yolo_inference

A wrapper for convenient inference with difference export formats for YoloV8.

Setup

Set up your environment with all the necessary dependencies by running:

$ pip install -r requirements.txt

Usage

from model import Model
from matplotlib.pyplot import imshow

model = Model("< path to onnx/tflite model >")
res = model("< image url/path >")
imshow(res.image_overlay)

Single image detection example at: example.ipynb
Full dataset mAP evaluation example at: map_eval.ipynb

Models

Some of our exported models are available for use here. For more variations, sign up for the DeGirum cloud platform to get access to our full ultralytics model zoo at cs.degirum.com!

Arguments

Model( 
    model_path:str, 
        # path to model, must be .onnx or .tflite file

    do_postprocessing:bool=True, 
        # do non-max-suppression and export results as Results objects
        # if False, export model output tensors directly

    do_bbox_decoding:bool=True, 
        # decode bounding box coordinates from model exported in DeGirum format (6 outputs)
        # MUST BE FALSE TO USE WITH MODELS EXPORTED VIA ULTRALYTICS REPO (1 output)

    labels:Optional[Union[str, dict[str,str]]]=None, 
        # either a path to a json file with a labels dictionary, or the dictionary object itself

    conf_threshold:float=0.25, 
        # confidence threshold for filtering results

    iou_threshold:float=0.75, 
        # IOU threshold for non-max suppression

    postprocessor_inputs:Optional[list[int]]=None
        # list of indices to rearrange model outputs before passing to postprocessor
        # if None, infer order of postprocessor inputs via model output shape
        # otherwise, manually override the order, in the following format:
        '''
          0. bounding box data [1, a, 64]
          1. bounding box data [1, b, 64]
          2. bounding box data [1, c, 64]
          3. probability data [1, a, num_classes]
          4. probability data [1, b, num_classes]
          5. probability data [1, c, num_classes]
        '''
        # NOTE: postprocessor input inference does not work with models that have 64 classes, 
        # as it relies on the differences in output shapes to check the order. for these models, 
        # you must input the indices manually

    )