Started working on new features, unit testing.

Author: ErnstRoell

.gitignore

@@ -1,3 +1,6 @@
+# Vscode setting.
+.vscode
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

README.md

@@ -1,3 +1,14 @@
+
+# Updates 
+This the development branch of DECT. Currently implementing/changing the following 
+features: 
+- Simplifying the usage, decoupling the torch_geometric style interface. 
+- Significant increase in unit testing. 
+- Adding proper poetry for development.
+- Providing jit compiled versions of all functions -> 20x speedups. 
+- Rendering of Point Clouds using Backprop. 
+
+
 # DECT - Differentiable Euler Characteristic Transform
 [![arXiv](https://img.shields.io/badge/arXiv-2310.07630-b31b1b.svg)](https://arxiv.org/abs/2310.07630) ![GitHub contributors](https://img.shields.io/github/contributors/aidos-lab/dect) ![GitHub](https://img.shields.io/github/license/aidos-lab/dect) [![Maintainability](https://api.codeclimate.com/v1/badges/82f86d7e2f0aae342055/maintainability)](https://codeclimate.com/github/aidos-lab/dect/maintainability)
 

dect/ect.py

@@ -0,0 +1,48 @@
+"""All ECT calculation functions."""
+
+from typing import TypeAlias
+
+import torch
+
+Tensor: TypeAlias = torch.Tensor
+
+
+def compute_ect_point_cloud(
+    x: Tensor,
+    v: Tensor,
+    radius: float,
+    resolution: int,
+    scale: float,
+) -> Tensor:
+    """
+    Computes the ECT of a point cloud.
+
+    Parameters
+    ----------
+    x : Tensor
+        The point cloud of shape [B,N,D] where B is the number of point clouds,
+        N is the number of points and D is the ambient dimension.
+    v : Tensor
+        The tensor of directions of shape [D,N], where D is the ambient
+        dimension and N is the number of directions.
+    radius : float
+        Radius of the interval to discretize the ECT into.
+    resolution : int
+        Number of steps to divide the lin interval into.
+    scale : Tensor
+        The multipicative factor for the sigmoid function.
+
+    Returns
+    -------
+    Tensor
+        The ECT of the point cloud of shape [B,N,R] where B is the number of
+        point clouds (thus ECT's), N is the number of direction and R is the
+        resolution.
+    """
+    lin = torch.linspace(
+        start=-radius, end=radius, steps=resolution, device=x.device
+    ).view(-1, 1, 1)
+    nh = (x @ v).unsqueeze(1)
+    ecc = torch.nn.functional.sigmoid(scale * torch.sub(lin, nh))
+    ect = torch.sum(ecc, dim=2)
+    return ect

dect/ect_compiled.py

@@ -0,0 +1,49 @@
+"""
+Contains all compiled versions, only use them if you are familiar with torch
+compile. Yields significant speed ups.
+"""
+
+from typing import Callable, TypeAlias
+import torch
+
+Tensor: TypeAlias = torch.Tensor
+
+from dect.ect import compute_ect_point_cloud as compute_ect_point_cloud_vanilla
+
+
+@torch.compile
+def compute_ect_point_cloud(
+    x: Tensor,
+    v: Tensor,
+    radius: float,
+    resolution: int,
+    scale: float,
+) -> Tensor:
+    """
+    Compiled function wrapper around the compute_ect_pointcloud.
+
+    Parameters
+    ----------
+    x : Tensor
+        The point cloud of shape [B,N,D] where B is the number of point clouds,
+        N is the number of points and D is the ambient dimension.
+    v : Tensor
+        The tensor of directions of shape [D,N], where D is the ambient
+        dimension and N is the number of directions.
+    radius : float
+        Radius of the interval to discretize the ECT into.
+    resolution : int
+        Number of steps to divide the lin interval into.
+    scale : Tensor
+        The multipicative factor for the sigmoid function.
+
+    Returns
+    -------
+    Tensor
+        The ECT of the point cloud of shape [B,N,R] where B is the number of
+        point clouds (thus ECT's), N is the number of direction and R is the
+        resolution.
+    """
+    return compute_ect_point_cloud_vanilla(
+        x=x, v=v, radius=radius, resolution=resolution, scale=scale
+    )

zz_dect_bu/__init__.py renamed to dect_bu/__init__.py

@@ -1,16 +1,16 @@
 import torch
 
+from dect.ect_compiled import compute_ect_point_cloud as ect_compiled
+from dect.ect import compute_ect_point_cloud
+from dect.directions import generate_uniform_directions
 
-def ect(x, *args, index):
-    print("x", x)
-    for i, arg in enumerate(args):
-        print("arg", (-1) ** (i + 1) * arg)
+v = generate_uniform_directions(num_thetas=10, d=3, seed=10)
 
+x = torch.rand(size=(1, 10, 3))
 
-ect(
-    torch.tensor(0),
-    torch.tensor(4),
-    torch.tensor(1),
-    torch.tensor(2),
-    index=torch.tensor(2),
-)
+
+ect1 = compute_ect_point_cloud(x, v, radius=1, resolution=64, scale=1)
+print(ect1)
+
+ect = ect_compiled(x, v, radius=1, resolution=64, scale=1)
+print(ect)