cellmap_flow.ai_annotate.staging

Where an AI annotation waits for review: a staging directory per annotation.

A model’s result is not written into the annotation volume until the user accepts it. Until then it lives in <corrections_dir>/.ai_annotate/<id>/: the image sent (input.png), the image the model returned (model.png, when it returned one), the mask’s objects on the write box’s voxels (mask_write.npy) and meta.json (the plan, the prompt, the model, the label name). Accepting writes the objects and discards the directory, as rejecting does. Nothing secret is staged: the prompt and ids are the user’s own, and no provider credential ever reaches these functions.

The id is the server’s (new_annotate_id) and is checked against ANNOTATE_ID_RE before it is joined into any path, so an id from a request can only ever name a directory under .ai_annotate.

Attributes

ANNOTATE_ID_RE
STAGING_DIRNAME
INPUT_PNG
MODEL_PNG
MASK_NPY
META_JSON
OVERLAY_ALPHA

Functions

new_annotate_id() → str	A fresh annotation id: 32 lowercase hex characters.
staging_dir(corrections_dir, annotate_id) → pathlib.Path	<corrections_dir>/.ai_annotate/<annotate_id>; the id is validated first.
stage(corrections_dir, annotate_id, *, plan, request, ...) → dict	Stage a model's result for review; returns the meta written to meta.json.
load(corrections_dir, annotate_id)	(meta, mask_write, input_image) of a staging.
overlay(input_image, mask_write, target_rgb) → PIL.Image.Image	input_image with the voxels the mask labels tinted target_rgb.
preview(corrections_dir, annotate_id) → dict	The review panel's images, base64 PNG: input_png, model_png (None
discard(corrections_dir, annotate_id) → None	Delete a staging; nothing happens when there is none.

Module Contents

cellmap_flow.ai_annotate.staging.ANNOTATE_ID_RE

cellmap_flow.ai_annotate.staging.STAGING_DIRNAME = '.ai_annotate'

cellmap_flow.ai_annotate.staging.INPUT_PNG = 'input.png'

cellmap_flow.ai_annotate.staging.MODEL_PNG = 'model.png'

cellmap_flow.ai_annotate.staging.MASK_NPY = 'mask_write.npy'

cellmap_flow.ai_annotate.staging.META_JSON = 'meta.json'

cellmap_flow.ai_annotate.staging.OVERLAY_ALPHA = 0.5

cellmap_flow.ai_annotate.staging.new_annotate_id() → str

A fresh annotation id: 32 lowercase hex characters.

cellmap_flow.ai_annotate.staging.staging_dir(corrections_dir, annotate_id) → pathlib.Path

<corrections_dir>/.ai_annotate/<annotate_id>; the id is validated first.

cellmap_flow.ai_annotate.staging.stage(corrections_dir, annotate_id, *, plan, request, result, mask_write, provider_id, model, label_name, volume_id=None) → dict

Stage a model’s result for review; returns the meta written to meta.json.

plan is the PlanePlan, request and result the backend’s SegmentRequest and SegmentResult, mask_write the mask’s objects on the write box (pipeline.mask_to_write_shape). Staging the same id again (a resend) replaces what was there; meta.json is written last, so a staging with one is complete.

cellmap_flow.ai_annotate.staging.load(corrections_dir, annotate_id)

(meta, mask_write, input_image) of a staging.

Raises AIAnnotateError (“refused”, 404) when there is none, as after an accept, a reject or a restart that cleared it.

cellmap_flow.ai_annotate.staging.overlay(input_image: PIL.Image.Image, mask_write: numpy.ndarray, target_rgb) → PIL.Image.Image

input_image with the voxels the mask labels tinted target_rgb.

The mask is the write box’s, laid out as the image is (to_screen) and scaled to it by nearest neighbour, so the overlay shows what accepting writes, not the model’s own image (portable.ai_annotate.overlay, which the browser’s review uses too).

cellmap_flow.ai_annotate.staging.preview(corrections_dir, annotate_id) → dict

The review panel’s images, base64 PNG: input_png, model_png (None when the model returned no image) and overlay_png.

Every image is re-encoded here, so nothing the model sent reaches the browser as it came.

cellmap_flow.ai_annotate.staging.discard(corrections_dir, annotate_id) → None

Delete a staging; nothing happens when there is none.