feat: add torchscript runtime and docker support

[CuriousDolphin]

Add Torchscript runtime, docker images and devcontainers.

Context

This pull request introduces comprehensive support for TorchScript runtime alongside the existing ONNX runtime, and adds devcontainer and dockerfile configurations for different development environments. These changes aim to provide more flexibility in model execution and simplify the development environment setup process.

KEY CHANGES

• Add devcontainer and docker configurations for multiple development environments:
  • CPU-only configuration with OnnxRuntime
  • GPU configuration with OnnxRuntime
  • GPU configuration with Torchscript support
  • GPU configuration with TensorRT support
• Enhance runtime system with support for multiple model formats:
  • Add TorchScript runtime implementation
  • Refactor ONNXRuntime implementation
  • Introduce BaseRuntime abstract class
• Update dependency management:
  • Split GPU dependencies into separate extras (onnx, torch, tensorrt)
  • Add new optional dependencies for different runtime configurations
• Improve documentation and examples:
  • Update README with new installation options
  • Add detailed runtime configuration documentation
  • Update playground notebook with new runtime examples
• fix CI Matrix tests:
  • now CI tests work also on windows and macos

Impact

• Development Environment: Affects how developers set up and interact with the development environment
• Runtime Performance: Impacts model inference performance through new runtime options
• Dependency Management: Changes how users install and manage SDK dependencies
• Documentation: Updates how users understand and implement the SDK features

Technical details

• Devcontainer Configuration: Implemented using JSON configurations with specific targets for different environments:

Dockerfile:

FROM python:3.12-slim-bullseye AS focoos-cpu
...
FROM ghcr.io/focoosai/deeplearning:base-cu12-cudnn9-py312-uv AS focoos-onnx
...
FROM focoos-onnx AS focoos-torch
....
FROM focoos-torch AS focoos-tensorrt

devcontainer.json

{
    "name": "Focoos CPU",
    "build": {
        "context": "../..",
        "dockerfile": "../../Dockerfile",
        "target": "focoos-cpu"
    }
}

• Runtime System: Introduced new abstract base class and TorchscriptRuntime:

class BaseRuntime:
    @abstractmethod
    def __call__(self, im: np.ndarray, conf_threshold: float) -> sv.Detections:
        pass

• Dependency Management Updated pyproject.toml with new extras:

[project.optional-dependencies]
cpu = ["onnxruntime==1.20.1"]
onnx = ["onnxruntime-gpu==1.20.1"]
tensorrt = ["onnxruntime-gpu==1.20.1","tensorrt==10.5.0"]
torch = ["torch==2.3.0","torchvision"]

The changes provide a more flexible and maintainable development environment while improving the SDK's performance and usability across different hardware configurations.

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
__init__.py	10	0	100%
focoos.py	116	0	100%
local_model.py	63	0	100%
ports.py	248	0	100%
runtime.py	158	97	38%	28, 37–38, 112, 116, 120, 130, 132–133, 135–138, 141–143, 146, 149–151, 154, 157–158, 161–162, 165, 190–194, 196–197, 200–203, 205–206, 208, 212–215, 219–220, 222–224, 226–230, 232–233, 235–237, 241, 251–252, 262–266, 268, 270–271, 273–279, 283–286, 290–291, 293–294, 296–300, 302–303, 305–306, 308, 318–319
utils
system.py	85	5	94%	220–224
TOTAL	1027	138	86%

Tests	Skipped	Failures	Errors	Time
109	0 💤	0 ❌	0 🔥	6.660s ⏱️

[fcdl94]

ONNXRuntime sucks