Skip to content

lacss.deploy

Attributes:

Name Type Description
model_urls

URLs for build-in pretrain models. e.g model_urls["default"].

Predictor dataclass

Main class interface for model deployment. This is the only class you need if you don't train your own model

Examples:

The most common use case is to use a build-in pretrained model.

import lacss.deploy

# look up the url of a build-in mode
url = lacss.deploy.model_urls["default"]

# create the predictor instance
predictor = lacss.deploy.Predictor(url)

# make a prediction
label = predictor.predict(image)

Parameters:

Name Type Description Default
url str | PathLike | tuple[Lacss, dict]

Model file (local or remote) or (module, parameters) tuple. If the parameters are a sequence instead of a dictionary, treat it as a model ensemble.

 required
f16 bool

If true, compute in f16 precision (instead of the default f32)

 False
grid_size int

Grid size for large images. Large image are broken down into grids to avoid GPU memeory overflow.

 544
step_size int

Should be slightly smaller than grid_size to allow some overlap between grids.

 480
grid_size_3d int

Grid_size for 3D input.

 256
step_size_3d int

Step_size for 3D input.

 192
mask_size int

Only use for bbox output. The dimension of the segmentation mask output.

 36
mc_step_size int

Only used for 3D contour output. The step size during the mesh generation.

 1
chain_approx int

Only used for 2D contour output. Polygon generation rule.

 CHAIN_APPROX_SIMPLE

__post_init__()

validate parameters

_call_model(patch, score_threshold, size_threshold, threshold, padding_value)

Core inference routine: 1. Pad input image to regular size and call model function 2. clean up results 3. compute contours

_compute_contours_2d(predictions, img_sz, threshold)

2D contour using cv2

_process_image(image, score_threshold, size_threshold, threshold, padding_value)

process a 2d/3d image in grids

predict(image, *, output_type='label', reshape_to=None, min_area=0, score_threshold=0.5, segmentation_threshold=0.5, nms_iou=0.4, remove_out_of_bound=None)

Predict segmentation.

Parameters:

Name Type Description Default
image ArrayLike

A ndarray of (h,w,c) or (d,h,w,c) format. c must be 1-3

 required

Other Parameters:

Name Type Description
output_type str

"label" | "contour" | "bbox"
reshape_to int | tuple[int] | ndarray | None

If not None, the input image will be resized internally before send to the model. The results will be resized back to the scale of the orginal input image.
min_area float

Minimum area of a valid prediction.
score_threshold float

Min score needed to be included in the output.
segmentation_threshold float

Threshold value for segmentation
nms_iou float

IOU threshold value for non-max-supression post-processing

Returns:

Type Description
dict

For "label" output:

  • pred_scores: The prediction scores of each instance.
  • pred_label: a 2D image label. 0 is background.
dict

For "contour" output:

  • pred_scores: The prediction scores of each instance.
  • pred_contours: a list of polygon arrays in x-y format.
dict

For "bbox" output (ie MaskRCNN):

  • pred_scores: The prediction scores of each instance.
  • pred_bboxes: The bounding-boxes of detected instances in y0x0y1x1 format
  • pred_masks: A 3d array representing (rescaled) segmentation mask within bboxes

_ensemble_model_fn(module, params, image)

A JIT compiled function to call Lacss model ensemble if params is a sequence (not a dict), assuming a model ensemble and generate ensembed prediction.