cellmap_flow.post.postprocessors

Attributes

logger

Classes

`PostProcessor`	Base class for post-processing methods.
`SigmoidPostprocessor`	Apply sigmoid activation to convert logits to probabilities.
`DefaultPostprocessor`	Base class for post-processing methods.
`ThresholdPostprocessor`	Base class for post-processing methods.
`FillHolesPostprocessor`	Threshold, then fill the background holes each foreground blob encloses.
`LabelPostprocessor`	An id per connected object of one channel's nonzero voxels (`segment.connected_components`).
`MortonSegmentationRelabeling`	Base class for post-processing methods.
`AffinityPostprocessor`	Objects from affinities by mutex watershed (`segment.mutex_watershed`).
`SimpleBlockwiseMerger`	Base class for post-processing methods.
`CellposeMasksPostprocessor`	Cellpose's masks, made from a Cellpose model's flows output.
`ChannelSelection`	Base class for post-processing methods.
`LambdaPostprocessor`	Base class for post-processing methods.

Functions

`get_postprocessors_list`(→ list[dict])	Returns a list of dictionaries containing the names and parameters of all subclasses of PostProcessor.
`get_postprocessors`(→ list[PostProcessor])	Get postprocessors from either dict or list format.

Module Contents

cellmap_flow.post.postprocessors.logger

class cellmap_flow.post.postprocessors.PostProcessor

Base class for post-processing methods.

property is_segmentation

problem_with(model_config)

Why this step cannot run on model_config’s output, or None when it can.

Checked by the dashboard before it applies a chain: a step that fails on every chunk does so inside the model’s server, where the page shows only an empty layer.

class cellmap_flow.post.postprocessors.SigmoidPostprocessor

Apply sigmoid activation to convert logits to probabilities.

property dtype

class cellmap_flow.post.postprocessors.DefaultPostprocessor(clip_min: float = -1.0, clip_max: float = 1.0, bias: float = 1.0, multiplier: float = 127.5)

Base class for post-processing methods.

clip_min

clip_max

bias

multiplier

property dtype

property is_segmentation

class cellmap_flow.post.postprocessors.ThresholdPostprocessor(threshold: float = 0.5)

Base class for post-processing methods.

threshold

property dtype

property is_segmentation

class cellmap_flow.post.postprocessors.FillHolesPostprocessor(threshold: float = 0.0)

Threshold, then fill the background holes each foreground blob encloses.

For compact single-instance organelles (a nucleus, say) that never have interior gaps; not for structures with a real lumen. It runs per chunk with no halo, so a hole that touches the chunk’s boundary is not enclosed within the chunk and stays unfilled.

threshold

property dtype

property is_segmentation

class cellmap_flow.post.postprocessors.LabelPostprocessor(channel: int = 0, connectivity: int = 1, min_size: int = 0, per_slice: bool = False)

An id per connected object of one channel’s nonzero voxels (segment.connected_components).

connectivity: 1 faces (the default), 2 faces and edges, 3 all neighbours. min_size: objects of fewer voxels become background. per_slice: each z slice labelled on its own, in 2D. The other channels pass through, as uint32.

channel = 0

connectivity

min_size = 0

per_slice

property dtype

property is_segmentation

class cellmap_flow.post.postprocessors.MortonSegmentationRelabeling(channel: int = 0)

Base class for post-processing methods.

channel = 0

num_previous_segments = 0

use_exact = True

property dtype

property is_segmentation

class cellmap_flow.post.postprocessors.AffinityPostprocessor(bias: float = 0.0, neighborhood: str = '[\n [1, 0, 0],\n [0, 1, 0],\n [0, 0, 1],\n [3, 0, 0],\n [0, 3, 0],\n [0, 0, 3],\n [9, 0, 0],\n [0, 9, 0],\n [0, 0, 9],\n ]')

Objects from affinities by mutex watershed (segment.mutex_watershed).

bias: the affinity above which neighbours join, and below which a fragment’s mean makes it background. neighborhood: the offset each channel compares, as the model was trained with. Ids are offset by the chunk’s Morton index, so they are unique across chunks.

bias

neighborhood

use_exact = True

num_previous_segments = 0

property dtype

property is_segmentation

property num_channels

class cellmap_flow.post.postprocessors.SimpleBlockwiseMerger(channel: int = 0, face_erosion_iterations: int = 0)

Base class for post-processing methods.

channel = 0

face_erosion_iterations = 0

use_exact = True

equivalences

chunk_slice_position_to_coords_id_dict

slices

keys_to_skip

equivalences_json(): The equivalences so far, read while no other chunk is adding to them.

calculate_equivalences()

property dtype

property is_segmentation

class cellmap_flow.post.postprocessors.CellposeMasksPostprocessor(flow_threshold: float = 0.4, cellprob_threshold: float = 0.0, min_size: int = 15, stitch_threshold: float = 0.0, niter: int = 200)

Cellpose’s masks, made from a Cellpose model’s flows output.

Cellpose turns its three output channels into objects by following each pixel’s flow to where the flows converge. Run here, on a server with output: flows (channels flow_y, flow_x and cell, the probability), its thresholds can be changed from the dashboard and the layer redrawn, where output: masks fixes them when the server starts. Put it first in the chain, on the flows as served. It needs Cellpose, which a Cellpose model’s server has.

Each slice is segmented on its own, as Cellpose segments a 2D image; stitch_threshold above 0 joins a slice’s object to the next slice’s one it overlaps by that IoU (Cellpose’s stitch3D), else each slice’s ids follow the previous slice’s. Ids are unique within the chunk: follow with MortonSegmentationRelabeling for ids unique across chunks.

Parameters:

flow_threshold – how far an object’s flows may be from those its shape implies before it is dropped (Cellpose’s; 0 keeps all).
cellprob_threshold – the cell probability, as a logit, over which a pixel is followed (Cellpose’s; 0 is a probability of 0.5).
min_size – objects of fewer pixels are removed (per slice; in 3D after stitching).
stitch_threshold – IoU joining objects across slices; 0 is off.
niter – flow-following steps (Cellpose’s default, 200).

uses_device = True

flow_threshold

cellprob_threshold

min_size = 15

stitch_threshold

niter = 200

problem_with(model_config): None for a Cellpose model (or a finetune of one) served with output: flows.

property dtype

property is_segmentation

property num_channels

class cellmap_flow.post.postprocessors.ChannelSelection(channels: str = '0')

Base class for post-processing methods.

channels

property num_channels

class cellmap_flow.post.postprocessors.LambdaPostprocessor(expression: str)

Base class for post-processing methods.

expression

property dtype

cellmap_flow.post.postprocessors.get_postprocessors_list() → list[dict]: Returns a list of dictionaries containing the names and parameters of all subclasses of PostProcessor.

cellmap_flow.post.postprocessors.get_postprocessors(elms) → list[PostProcessor]: Get postprocessors from either dict or list format.