Kornia

Warnings

• gotcha Kornia's keypoint and bounding box operations, while using float coordinates, often assume an underlying pixel-center (index-based) convention (e.g., pixel (3,4) center is at (3.5, 4.5)). This can lead to slight inaccuracies or off-by-one errors if precise sub-pixel float coordinates are expected, potentially affecting model accuracy or geometric transformations.
  Fix: Refer to Kornia's documentation on coordinate conventions for keypoints and bounding boxes. Be mindful of how your input coordinates align with Kornia's assumed convention, especially when integrating with other libraries or performing precise geometric tasks.
• gotcha While `pip install kornia` is the primary installation, functionalities like `kornia.io.load_image` (for robust image I/O) internally rely on `kornia_rs`. If you encounter errors related to image loading, ensure `kornia_rs` is also installed (e.g., `pip install kornia kornia_rs`).
  Fix: Explicitly install `kornia_rs`: `pip install kornia_rs`.
• gotcha When using `kornia.augmentation` for differentiable data augmentation, distinguish between `torch.nn.Parameter` (for parameters that should be optimized/differentiated) and `torch.Tensor` (for static parameters). Incorrect usage can lead to unintended optimization behavior in meta-learning or differentiable augmentation pipelines.
  Fix: Carefully define your augmentation parameters; use `nn.Parameter` for trainable parameters and regular `torch.Tensor` for fixed ones. Consult the differentiable data augmentation tutorials for examples.
• breaking Kornia continuously evolves. Older versions of PyTorch might not be compatible with the latest Kornia releases. Specifically, Kornia 0.8.x requires PyTorch >=2.0.0.
  Fix: Ensure your PyTorch version is up-to-date and compatible with your Kornia version. For Kornia 0.8.x, ensure `torch>=2.0.0` is installed. Check the official Kornia documentation or `pyproject.toml` for precise `torch` version requirements for your specific Kornia version.

Install

• pip install kornia Install stable version

Imports

• kornia

  import kornia as K

  Standard alias for Kornia.
• AugmentationSequential

  from kornia.augmentation import AugmentationSequential

  Most functionalities are within Kornia's submodules, like `augmentation`.
• rgb_to_grayscale

  from kornia.color import rgb_to_grayscale

  Color conversions are found in the `kornia.color` submodule.

Quickstart

This quickstart demonstrates basic Kornia usage: creating a dummy image tensor, converting it to grayscale, and applying a Gaussian blur filter. It highlights common imports and tensor operations. For actual image loading, libraries like Pillow or OpenCV are typically used, followed by `K.image_to_tensor`.

import torch
import kornia as K
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt

# Simulate loading an image or create a dummy tensor
# In a real scenario, replace this with actual image loading (e.g., with Pillow/OpenCV)
# and conversion using K.image_to_tensor.
# Example with a dummy tensor:
image_tensor = torch.rand(1, 3, 256, 256, dtype=torch.float32)

# Convert image to grayscale
grayscale_tensor = K.color.rgb_to_grayscale(image_tensor)

# Apply a Gaussian blur
blurred_tensor = K.filters.gaussian_blur2d(grayscale_tensor, kernel_size=(7, 7), sigma=(1.5, 1.5))

print(f"Original image shape: {image_tensor.shape}")
print(f"Grayscale image shape: {grayscale_tensor.shape}")
print(f"Blurred image shape: {blurred_tensor.shape}")

# --- Optional: Visualization (requires matplotlib and numpy) ---
# Convert tensors to NumPy arrays for display
# You might need to detach from GPU and move to CPU if applicable:
# grayscale_np = K.tensor_to_image(grayscale_tensor.detach().cpu())
# blurred_np = K.tensor_to_image(blurred_tensor.detach().cpu())

# fig, axs = plt.subplots(1, 2, figsize=(10, 5))
# axs[0].imshow(grayscale_np)
# axs[0].set_title('Grayscale Image')
# axs[0].axis('off')
# axs[1].imshow(blurred_np)
# axs[1].set_title('Blurred Image')
# axs[1].axis('off')
# plt.show()