SAVA ML TOOLBOX

SAVA ML TOOLBOX is a Python library for Machine Learning tasks. It provides a collection of tools and utilities to simplify the process of deploying machine learning models for SAVA project.

Installation

Clone the repository

git clone https://github.com/username/sava-ml-toolbox.git

And navigate into the folder

cd sava-ml-toolbox

Create environment (Optional)

Create a virtual environment in python with venv:

python3 -m venv .env

And activate it on Windows:

.env\Scripts\activate 

And activate it on Linux/Mac:

source .env/bin/activate 

Install the library

Navigate to the root directory of the project (where the setup.py file is located), and run:

pip install -e .

Usage

Pick one of the configurations from the config folder or create one in YAML format, with the following template:

# Runtime configuration
model_path: <path_to_your_model>
runtime: onnxruntime | tensorrt
providers:
  - <provider_1>
  - <provider_2>
batch_size: 1 # For now, 1 is the only accepted value

# Model configuration
conf_thres: <confidence_threshold>
iou_thres: <iou_threshold>
num_masks: <number of masks for the segmentation part>
use_sahi_optim: True | False

# SAHI related configuration
slice_height: <int sliding window height> 
slice_width: <int sliding window width>  
overlap_height_ratio: <vertical overlapping value 0<x<1 >
overlap_width_ratio: <horizontal overlapping value 0<x<1 >

# Classes
classes:
  - 0: class_0
  - 1: class_1
  - ...

Then you can use the Inference Engine as follow:

from sava_ml_toolbox.inference import InferenceEngine
from sava_ml_toolbox.utils import draw_detections
# ...

config = "path/to/config.yaml"

# Create Inference Engine
inference_engine = InferenceEngine(config)

# Load the sample image as PIL Image
img = Image.open(<image_path>)

# Perform inference
results = inference_engine.predict(img)

# Draw the results
out_img = inference_engine.draw_results(img, results)

# ...

The output of the predict function is a class which contains the following attributes:

• xywh - 2D bounding box
• xyzwhd - 3D bounding box
• segm - Segmentation mask
• class_id - Class ID of the detected object
• score - Confidence score of the detection
• pc - Point cloud data

Export YOLO family to ONNX

After installing ultralitics through pip, the following command allows to export yolo models:

yolo export model=yolov8x-seg.pt format=onnx opset=15

License