Module gpu-transform-camera

Model isha-org:gpu-transform-camera:gpu-transform

Module to redirect transforms from cameras on the vision service to use GPU CUDA instead of CPU to improve processing speed and efficiency.

├── Makefile
├── module.tar.gz
├── reload.sh
├── run.sh
├── meta.json
├── README.md
├── requirements.txt
├── setup.sh
|-- build.sh
├── src
│   ├── __init__.py
│   ├── camera_module.py
│   ├── config.py
│   ├── transform_pipeline.py
│   └── utils
│       ├── __init__.py
│       └── gpu_utils.py
└── tests
    └── test_transform_pipeline.py

Run tests through pytest with:

 pytest tests/ -v 

---

Features

• GPU-accelerated transforms: Uses CUDA to speed up image processing from camera feeds.
• Modular pipeline: Easily extend or modify transform steps.
• Test suite: Includes integration and unit tests for reliability.

---

Installation

For standard users (x86/linux/mac)

Create venv and install using

make install

For NVIDIA Jetson or Jetpack6 Users

Create venv and install using

make install-jetpack-

This sets up the Python venv and installs all dependencies for use on jetpack using pytorch wheels

Ensure you are using Python version 3.10 or above for jetpack.

Usage

You can run the main pipeline or integrate the module into your own code.

Example:

from src.transform_pipeline import GPUTransformPipeline

config = [
    {
        "type": "resize",
        "attributes": {
            "width_px": 224,
            "height_px": 224
        }
    },
    {
        "type": "normalize",
        "attributes": {
            "mean": [0.485, 0.456, 0.406],
            "std": [0.229, 0.224, 0.225]
        }
    }
]

pipeline = GPUTransformPipeline(config)

Or run scripts directly:

python src/main.py

---

Running Tests

Tests are located in src/tests/.
To run all tests:

pytest src/tests/ -v

---

Transform Pipeline Configuration

The transform pipeline is configured with a list of transform steps. Each step is a dictionary with a type and an attributes dictionary.

Supported Transform Types and Attributes

Transform Type	Attribute Name(s)	Type	Required	Description
resize	width_px, height_px	int	Yes	Output width and height in pixels
normalize	mean, std	list	Yes	Mean and std lists for normalization
grayscale	(none)		-	Converts image to grayscale
rotate	angle_degs	int/float	Yes	Angle in degrees to rotate the image
crop	x_min_px, y_min_px, x_max_px, y_max_px, overlay_crop_box	float, float, float, float, bool	Yes, Yes, Yes, Yes, No	Crop box coordinates and optional overlay flag

Example Configuration

[
  {
    "type": "resize",
    "attributes": {
      "width_px": 224,
      "height_px": 224
    }
  },
  {
    "type": "normalize",
    "attributes": {
      "mean": [0.485, 0.456, 0.406],
      "std": [0.229, 0.224, 0.225]
    }
  },
  {
    "type": "rotate",
    "attributes": {
      "angle_degs": 90
    }
  },
  {
    "type": "crop",
    "attributes": {
      "x_min_px": 10.0,
      "y_min_px": 20.0,
      "x_max_px": 200.0,
      "y_max_px": 220.0,
      "overlay_crop_box": false
    }
  },
  {
    "type": "grayscale",
    "attributes": {}
  },
  {
    "type": "to_tensor",
    "attributes": {}
  }
]

---

Makefile Commands

You can use the provided Makefile to build and clean the module:

Command	Description
make install	Installs dependencies for standard (x86/Linux/Mac) environments.
make install-jetpack-	Installs JetPack-specific dependencies and custom wheels.
make module	Builds the module and creates module.tar.gz for deployment.
make clean_module	Removes build artifacts, including module.tar.gz.

Example usage:

make module        # Build the module
make clean_module  

## Notes

- CUDA and compatible GPU required for acceleration.
- This module can support either JPEG or png image types from the source camera, but the entire JPEG pipeline is GPU accelerated vs only parts of the png pipeline are GPU supported. 
---