micro_sam.automatic_segmentation: embeddings

[aloboa]

I understand that, in micro_sam.automatic_segmentation,

  -e EMBEDDING_PATH, --embedding_path EMBEDDING_PATH
                        The path where the embeddings will be saved.

if embeddings do not exist, they are are created in that folder, but that if they exist, they are read from that folder

Correct?

[constantinpape]

Yes, that is correct.

Note that you don't have to give that argument at all, in that case the embeddings will just be computed internally and will not be saved to a file.

[aloboa]

Note that you don't have to give that argument at all, in that case the embeddings will just be computed internally and will not be saved to a file.

My understanding is that embeddings must be calculated for each image (and model) and then the same embeddings can be used with different values of
-pred_iou_thresh, --stability_score_thresh and --box_nms_thresh, to speed up the processing, correct?

BTW -pred_iou_thresh, --stability_score_thresh and --box_nms_thresh are not listed by the -h option:

$ micro_sam.automatic_segmentation -h
usage: micro_sam.automatic_segmentation [-h] -i INPUT_PATH -o OUTPUT_PATH [-e EMBEDDING_PATH] [--pattern PATTERN] [-k KEY]
                                        [-m MODEL_TYPE] [-c CHECKPOINT] [--tile_shape TILE_SHAPE [TILE_SHAPE ...]]
                                        [--halo HALO [HALO ...]] [-n NDIM] [--mode MODE] [--annotate] [-d DEVICE] [-v]

[constantinpape]

My understanding is that embeddings must be calculated for each image (and model) and then the same embeddings can be used with different values of
-pred_iou_thresh, --stability_score_thresh and --box_nms_thresh, to speed up the processing, correct?

Yes, that is correct.

BTW -pred_iou_thresh, --stability_score_thresh and --box_nms_thresh are not listed by the -h option:

Yes, this is because they are part of keyword arguments that depend on whether AMG or AIS is used. It is difficult to expose them as first class citizen in the CLI.

[constantinpape]

I think the questions here are all resolved.

[aloboa]

I think the questions here are all resolved.

Yes, but please see https://forum.image.sc/t/specifying-embeddings-in-suite-of-micro-sam-automatic-segmentation-runs/110549.

BTW -pred_iou_thresh, --stability_score_thresh and --box_nms_thresh are not listed by the -h option:

Yes, this is because they are part of keyword arguments that depend on whether AMG or AIS is used. It is difficult to expose them as first class citizen in the CLI.

It is more difficult to use the options if you do not even know they exist. An alternative would be having two separate commands, one for AMG and another for AIS.