Spaces:
Running
Running
| """Centralized tensor presentation formatting obeying the 16-element budget. | |
| Tensor-value invariant: Wherever the visual grammar claims to display the value | |
| of a quantity, adjoint, gradient contribution, or numerical backward computation, | |
| it displays a representation of the actual numerical value. Shape information is | |
| metadata and must never substitute for a value. | |
| """ | |
| from dataclasses import dataclass | |
| import numpy as np | |
| from number_format import format_num, DEFAULT_FORMAT_CONFIG, NumberFormatConfig | |
| DEFAULT_TENSOR_ELEMENT_BUDGET = 16 | |
| class TensorDisplay: | |
| text: str | |
| shape: tuple[int, ...] | |
| truncated: bool | |
| element_count: int | |
| def _format_vector(arr, truncated=False, config: NumberFormatConfig | None = None): | |
| if not truncated: | |
| items = [format_num(float(x), config=config) for x in arr] | |
| return f"[{', '.join(items)}]" | |
| n = len(arr) | |
| # Symmetric truncation: 2 at start, 1 at end (or 2 at end if length allows) | |
| # E.g. [a0, a1, …, an-1] | |
| first = [format_num(float(x), config=config) for x in arr[:2]] | |
| last = [format_num(float(x), config=config) for x in arr[-1:]] | |
| return f"[{', '.join(first)}, …, {', '.join(last)}]" | |
| def _format_matrix(arr, truncated=False, config: NumberFormatConfig | None = None): | |
| m, n = arr.shape | |
| if not truncated: | |
| rows = [] | |
| for r in arr: | |
| row_items = [format_num(float(x), config=config) for x in r] | |
| rows.append(f"[{', '.join(row_items)}]") | |
| return f"[{', '.join(rows)}]" | |
| # Truncated matrix: preserve beginning and ending rows/columns | |
| # Rows to show: if m > 2, show row 0, row 1, ..., row m-1 | |
| # Cols to show: if n > 2, show col 0, col 1, ..., col n-1 | |
| def format_row(r): | |
| if n > 2: | |
| return f"[{format_num(float(r[0]), config=config)}, {format_num(float(r[1]), config=config)}, …, {format_num(float(r[-1]), config=config)}]" | |
| else: | |
| cols = [format_num(float(x), config=config) for x in r] | |
| return f"[{', '.join(cols)}]" | |
| row_strs = [] | |
| if m <= 2: | |
| for r in arr: | |
| row_strs.append(format_row(r)) | |
| else: | |
| # First row (or two rows if m >= 4) | |
| num_top = 2 if m >= 4 else 1 | |
| for i in range(num_top): | |
| row_strs.append(format_row(arr[i])) | |
| v_ellipsis = ["⋮", "⋮", "⋱", "⋮"] if n > 2 else ["⋮"] * n | |
| row_strs.append(f"[{', '.join(v_ellipsis)}]") | |
| row_strs.append(format_row(arr[-1])) | |
| return f"[{', '.join(row_strs)}]" | |
| def _format_nd(arr, truncated=False, config: NumberFormatConfig | None = None): | |
| if arr.ndim == 1: | |
| return _format_vector(arr, truncated=truncated, config=config) | |
| if arr.ndim == 2: | |
| return _format_matrix(arr, truncated=truncated, config=config) | |
| if not truncated: | |
| sub_strs = [_format_nd(sub, truncated=False, config=config) for sub in arr] | |
| return f"[{', '.join(sub_strs)}]" | |
| else: | |
| if len(arr) <= 2: | |
| sub_strs = [_format_nd(sub, truncated=True, config=config) for sub in arr] | |
| return f"[{', '.join(sub_strs)}]" | |
| sub_first = _format_nd(arr[0], truncated=True, config=config) | |
| sub_last = _format_nd(arr[-1], truncated=True, config=config) | |
| return f"[{sub_first}, …, {sub_last}]" | |
| def make_tensor_display(val, element_budget=DEFAULT_TENSOR_ELEMENT_BUDGET, config: NumberFormatConfig | None = None) -> TensorDisplay: | |
| """Produces structured TensorDisplay distinguishing text representation from shape metadata.""" | |
| if val is None: | |
| return TensorDisplay(text="—", shape=(), truncated=False, element_count=0) | |
| if isinstance(val, (int, float, np.floating, np.integer)): | |
| val_f = float(val) | |
| return TensorDisplay(text=format_num(val_f, config=config), shape=(), truncated=False, element_count=1) | |
| if not isinstance(val, np.ndarray): | |
| val = np.asarray(val, dtype=np.float64) | |
| if val.ndim == 0: | |
| return TensorDisplay(text=format_num(float(val), config=config), shape=(), truncated=False, element_count=1) | |
| size = val.size | |
| shape = val.shape | |
| truncated = size > element_budget | |
| if val.ndim == 1: | |
| text = _format_vector(val, truncated=truncated, config=config) | |
| elif val.ndim == 2: | |
| text = _format_matrix(val, truncated=truncated, config=config) | |
| else: | |
| text = _format_nd(val, truncated=truncated, config=config) | |
| return TensorDisplay(text=text, shape=shape, truncated=truncated, element_count=size) | |
| def format_tensor_value(val, element_budget=DEFAULT_TENSOR_ELEMENT_BUDGET, include_shape_if_truncated=True, config: NumberFormatConfig | None = None) -> str: | |
| """Returns canonical string representation of tensor value. | |
| If truncated and include_shape_if_truncated is True, appends shape metadata. | |
| Shape information is NEVER substituted for the value itself. | |
| """ | |
| disp = make_tensor_display(val, element_budget=element_budget, config=config) | |
| if disp.truncated and include_shape_if_truncated: | |
| return f"{disp.text} shape {disp.shape}" | |
| return disp.text | |
| def format_tensor_shape(val) -> str: | |
| """Format shape metadata only.""" | |
| if val is None: | |
| return "" | |
| if isinstance(val, (int, float, np.floating, np.integer)): | |
| return "()" | |
| arr = np.asarray(val) | |
| return str(arr.shape) | |