multipers 2.2.3__cp310-cp310-win_amd64.whl

multipers/multiparameter_module_approximation.pyx

@@ -0,0 +1,217 @@
+"""!
+@package mma
+@brief Files containing the C++ cythonized functions.
+@author David Loiseaux
+@copyright Copyright (c) 2022 Inria.
+"""
+
+# distutils: language = c++
+
+###########################################################################
+## PYTHON LIBRARIES
+import gudhi as gd
+import numpy as np
+from typing import List
+from joblib import Parallel, delayed
+import sys
+
+###########################################################################
+## CPP CLASSES
+from libc.stdint cimport intptr_t
+from libc.stdint cimport uintptr_t
+
+###########################################################################
+## CYTHON TYPES
+from libcpp.vector cimport vector
+from libcpp.utility cimport pair
+#from libcpp.list cimport list as clist
+from libcpp cimport bool
+from libcpp cimport int
+from typing import Iterable,Optional, Literal
+from cython.operator import dereference
+#########################################################################
+## Multipersistence Module Approximation Classes
+from multipers.mma_structures cimport *
+from multipers.filtrations cimport *
+from multipers.filtration_conversions cimport *
+cimport numpy as cnp
+
+
+#########################################################################
+## Small hack for typing
+from multipers.simplex_tree_multi import is_simplextree_multi, SimplexTreeMulti_type
+from multipers.slicer import Slicer_type, is_slicer
+from multipers._slicer_meta import Slicer
+from multipers.mma_structures import *
+from multipers.mma_structures import PyModule_type
+from typing import Union
+import multipers
+import multipers.io as mio
+from multipers.slicer cimport _multiparameter_module_approximation_f32, _multiparameter_module_approximation_f64
+
+
+
+def module_approximation_from_slicer(
+        slicer:Slicer_type,
+        box:Optional[np.ndarray]=None,
+        max_error=-1,
+        bool complete=True,
+        bool threshold=False,
+        bool verbose=False,
+        list[float] direction = [],
+        )->PyModule_type:
+
+    cdef Module[float] mod_f32
+    cdef Module[double] mod_f64
+    cdef intptr_t ptr
+    if not slicer.is_vine:
+        print(r"Got a non-vine slicer as an input. Use `vineyard=True` to remove this copy.", file=sys.stderr)
+        slicer = Slicer(slicer, vineyard=True)
+    direction_ = np.asarray(direction, dtype=slicer.dtype)
+    if slicer.dtype == np.float32:
+        approx_mod = PyModule_f32()
+        if box is None:
+            box = slicer.filtration_bounds()
+        mod_f32 = _multiparameter_module_approximation_f32(slicer,_py21c_f32(direction_), max_error,Box[float](box),threshold, complete, verbose)
+        ptr = <intptr_t>(&mod_f32)
+    elif slicer.dtype == np.float64:
+        approx_mod = PyModule_f64()
+        if box is None:
+            box = slicer.filtration_bounds()
+        mod_f64 = _multiparameter_module_approximation_f64(slicer,_py21c_f64(direction_), max_error,Box[double](box),threshold, complete, verbose)
+        ptr = <intptr_t>(&mod_f64)
+    else:
+        raise ValueError(f"Slicer must be float-like. Got {slicer.dtype}.")
+
+    approx_mod._set_from_ptr(ptr)
+
+    return approx_mod
+
+def module_approximation(
+        input:Union[SimplexTreeMulti_type,Slicer_type, tuple],
+        box:Optional[np.ndarray]=None,
+        float max_error=-1, 
+        int nlines=500,
+        slicer_backend:Literal["matrix","clement","graph"]="matrix",
+        minpres:Optional[Literal["mpfree"]]=None,
+        degree:Optional[int]=None,
+        bool complete=True, 
+        bool threshold=False, 
+        bool verbose=False,
+        bool ignore_warning=False,
+        id="",
+        list[float] direction = [],
+        list[int] swap_box_coords = [],
+        *,
+        int n_jobs = -1,
+        )->PyModule_type:
+    """Computes an interval module approximation of a multiparameter filtration.
+
+    Parameters
+    ----------
+    input: SimplexTreeMulti or Slicer-like.
+        Holds the multifiltered complex.
+    max_error: positive float
+        Trade-off between approximation and computational complexity.
+        Upper bound of the module approximation, in bottleneck distance, 
+        for interval-decomposable modules.
+    nlines: int = 200
+        Alternative to max_error;
+        specifies the number of persistence computation used for the approximation.
+    box : (Optional) pair of list of floats
+        Defines a rectangle on which to compute the approximation.
+        Format : [x,y], This defines a rectangle on which we draw the lines,
+        uniformly drawn (with a max_error step).
+        The first line is `x`. 
+        **Warning**: For custom boxes, and directions, you **must** ensure
+        that the first line captures a generic barcode.
+    direction: float[:] = []
+        If given, the line are drawn with this angle.
+        **Warning**: You must ensure that the first line, defined by box,
+        captures a generic barcode.
+    slicer_backend: Either "matrix","clement", or "graph".
+        If a simplextree is given, it is first converted to this structure,
+        with different choices of backends.
+    minpres: (Optional) "mpfree" only for the moment.
+        If given, and the input is a simplextree, 
+        computes a minimal presentation before starting the computation.
+        A degree has to be given.
+    degree: int Only required when minpres is given.
+        Homological degree of the minimal degree.
+    threshold: bool
+        When true, intersects the module support with the box,
+        i.e. no more infinite summands.
+    verbose: bool
+        Prints C++ infos.
+    ignore_warning : bool
+        Unless set to true, prevents computing on more than 10k lines. 
+        Useful to prevent a segmentation fault due to "infinite" recursion.
+    Returns
+    -------
+    PyModule
+        An interval decomposable module approximation of the module defined by the
+        homology of this multi-filtration.
+    """
+    if isinstance(input, tuple) or isinstance(input, list):
+        assert all(s.is_minpres for s in input), "Modules cannot be merged unless they are minimal presentations."
+        assert np.unique([s.minpres_degree for s in input]).shape[0] == len(input), "Multiple modules are at the same degree, cannot merge modules" 
+        if len(input) == 0:
+            return PyModule_f64()
+        if n_jobs <= 1: 
+            modules = tuple(module_approximation(slicer, box, max_error, nlines, slicer_backend, minpres, degree, complete, threshold, verbose, ignore_warning, id, direction, swap_box_coords) for slicer in input)
+        else:
+            modules = tuple(Parallel(n_jobs=n_jobs, prefer="threads")(
+                delayed(module_approximation)(slicer, box, max_error, nlines, slicer_backend, minpres, degree, complete, threshold, verbose, ignore_warning, id, direction, swap_box_coords)
+                for slicer in input
+            ))
+        mod = PyModule_f64().set_box(PyBox_f64(*modules[0].get_box()))
+        for i,m in enumerate(modules):
+            mod.merge(m, input[i].minpres_degree)
+        return mod
+    if box is None:
+        if is_simplextree_multi(input):
+            box = input.filtration_bounds()
+        else:
+            box = input.filtration_bounds()
+    box = np.asarray(box)
+
+    # empty coords
+    zero_idx = box[1] == box[0]
+    if np.any(zero_idx):
+        box[1] += zero_idx
+
+    for i in swap_box_coords:
+        box[0,i], box[1,i] = box[1,i], box[0,i]
+    num_parameters = box.shape[1]
+    if num_parameters <=0:
+        num_parameters = box.shape[1]
+    assert len(direction) == 0 or len(direction) == len(box[0]), f"Invalid line direction, has to be 0 or {num_parameters=}"
+
+    prod = sum(np.abs(box[1] - box[0])[:i].prod() * np.abs(box[1] - box[0])[i+1:].prod() for i in range(0,num_parameters))
+
+    if max_error <= 0:
+        max_error = (prod/nlines)**(1/(num_parameters-1))
+
+    if not ignore_warning and prod >= 10_000:
+        raise ValueError(f"""
+Warning : the number of lines (around {np.round(prod)}) may be too high. 
+Try to increase the precision parameter, or set `ignore_warning=True` to compute this module. 
+Returning the trivial module."""
+        )
+    if is_simplextree_multi(input):
+        input = multipers.Slicer(input,backend=slicer_backend, vineyard=True)
+    assert is_slicer(input), "First argument must be a simplextree or a slicer !"
+    return module_approximation_from_slicer(
+            slicer=input,
+            box=box,
+            max_error=max_error,
+            complete=complete,
+            threshold=threshold,
+            verbose=verbose,
+            direction=direction,
+            )
+
+
+
+
+

multipers/pickle.py

@@ -0,0 +1,53 @@
+import numpy as np
+
+def save_with_axis(path:str, signed_measures):
+	np.savez(path, 
+		**{f"{i}_{axis}_{degree}":np.c_[sm_of_degree[0],sm_of_degree[1][:,np.newaxis]] for i,sm in enumerate(signed_measures) for axis,sm_of_axis in enumerate(sm) for degree,sm_of_degree in enumerate(sm_of_axis)},
+	)
+
+def save_without_axis(path:str, signed_measures):
+    np.savez(path, 
+        **{f"{i}_{degree}":np.c_[sm_of_degree[0],sm_of_degree[1][:,np.newaxis]] for i,sm in enumerate(signed_measures) for degree,sm_of_degree in enumerate(sm)},
+    )
+
+def get_sm_with_axis(sms,idx,axis,degree):
+    sm = sms[f"{idx}_{axis}_{degree}"]
+    return (sm[:,:-1],sm[:,-1])
+def get_sm_without_axis(sms,idx,degree):
+    sm = sms[f"{idx}_{degree}"]
+    return (sm[:,:-1],sm[:,-1])
+
+
+def load_without_axis(sms):
+	indices = np.array([[int(i) for i in key.split('_')] for key in sms.keys()], dtype=int)
+	num_data,num_degrees = indices.max(axis=0)+1
+	signed_measures_reconstructed = [[get_sm_without_axis(sms,idx,degree) for degree in range(num_degrees)] for idx in range(num_data)]
+	return signed_measures_reconstructed
+# test : np.all([np.array_equal(a[0],b[0]) and np.array_equal(a[1],b[1]) and len(a) == len(b) == 2 for x,y in zip(signed_measures_reconstructed,signed_measures_reconstructed) for a,b in zip(x,y)])
+
+def load_with_axis(sms):
+	indices = np.array([[int(i) for i in key.split('_')] for key in sms.keys()], dtype=int)
+	num_data,num_axis,num_degrees = indices.max(axis=0)+1
+	signed_measures_reconstructed = [[[get_sm_with_axis(sms,idx,axis,degree) for degree in range(num_degrees)] for axis in range(num_axis)] for idx in range(num_data)]
+	return signed_measures_reconstructed
+
+def save(path:str, signed_measures):
+	if isinstance(signed_measures[0][0], tuple):
+		save_without_axis(path=path,signed_measures=signed_measures)
+	else:
+		save_with_axis(path=path,signed_measures=signed_measures)
+
+def load(path:str):
+	sms = np.load(path)
+	item=None
+	for i in sms.keys():
+		item=i
+		break
+	n = len(item.split('_'))
+	match n:
+		case 2:
+			return load_without_axis(sms)
+		case 3:
+			return load_with_axis(sms)
+		case _:
+			raise Exception("Invalid Signed Measure !")

multipers/plots.py

@@ -0,0 +1,334 @@
+from typing import Optional
+
+import matplotlib.pyplot as plt
+import numpy as np
+
+
+def _plot_rectangle(rectangle: np.ndarray, weight, **plt_kwargs):
+    rectangle = np.asarray(rectangle)
+    x_axis = rectangle[[0, 2]]
+    y_axis = rectangle[[1, 3]]
+    color = "blue" if weight > 0 else "red"
+    plt.plot(x_axis, y_axis, c=color, **plt_kwargs)
+
+
+def _plot_signed_measure_2(
+    pts, weights, temp_alpha=0.7, threshold=(np.inf, np.inf), **plt_kwargs
+):
+    import matplotlib.colors
+
+    pts = np.clip(pts, a_min=-np.inf, a_max=np.asarray(threshold)[None, :])
+    weights = np.asarray(weights)
+    color_weights = np.array(weights, dtype=float)
+    neg_idx = weights < 0
+    pos_idx = weights > 0
+    if np.any(neg_idx):
+        current_weights = -weights[neg_idx]
+        min_weight = np.max(current_weights)
+        color_weights[neg_idx] /= min_weight
+        color_weights[neg_idx] -= 1
+    else:
+        min_weight = 0
+
+    if np.any(pos_idx):
+        current_weights = weights[pos_idx]
+        max_weight = np.max(current_weights)
+        color_weights[pos_idx] /= max_weight
+        color_weights[pos_idx] += 1
+    else:
+        max_weight = 1
+
+    bordeaux = np.array([0.70567316, 0.01555616, 0.15023281, 1])
+    light_bordeaux = np.array([0.70567316, 0.01555616, 0.15023281, temp_alpha])
+    bleu = np.array([0.2298057, 0.29871797, 0.75368315, 1])
+    light_bleu = np.array([0.2298057, 0.29871797, 0.75368315, temp_alpha])
+    norm = plt.Normalize(-2, 2)
+    cmap = matplotlib.colors.LinearSegmentedColormap.from_list(
+        "", [bordeaux, light_bordeaux, "white", light_bleu, bleu]
+    )
+    plt.scatter(
+        pts[:, 0], pts[:, 1], c=color_weights, cmap=cmap, norm=norm, **plt_kwargs
+    )
+    plt.scatter([], [], color=bleu, label="positive mass", **plt_kwargs)
+    plt.scatter([], [], color=bordeaux, label="negative mass", **plt_kwargs)
+    plt.legend()
+
+
+def _plot_signed_measure_4(
+    pts,
+    weights,
+    x_smoothing: float = 1,
+    area_alpha: bool = True,
+    threshold=(np.inf, np.inf),
+    alpha=None,
+    **plt_kwargs,  # ignored ftm
+):
+    # compute the maximal rectangle area
+    pts = np.clip(pts, a_min=-np.inf, a_max=np.array((*threshold, *threshold))[None, :])
+    alpha_rescaling = 0
+    for rectangle, weight in zip(pts, weights):
+        if rectangle[2] > x_smoothing * rectangle[0]:
+            alpha_rescaling = max(
+                alpha_rescaling,
+                (rectangle[2] / x_smoothing - rectangle[0])
+                * (rectangle[3] - rectangle[1]),
+            )
+    # draw the rectangles
+    for rectangle, weight in zip(pts, weights):
+        # draw only the rectangles that have not been reduced to the empty set
+        if rectangle[2] > x_smoothing * rectangle[0]:
+            # make the alpha channel proportional to the rectangle's area
+            if area_alpha:
+                _plot_rectangle(
+                    rectangle=[
+                        rectangle[0],
+                        rectangle[1],
+                        rectangle[2] / x_smoothing,
+                        rectangle[3],
+                    ],
+                    weight=weight,
+                    alpha=(
+                        (rectangle[2] / x_smoothing - rectangle[0])
+                        * (rectangle[3] - rectangle[1])
+                        / alpha_rescaling
+                        if alpha is None
+                        else alpha
+                    ),
+                    **plt_kwargs,
+                )
+            else:
+                _plot_rectangle(
+                    rectangle=[
+                        rectangle[0],
+                        rectangle[1],
+                        rectangle[2] / x_smoothing,
+                        rectangle[3],
+                    ],
+                    weight=weight,
+                    alpha=1 if alpha is None else alpha,
+                    **plt_kwargs,
+                )
+
+
+def plot_signed_measure(signed_measure, threshold=None, ax=None, **plt_kwargs):
+    if ax is None:
+        ax = plt.gca()
+    else:
+        plt.sca(ax)
+    pts, weights = signed_measure
+    pts = np.asarray(pts)
+    num_pts = pts.shape[0]
+    num_parameters = pts.shape[1]
+    if threshold is None:
+        if num_pts == 0:
+            threshold = (np.inf, np.inf)
+        else:
+            if num_parameters == 4:
+                pts_ = np.concatenate([pts[:, :2], pts[:, 2:]], axis=0)
+            else:
+                pts_ = pts
+            threshold = np.max(np.ma.masked_invalid(pts_), axis=0)
+    if isinstance(pts, np.ndarray):
+        pass
+    else:
+        import torch
+
+        if isinstance(pts, torch.Tensor):
+            pts = pts.detach().numpy()
+        else:
+            raise Exception("Invalid measure type.")
+
+    assert num_parameters in (2, 4)
+    if num_parameters == 2:
+        _plot_signed_measure_2(
+            pts=pts, weights=weights, threshold=threshold, **plt_kwargs
+        )
+    else:
+        _plot_signed_measure_4(
+            pts=pts, weights=weights, threshold=threshold, **plt_kwargs
+        )
+
+
+def plot_signed_measures(signed_measures, threshold=None, size=4):
+    num_degrees = len(signed_measures)
+    if num_degrees <= 1:
+        axes = [plt.gca()]
+    else:
+        fig, axes = plt.subplots(
+            nrows=1, ncols=num_degrees, figsize=(num_degrees * size, size)
+        )
+    for ax, signed_measure in zip(axes, signed_measures):
+        plot_signed_measure(signed_measure=signed_measure, ax=ax, threshold=threshold)
+    plt.tight_layout()
+
+
+def plot_surface(
+    grid,
+    hf,
+    fig=None,
+    ax=None,
+    cmap: Optional[str] = None,
+    discrete_surface=False,
+    has_negative_values=False,
+    **plt_args,
+):
+    import matplotlib
+
+    if ax is None:
+        ax = plt.gca()
+    else:
+        plt.sca(ax)
+    if hf.ndim == 3 and hf.shape[0] == 1:
+        hf = hf[0]
+    assert hf.ndim == 2, "Can only plot a 2d surface"
+    fig = plt.gcf() if fig is None else fig
+    if cmap is None:
+        if discrete_surface:
+            cmap = matplotlib.colormaps["gray_r"]
+        else:
+            cmap = matplotlib.colormaps["plasma"]
+    if discrete_surface:
+        if has_negative_values:
+            bounds = np.arange(-5, 6, 1, dtype=int)
+        else:
+            bounds = np.arange(0, 11, 1, dtype=int)
+        norm = matplotlib.colors.BoundaryNorm(bounds, cmap.N, extend="max")
+        im = ax.pcolormesh(grid[0], grid[1], hf.T, cmap=cmap, norm=norm, **plt_args)
+        cbar = fig.colorbar(
+            matplotlib.cm.ScalarMappable(cmap=cmap, norm=norm),
+            spacing="proportional",
+            ax=ax,
+        )
+        cbar.set_ticks(ticks=bounds, labels=bounds)
+        return im
+    im = ax.pcolormesh(grid[0], grid[1], hf.T, cmap=cmap, **plt_args)
+    return im
+
+
+def plot_surfaces(HF, size=4, **plt_args):
+    grid, hf = HF
+    assert (
+        hf.ndim == 3
+    ), f"Found hf.shape = {hf.shape}, expected ndim = 3 : degree, 2-parameter surface."
+    num_degrees = hf.shape[0]
+    fig, axes = plt.subplots(
+        nrows=1, ncols=num_degrees, figsize=(num_degrees * size, size)
+    )
+    if num_degrees == 1:
+        axes = [axes]
+    for ax, hf_of_degree in zip(axes, hf):
+        plot_surface(grid=grid, hf=hf_of_degree, fig=fig, ax=ax, **plt_args)
+    plt.tight_layout()
+
+
+def _rectangle(x, y, color, alpha):
+    """
+    Defines a rectangle patch in the format {z | x  ≤ z ≤ y} with color and alpha
+    """
+    from matplotlib.patches import Rectangle as RectanglePatch
+
+    return RectanglePatch(
+        x, max(y[0] - x[0], 0), max(y[1] - x[1], 0), color=color, alpha=alpha
+    )
+
+
+def _d_inf(a, b):
+    if type(a) != np.ndarray or type(b) != np.ndarray:
+        a = np.array(a)
+        b = np.array(b)
+    return np.min(np.abs(b - a))
+
+
+def plot2d_PyModule(
+    corners,
+    box,
+    *,
+    dimension=-1,
+    separated=False,
+    min_persistence=0,
+    alpha=1,
+    verbose=False,
+    save=False,
+    dpi=200,
+    shapely=True,
+    xlabel=None,
+    ylabel=None,
+    cmap=None,
+):
+    import matplotlib
+
+    try:
+        from shapely import union_all
+        from shapely.geometry import Polygon as _Polygon
+        from shapely.geometry import box as _rectangle_box
+
+        shapely = True and shapely
+    except ImportError:
+        from warnings import warn
+
+        shapely = False
+        warn(
+            "Shapely not installed. Fallbacking to matplotlib. The plots may be inacurate."
+        )
+    cmap = (
+        matplotlib.colormaps["Spectral"] if cmap is None else matplotlib.colormaps[cmap]
+    )
+    box = list(box)
+    if not (separated):
+        # fig, ax = plt.subplots()
+        ax = plt.gca()
+        ax.set(xlim=[box[0][0], box[1][0]], ylim=[box[0][1], box[1][1]])
+    n_summands = len(corners)
+    for i in range(n_summands):
+        trivial_summand = True
+        list_of_rect = []
+        for birth in corners[i][0]:
+            if len(birth) == 1:
+                birth = np.asarray([birth[0]] * 2)
+            birth = np.asarray(birth).clip(min=box[0])
+            for death in corners[i][1]:
+                if len(death) == 1:
+                    death = np.asarray([death[0]] * 2)
+                death = np.asarray(death).clip(max=box[1])
+                if death[1] > birth[1] and death[0] > birth[0]:
+                    if trivial_summand and _d_inf(birth, death) > min_persistence:
+                        trivial_summand = False
+                    if shapely:
+                        list_of_rect.append(
+                            _rectangle_box(birth[0], birth[1], death[0], death[1])
+                        )
+                    else:
+                        list_of_rect.append(
+                            _rectangle(birth, death, cmap(i / n_summands), alpha)
+                        )
+        if not (trivial_summand):
+            if separated:
+                fig, ax = plt.subplots()
+                ax.set(xlim=[box[0][0], box[1][0]], ylim=[box[0][1], box[1][1]])
+            if shapely:
+                summand_shape = union_all(list_of_rect)
+                if type(summand_shape) is _Polygon:
+                    xs, ys = summand_shape.exterior.xy
+                    ax.fill(xs, ys, alpha=alpha, fc=cmap(i / n_summands), ec="None")
+                else:
+                    for polygon in summand_shape.geoms:
+                        xs, ys = polygon.exterior.xy
+                        ax.fill(xs, ys, alpha=alpha, fc=cmap(i / n_summands), ec="None")
+            else:
+                for rectangle in list_of_rect:
+                    ax.add_patch(rectangle)
+            if separated:
+                if xlabel:
+                    plt.xlabel(xlabel)
+                if ylabel:
+                    plt.ylabel(ylabel)
+                if dimension >= 0:
+                    plt.title(rf"$H_{dimension}$ $2$-persistence")
+    if not (separated):
+        if xlabel is not None:
+            plt.xlabel(xlabel)
+        if ylabel is not None:
+            plt.ylabel(ylabel)
+        if dimension >= 0:
+            plt.title(rf"$H_{dimension}$ $2$-persistence")
+    return