_frame + self.ncols = self.tr.ncols + self.th = fmt.is_truncated_horizontally + self.tv = fmt.is_truncated_vertically + self.row_levels = 1 # single-level index, always shown + self.elements = [] + + def write(self, s, indent=0): + self.elements.append(" " * indent + s) + + def _cell(self, s, kind, indent): + rs = _html_escape(str(s)).strip().replace(" ", "  ") + self.write(f"<{kind}>{rs}", indent) + + def write_tr(self, line, indent, header=False, align=None, nindex_levels=0): + self.write("" if align is None else f'', indent) + inner = indent + self.indent_delta + for i, s in enumerate(line): + self._cell(s, "th" if (header or i < nindex_levels) else "td", inner) + self.write("", indent) + + def _col_header(self, indent): + row = ["" if self.tr.col_name is None else str(self.tr.col_name)] + row.extend(_format_labels([col.label for col in self.tr.columns])) + if self.th: + row.insert(self.row_levels + self.fmt.tr_col_num, "...") + self.write_tr(row, indent, header=True, align=self.fmt.justify) + + def _body(self, indent): + index_labels = _format_labels(list(self.tr.index)) + col_cells = [ + _make_fixed_width(self.fmt._format_col_raw(col), "right") for col in self.tr.columns + ] + row = [] + for i in range(self.tr.nrows): + if self.tv and i == self.fmt.tr_row_num: + self.write_tr(["..."] * len(row), indent, nindex_levels=self.row_levels) + row = [index_labels[i], *(col_cells[j][i] for j in range(self.ncols))] + if self.th: + row.insert(self.fmt.tr_col_num + self.row_levels, "...") + self.write_tr(row, indent, nindex_levels=self.row_levels) + + def _table(self, indent=0): + classes = "dataframe" + if not _display_option("html.use_mathjax"): + classes = "dataframe tex2jax_ignore mathjax_ignore" + # pandas keeps the attribute for any non-None border, including 0. + border = _display_option("html.border") + border_attr = "" if border is None else f' border="{border}"' + self.write(f'', indent) + self.write("", indent + self.indent_delta) + self._col_header(indent + 2 * self.indent_delta) + self.write("", indent + self.indent_delta) + self.write("", indent + self.indent_delta) + self._body(indent + 2 * self.indent_delta) + self.write("", indent + self.indent_delta) + self.write("", indent) + + def render(self): + self.write("
") + self.write(_HTML_STYLE) + self._table(0) + self.write("
") + return "\n".join(self.elements) + + +def _render_html(frame): + fmt = _TextFormatter( + frame, + _display_option("max_rows"), + _display_option("min_rows"), + _display_option("max_columns"), + ) + return _HtmlBuilder(fmt).render() + + def _get_matrix_dataframe(matrix, max_rows, min_rows, max_columns, *, mask=None): - if not has_pandas: - return if max_rows is None: # pragma: no branch - max_rows = pd.options.display.max_rows + max_rows = _display_option("max_rows") if min_rows is None: # pragma: no branch - min_rows = pd.options.display.min_rows + min_rows = _display_option("min_rows") if max_columns is None: # pragma: no branch max_columns = _get_max_columns() rows, row_groups = _get_chunk(matrix._nrows, min_rows, max_rows) @@ -264,12 +934,12 @@ def _get_matrix_dataframe(matrix, max_rows, min_rows, max_columns, *, mask=None) column_offset, mask=mask, ) - df = pd.DataFrame(arr, columns=columns, index=rows) + present = _count_present(arr) + truncated = (len(rows), len(columns)) != matrix.shape if ( (mask is None or mask.structure) - and df.shape != matrix.shape - and min(matrix._nvals, max_rows if matrix._nvals <= max_rows else min_rows) - > 2 * df.count().sum() + and truncated + and min(matrix._nvals, max_rows if matrix._nvals <= max_rows else min_rows) > 2 * present ): # The data is sparse and it's better to show in COO format. # SS, SuiteSparse-specific: head @@ -283,47 +953,43 @@ def _get_matrix_dataframe(matrix, max_rows, min_rows, max_columns, *, mask=None) vals = np.zeros(vals.size, dtype=np.uint8) else: vals = np.ones(vals.size, dtype=np.uint8) - df = pd.DataFrame({"row": rows, "col": cols, "val": vals}) - if num_rows < matrix._nvals: - df.loc["..."] = ["..."] * 3 - return df - if mask is not None and not mask.structure and df.shape != matrix.shape: + return _make_coo_frame( + [("row", rows), ("col", cols), ("val", vals)], num_rows < matrix._nvals + ) + if mask is not None and not mask.structure and truncated: # This performs more calculation and uses more memory than I would prefer. # Perhaps we could use the efficient "constant vector or matrix" trick. nonzero = matrix.apply(unary.one["UINT8"]).new(mask=matrix.V, name="") num_rows = matrix._nvals if matrix._nvals <= max_rows else min_rows - if min(nonzero._nvals, num_rows) > 2 * df.count().sum(): + if min(nonzero._nvals, num_rows) > 2 * present: rows, cols, vals = nonzero.ss.head(num_rows, sort=True) if mask.complement: if not vals.flags.writeable: # pragma: no cover (safety) vals = vals.copy() vals[:] = 0 - df = pd.DataFrame({"row": rows, "col": cols, "val": vals}) - if num_rows < nonzero._nvals: - df.loc["..."] = ["..."] * 3 - return df - return df.where(pd.notna(df), "") + return _make_coo_frame( + [("row", rows), ("col", cols), ("val", vals)], num_rows < nonzero._nvals + ) + return _make_dense_frame(arr, columns, rows) def _get_vector_dataframe(vector, max_rows, min_rows, max_columns, *, mask=None): - if not has_pandas: - return if max_rows is None: # pragma: no branch - max_rows = pd.options.display.max_rows + max_rows = _display_option("max_rows") if min_rows is None: # pragma: no branch - min_rows = pd.options.display.min_rows + min_rows = _display_option("min_rows") if max_columns is None: # pragma: no branch max_columns = _get_max_columns() columns, column_groups = _get_chunk(vector._size, max_columns, max_columns) arr = np.full((1, len(columns)), np.nan, dtype=object) for column_group, column_offset in column_groups: _update_vector_array(arr, vector, column_group, column_offset, mask=mask) - df = pd.DataFrame(arr, columns=columns, index=[""]) + present = _count_present(arr) + truncated = len(columns) != vector._size if ( (mask is None or mask.structure) - and df.size != vector._size - and min(vector._nvals, max_rows if vector._nvals <= max_rows else min_rows) - > 2 * df.count().sum() + and truncated + and min(vector._nvals, max_rows if vector._nvals <= max_rows else min_rows) > 2 * present ): # The data is sparse and it's better to show in COO format. # SS, SuiteSparse-specific: head @@ -334,26 +1000,20 @@ def _get_vector_dataframe(vector, max_rows, min_rows, max_columns, *, mask=None) vals = np.zeros(vals.size, dtype=np.uint8) else: vals = np.ones(vals.size, dtype=np.uint8) - df = pd.DataFrame({"index": indices, "val": vals}) - if num_rows < vector._nvals: - df.loc["..."] = ["..."] * 2 - return df - if mask is not None and not mask.structure and df.size != vector._size: + return _make_coo_frame([("index", indices), ("val", vals)], num_rows < vector._nvals) + if mask is not None and not mask.structure and truncated: # This performs more calculation and uses more memory than I would prefer. # Perhaps we could use the efficient "constant vector or matrix" trick. nonzero = vector.apply(unary.one["UINT8"]).new(mask=vector.V, name="") num_rows = vector._nvals if vector._nvals <= max_rows else min_rows - if min(nonzero._nvals, num_rows) > 2 * df.count().sum(): + if min(nonzero._nvals, num_rows) > 2 * present: indices, vals = nonzero.ss.head(num_rows, sort=True) if mask.complement: if not vals.flags.writeable: # pragma: no cover (safety) vals = vals.copy() vals[:] = 0 - df = pd.DataFrame({"index": indices, "val": vals}) - if num_rows < nonzero._nvals: - df.loc["..."] = ["..."] * 2 - return df - return df.where(pd.notna(df), "") + return _make_coo_frame([("index", indices), ("val", vals)], num_rows < nonzero._nvals) + return _make_dense_frame(arr, columns, [""]) def get_format(x, is_transposed=False): @@ -436,14 +1096,9 @@ def vector_expression_header_html(matrix, expr): return create_header_html(name, keys, vals) -def _format_html(name, header, df, collapse): - if has_pandas: - state = "" if collapse else " open" - with pd.option_context("display.show_dimensions", False, "display.large_repr", "truncate"): - details = df._repr_html_() - else: - state = "" - details = "(Install pandas to see a preview of the data)" +def _format_html(name, header, frame, collapse): + state = "" if collapse else " open" + details = _render_html(frame) return ( "
" f"{CSS_STYLE}" @@ -667,17 +1322,12 @@ def format_matrix(matrix, *, max_rows=None, min_rows=None, max_columns=None, mas name, keys, vals, - lower_border=has_pandas, + lower_border=True, name=matrix.name if mask is None else mask.name, ) - if has_pandas: - df = _get_matrix_dataframe(matrix, max_rows, min_rows, max_columns, mask=mask) - if 0 not in matrix.shape: - with pd.option_context( - "display.show_dimensions", False, "display.large_repr", "truncate" - ): - df_repr = df.__repr__() - return f"{header}\n{df_repr}" + if 0 not in matrix.shape: + frame = _get_matrix_dataframe(matrix, max_rows, min_rows, max_columns, mask=mask) + return f"{header}\n{_render_text(frame)}" return header @@ -687,20 +1337,16 @@ def format_vector(vector, *, max_rows=None, min_rows=None, max_columns=None, mas name, keys, vals, - lower_border=has_pandas, + lower_border=True, name=vector.name if mask is None else mask.name, ) - if has_pandas: - df = _get_vector_dataframe(vector, max_rows, min_rows, max_columns, mask=mask) - if vector._size > 0: - if df.columns[0] != "index": - df.columns.name = "index" - df.index = ["value"] - with pd.option_context( - "display.show_dimensions", False, "display.large_repr", "truncate" - ): - df_repr = df.__repr__() - return f"{header}\n{df_repr}" + if vector._size > 0: + frame = _get_vector_dataframe(vector, max_rows, min_rows, max_columns, mask=mask) + if frame.columns[0].label != "index": + # Dense vectors label the corner "index" and the single row "value". + frame.col_name = "index" + frame.index = ["value"] + return f"{header}\n{_render_text(frame)}" return header diff --git a/graphblas/tests/test_formatting.py b/graphblas/tests/test_formatting.py index a6522dcef..5db849ea8 100644 --- a/graphblas/tests/test_formatting.py +++ b/graphblas/tests/test_formatting.py @@ -146,38 +146,24 @@ def t(): def test_no_pandas_repr(A, C, v, w): - # This is a bit of a hack... + # The rich repr is hand-rendered, so it no longer depends on pandas: with + # pandas marked absent the output is byte-identical to the pandas-present + # output, data grid and all. (When pandas is genuinely absent both branches + # use the same pandas-free path, so this still exercises that code.) + objs = [A, A.T, C, C.S, ~C.V, v, v.S, ~w.V, w] + expected = [repr(x) for x in objs] has_pandas_prev = formatting.has_pandas formatting.has_pandas = False try: - repr_printer(A, "A", indent=8) - assert repr(A) == ( - '"A_1" nvals nrows ncols dtype format\n' - "gb.Matrix 3 1 5 INT64 bitmapr" - ) - repr_printer(A.T, "A.T", indent=8) - assert repr(A.T) == ( - '"A_1.T" nvals nrows ncols dtype format\n' - "gb.TransposedMatrix 3 5 1 INT64 bitmapc" - ) - repr_printer(C.S, "C.S", indent=8) - assert repr(C.S) == ( - '"C.S" nvals nrows ncols dtype format\n' - "StructuralMask\n" - "of gb.Matrix 8 70 77 INT64 hypercsr" - ) - repr_printer(v, "v", indent=8) - assert repr(v) == ( - '"v" nvals size dtype format\ngb.Vector 3 5 FP64 bitmap' - ) - repr_printer(~w.V, "~w.V", indent=8) - assert repr(~w.V) == ( - '"~w.V" nvals size dtype format\n' - "ComplementedValueMask\n" - "of gb.Vector 4 77 INT64 bitmap" - ) + actual = [repr(x) for x in objs] finally: formatting.has_pandas = has_pandas_prev + assert actual == expected + # The data grid is rendered, not just the header: the border line and the + # values are present. + lines = repr(A).split("\n") + assert lines[2].startswith("----") + assert lines[-1] == "0 0 1 2" @pytest.mark.skipif("not pd") @@ -517,137 +503,21 @@ def test_scalar_repr(s, t): def test_no_pandas_repr_html(A, C, v, w): - # This is a bit of a hack... + # _repr_html_ is hand-rendered too: marking pandas absent yields output that + # is byte-identical to the pandas-present output, data table included. + objs = [A, A.T, C, C.S, ~C.V, v, v.S, ~w.V, w] + expected = [repr_html(x) for x in objs] has_pandas_prev = formatting.has_pandas formatting.has_pandas = False try: - html_printer(A, "A", indent=8) - assert repr_html(A) == ( - "
" - f"{CSS_STYLE}" - '
A1
\n' - '\n' - " \n" - ' \n' - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - "
gb.Matrix
nvals
nrows
ncols
dtype
format
315INT64bitmapr
\n" - "
\n" - "(Install pandas to see a preview of the data)
" - ) - html_printer(A.T, "A.T", indent=8) - assert repr_html(A.T) == ( - "
" - f"{CSS_STYLE}" - '
A1.T
\n' - '\n' - " \n" - ' \n' - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - "
gb.TransposedMatrix
nvals
nrows
ncols
dtype
format
351INT64bitmapc
\n" - "
\n" - "(Install pandas to see a preview of the data)
" - ) - html_printer(C.S, "C.S", indent=8) - assert repr_html(C.S) == ( - "
" - f"{CSS_STYLE}" - '
C.S
\n' - '\n' - " \n" - ' \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - "
StructuralMask\n'
-            "of\n"
-            "gb.Matrix
nvals
nrows
ncols
dtype
format
87077INT64hypercsr
\n" - "
\n" - "(Install pandas to see a preview of the data)
" - ) - html_printer(v, "v", indent=8) - assert repr_html(v) == ( - "
" - f"{CSS_STYLE}" - '
v
\n' - '\n' - " \n" - ' \n' - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - "
gb.Vector
nvals
size
dtype
format
35FP64bitmap
\n" - "
\n" - "(Install pandas to see a preview of the data)
" - ) - html_printer(~w.V, "~w.V", indent=8) - assert repr_html(~w.V) == ( - "
" - f"{CSS_STYLE}" - '
~w.V
\n' - '\n' - " \n" - ' \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - " \n" - "
ComplementedValueMask\n'
-            "of\n"
-            "gb.Vector
nvals
size
dtype
format
477INT64bitmap
\n" - "
\n" - "(Install pandas to see a preview of the data)
" - ) + actual = [repr_html(x) for x in objs] finally: formatting.has_pandas = has_pandas_prev + assert actual == expected + # The data table is rendered, not the "install pandas" placeholder. + html = repr_html(A) + assert "install" not in html.lower() + assert '' in html @pytest.mark.skipif("not pd") From 8f391d1ea84e5f27d51b27bf00bddf98ad7439ba Mon Sep 17 00:00:00 2001 From: Erik Welch Date: Fri, 31 Jul 2026 14:34:59 -0500 Subject: [PATCH 2/3] Un-gate formatting tests that no longer need pandas The rich repr renders without pandas as of the previous commit, so 29 of the 37 skipif("not pd") gates in test_formatting.py are obsolete; in a pandas-free environment those tests now run instead of skipping (12 -> 41 passing). The 8 tests exercising pd.option_context stay gated, as do the three autocompute tests with their own runtime skip. Several un-gated tests assert the full pandas-style HTML verbatim and pass without pandas, confirming the hand renderer is byte-identical. Verified with pandas installed (unchanged: the removed gates never fired there) and with pandas blocked at sys.meta_path, where test_formatting.py goes from 12 passed / 40 skipped to 41 passed / 11 skipped, with no failures. --- graphblas/tests/test_formatting.py | 29 ----------------------------- 1 file changed, 29 deletions(-) diff --git a/graphblas/tests/test_formatting.py b/graphblas/tests/test_formatting.py index 5db849ea8..62c722756 100644 --- a/graphblas/tests/test_formatting.py +++ b/graphblas/tests/test_formatting.py @@ -166,7 +166,6 @@ def test_no_pandas_repr(A, C, v, w): assert lines[-1] == "0 0 1 2" -@pytest.mark.skipif("not pd") def test_matrix_repr_small(A, B): repr_printer(A, "A") assert repr(A) == ( @@ -198,7 +197,6 @@ def test_matrix_repr_small(A, B): ) -@pytest.mark.skipif("not pd") def test_matrix_mask_repr_small(A): repr_printer(A.S, "A.S") assert repr(A.S) == ( @@ -390,7 +388,6 @@ def test_matrix_mask_repr_large(C): ) -@pytest.mark.skipif("not pd") def test_vector_repr_small(v): repr_printer(v, "v") assert repr(v) == ( @@ -415,7 +412,6 @@ def test_vector_repr_large(w): ) -@pytest.mark.skipif("not pd") def test_vector_mask_repr_small(v): repr_printer(v.S, "v.S") assert repr(v.S) == ( @@ -520,7 +516,6 @@ def test_no_pandas_repr_html(A, C, v, w): assert '
' in html -@pytest.mark.skipif("not pd") def test_matrix_repr_html_small(A, B): html_printer(A, "A") assert repr_html(A) == ( @@ -715,7 +710,6 @@ def test_matrix_repr_html_small(A, B): ) -@pytest.mark.skipif("not pd") def test_matrix_mask_repr_html_small(A): html_printer(A.S, "A.S") assert repr_html(A.S) == ( @@ -1991,7 +1985,6 @@ def test_matrix_mask_repr_html_large(C): ) -@pytest.mark.skipif("not pd") def test_vector_repr_html_small(v): html_printer(v, "v") assert repr_html(v) == ( @@ -2137,7 +2130,6 @@ def test_vector_repr_html_large(w): ) -@pytest.mark.skipif("not pd") def test_vector_mask_repr_html_small(v): html_printer(v.S, "v.S") assert repr_html(v.S) == ( @@ -2759,7 +2751,6 @@ def test_apply_repr(v): ) -@pytest.mark.skipif("not pd") def test_apply_repr_html(v): html_printer(v.apply(unary.one), "v.apply(unary.one)") assert repr_html(v.apply(unary.one)) == ( @@ -2792,7 +2783,6 @@ def test_mxm_repr(A, B): ) -@pytest.mark.skipif("not pd") def test_mxm_repr_html(A, B): html_printer(A.mxm(B), "A.mxm(B)") assert repr_html(A.mxm(B)) == ( @@ -2827,7 +2817,6 @@ def test_mxv_repr(A, v): ) -@pytest.mark.skipif("not pd") def test_mxv_repr_html(A, v): html_printer(A.mxv(v), "A.mxv(v)") assert repr_html(A.mxv(v)) == ( @@ -2850,7 +2839,6 @@ def test_mxv_repr_html(A, v): ) -@pytest.mark.skipif("not pd") def test_matrix_reduce_columns_repr_html(A): # This is implemented using the transpose of A, so make sure we're oriented correctly! html_printer(A.reduce_columnwise(), "A.reduce_columnwise()") @@ -2884,7 +2872,6 @@ def test_matrix_reduce_repr(C, v): ) -@pytest.mark.skipif("not pd") def test_matrix_reduce_repr_html(C, v): html_printer(C.reduce_scalar(), "C.reduce_scalar()", indent=8) assert repr_html(C.reduce_scalar()) == ( @@ -2905,7 +2892,6 @@ def test_matrix_reduce_repr_html(C, v): ) -@pytest.mark.skipif("not pd") def test_matrix_huge(): M = Matrix(int, nrows=2**60, ncols=2**60, name="M") repr_printer(M, "M") @@ -2931,7 +2917,6 @@ def test_matrix_huge(): assert M.isequal(M2) -@pytest.mark.skipif("not pd") def test_matrix_huge_html(): M = Matrix(int, nrows=2**60, ncols=2**60, name="M") html_printer(M, "M") @@ -3124,7 +3109,6 @@ def test_matrix_huge_html(): ) -@pytest.mark.skipif("not pd") def test_vector_huge(): v = Vector(int, size=2**60) repr_printer(v, "v") @@ -3139,7 +3123,6 @@ def test_vector_huge(): assert v2.isequal(v) -@pytest.mark.skipif("not pd") def test_vector_huge_html(): v = Vector(int, size=2**60) html_printer(v, "v") @@ -3277,7 +3260,6 @@ def test_vector_huge_html(): ) -@pytest.mark.skipif("not pd") def test_sparse_vector_repr(): v = Vector.from_coo([100 * i for i in range(100)], [10 * i for i in range(100)], name="v") repr_printer(v, "v") @@ -3431,7 +3413,6 @@ def test_sparse_vector_repr(): ) -@pytest.mark.skipif("not pd") def test_sparse_matrix_repr(): A = Matrix.from_coo( [100 * i for i in range(100)], [10 * i for i in range(100)], list(range(100)), name="A" @@ -3605,7 +3586,6 @@ def test_sparse_matrix_repr(): ) -@pytest.mark.skipif("not pd") def test_infix_expr_repr_html(A, B, v): html_printer(v & v, "v & v") assert repr_html(v & v) == ( @@ -3804,7 +3784,6 @@ def test_infix_expr_repr_html(A, B, v): ) -@pytest.mark.skipif("not pd") def test_infix_expr_repr(A, B, v): repr_printer(v & v, "v & v") assert repr(v & v) == ( @@ -3880,7 +3859,6 @@ def test_infix_expr_repr(A, B, v): ) -@pytest.mark.skipif("not pd") def test_inner_outer_repr_html(v): html_printer(v.inner(v), "v.inner(v)") assert repr_html(v.inner(v)) == ( @@ -3922,7 +3900,6 @@ def test_inner_outer_repr_html(v): ) -@pytest.mark.skipif("not pd") def test_inner_outer_repr(v): # XXX: hmm, having `(GrB_Matrix)` here isn't so pretty repr_printer(v.inner(v), "v.inner(v)") @@ -4348,7 +4325,6 @@ def test_autocompute_html(A, B, v): ) -@pytest.mark.skipif("not pd") def test_display_nan(): v = Vector.from_coo([0, 1], [1.0, np.nan], size=3, name="v") repr_printer(v, "v") @@ -4482,7 +4458,6 @@ def test_display_nan(): ) -@pytest.mark.skipif("not pd") def test_large_iso(): A = Matrix(int, nrows=2**60, ncols=2**60) A[:, :] << 1 @@ -4711,7 +4686,6 @@ def test_index_expr_matrix_html(A): ) -@pytest.mark.skipif("not pd") def test_scalar_as_vector(): s = Scalar.from_value(5, is_cscalar=False) # pragma: is_grbscalar v = s._as_vector() @@ -4825,7 +4799,6 @@ def test_index_expr_autocompute(v): ) -@pytest.mark.skipif("not pd") def test_udt(): record_dtype = np.dtype([("x", np.bool_), ("y", np.int64)], align=True) udt = dtypes.register_anonymous(record_dtype, "record_dtype") @@ -4875,7 +4848,6 @@ def test_udt(): ) -@pytest.mark.skipif("not pd") def test_empty(): v = Vector(int, 0) repr_printer(v, "v") @@ -4900,7 +4872,6 @@ def test_empty(): ) -@pytest.mark.skipif("not pd") def test_vector_as_matrix(): v = Vector.from_coo([1], [2], name="v_A") A = v._as_matrix() From 730498f7e7f6900dc5397af6a1b3baf5bdd0d0f8 Mon Sep 17 00:00:00 2001 From: Erik Welch Date: Fri, 31 Jul 2026 14:34:59 -0500 Subject: [PATCH 3/3] Un-gate three pandas-free autocompute tests in test_formatting.py Since the pandas-free repr, test_autocompute, test_autocompute_html, and test_index_expr_autocompute only assert repr/_repr_html_ output that renders byte-identically without pandas, so the runtime "if not pd: skip" guards are obsolete. Removed them. The eight large-repr tests that call pd.option_context stay gated (they genuinely need pandas to exercise user display options). Verified with pandas installed (unchanged: the removed guards never fired there) and with pandas blocked at sys.meta_path, where test_formatting.py goes from 41 passed / 11 skipped to 44 passed / 8 skipped; the three moved from skip to pass, with no failures. --- graphblas/tests/test_formatting.py | 6 ------ 1 file changed, 6 deletions(-) diff --git a/graphblas/tests/test_formatting.py b/graphblas/tests/test_formatting.py index 62c722756..264dc0502 100644 --- a/graphblas/tests/test_formatting.py +++ b/graphblas/tests/test_formatting.py @@ -3920,8 +3920,6 @@ def test_inner_outer_repr(v): @autocompute def test_autocompute(A, B, v): - if not pd: # pragma: no cover (import) - pytest.skip("needs pandas") repr_printer(A & A, "A & A") assert repr(A & A) == ( "gb.MatrixEwiseMultExpr nrows ncols left_dtype right_dtype\n" @@ -4002,8 +4000,6 @@ def test_autocompute(A, B, v): @autocompute def test_autocompute_html(A, B, v): - if not pd: # pragma: no cover (import) - pytest.skip("needs pandas") html_printer(A & A, "A & A") assert repr_html(A & A) == ( "
" @@ -4744,8 +4740,6 @@ def test_scalar_as_vector(): @autocompute def test_index_expr_autocompute(v): - if not pd: # pragma: no cover (import) - pytest.skip("needs pandas") html_printer(v[[0, 1]], "v[[0, 1]]") assert repr_html(v[[0, 1]]) == ( "
"