PyInterpX - Higher Order Interpolation in 3D for Torch

Overview

PyInterpX is a compact library designed for advanced 3D interpolation using higher order polynomial bases, which is not currently supported by PyTorch's torch.nn.functional.interpolate() method. This enhancement allows for more precise and customized interpolation processes in 3D spaces, catering to specialized applications requiring beyond-linear data manipulation.

Quick Start

To get started with PyInterpX:

1. Install the library using pip:

   pip install pyinterpx

2. Import interp from PyInterpX and PyTorch in your script or notebook:

   from pyinterpx.Interpolation import interp
   import torch

3. Utilize the interpolation function with a 6x6x6 kernel, polynomials up to the third power, and 25 channels:

   points, power, channels = 6, 3, 25
   Interp = interp(points, power, channels)
   x = torch.rand(2, 25, 10, 10, 10)
   Interp(x)

Key Features

• Fast: Optimized for high performance across any device.

• CPU and GPU Compatible: Functions seamlessly on both CPU and GPU environments.

  points, power, channels = 6, 3, 25
  # Running on GPU for even faster computations 
  interp = interp(points, power, channels, device="cuda:0")

• Precise: Supports various data types for precise computation.

  points, power, channels = 6, 3, 25
  # Using double for more precision 
  interp = interp(points, power, channels, dtype=torch.double)

• Integrated with PyTorch: Easily integrates within the PyTorch ecosystem.

  # A simple model which uses interpolation at some layer
  class Model(torch.nn.Module):
      def __init__(self):
          super(Model, self).__init__()
          points, power, channels = 6, 3, 25
          # Setting up interpolation 
          self.interpolation = interp(points, power, channels)
  
          self.convs = torch.nn.Sequential(
              torch.nn.Conv3d(25, 64, kernel_size=3, padding=1),
              torch.nn.ReLU(inplace=True),
          )
  
      def forward(self, x):
          x = self.convs(x)
          x = self.interpolation(x)
          return x

• Choose simply the grid alignment you like.

  points, power, channels = 6, 3, 25
  interp = interp(points, power, channels, dtype=torch.double,align_corners = False)

  

  or if you do not want to have any aligment with the input grid

  points, power, channels = 6, 3, 25
  interp = interp(points, power, channels, dtype=torch.double,align_corners = True)

  

• Choose the enhacement factor you like

  factor = 4
  points, power, channels = 6, 3, 25
  interp = interp(points, power, channels, dtype=torch.double,align_corners = False,factor = factor)

  

  factor = 16
  points, power, channels = 6, 3, 25
  interp = interp(points, power, channels, dtype=torch.double,align_corners = False,factor = factor)

  

Prerequisites

Before installing PyInterpX, ensure you meet the following prerequisites:

• Python 3.8 or higher
• pip package manager

License

PyInterpX is open-sourced under the MIT License. For more details, see the LICENSE file.

Contact

For inquiries or support, reach out to Thomas Helfer at thomashelfer@live.de.