Annotating and Training

1. Start a Desktop session in OnDemand

2. In the desktop terminal, run:
Copy

module load sleap/1.4.1
sleap-label

3. Once the SLEAP UI opens, click File > Open Project… and open the file /data/groups/wheelenj/mosquitoes/L1_L2_L4.labels.v001.slp
1. Might take a bit to load…
4. Add new videos via File > Add Videos… and select a new MP4
1. Check grayscale, then click Import
5. Click the Videos tab in the right-side pane
1. Double click the video you want to annotate
6. Click the Labeling Suggestions tab
1. In Generate Suggestions, choose Method: Sample; Samples Per Video: 20, Sampling method: stride
2. Change Target to current video
3. Click Generate Suggestions
4. Blue empty boxes will show up in the video nav bar
7. Click a blue suggestion. If there’s a mosquito:
1. Click the Instance tab
2. Click New Instance
3. Move the nodes to annotate
1. Examples (not to scale)

L1

L3
8. Press the space bar to move to the next suggested frame

Below are old notes…ignore for now.

Commands

# from BOSE
rsync -Ppvrz wheelenj@wheeler.nas.uwec.edu:/volume1/WheelerLab_data/InVision/20250616* .

# from desktop session terminal
module load sleap/1.4.1
sleap-label

File > Open Project… /data/groups/wheelenj/mosquitoes/L1_L2_L4.labels.v001.slp

File > Add Videos… /data/groups/wheelenj/mosquitoes/20250616a01controlard_20250616_140037.24568709/000000.mp4

Predict > Run Inference…

Save configuration files… `/data/groups/wheelenj/mosquitoes/models/

• inference-script.sh
Copy

#!/bin/bash
sleap-track /data/groups/wheelenj/mosquitoes/L1_L2_L4.labels.v001.slp --video.index 26 --frames 0,-2331 -m /data/groups/wheelenj/mosquitoes/models/250603_091255.single_instance.n=171/training_config.json --batch_size 4 --tracking.tracker none -o L1_L2_L4.labels.v001.slp.predictions.slp --verbosity json --no-empty-frames

• jobs.yaml
Copy

inference:
- data_path: L1_L2_L4.labels.v001.slp
  models:
  - /data/groups/wheelenj/mosquitoes/models/250603_091255.single_instance.n=171/training_config.json
  only_suggested_frames: false
  output_path: L1_L2_L4.labels.v001.slp.predictions.slp
  tracking:
    tracking.tracker: none
  type: video
training: []

• train-script.sh
Copy

#!/bin/bash

Inference processed at ~22 FPS

Terminal STDOUT:

In general very good performance with one consistent outlier/error:

Resource usage:

Now to try running from command line instead of GUI:

#!/bin/bash
sleap-track {input_data.mp4} -m /data/groups/wheelenj/mosquitoes/models/250603_091255.single_instance.n=171/training_config.json --batch_size 4 --tracking.tracker none --verbosity json --no-empty-frames

sleap-track /data/groups/wheelenj/mosquitoes/20250616a01cueard_20250616_140456.24568709/000000.mp4 -m /data/groups/wheelenj/mosquitoes/models/250603_091255.single_instance.n=171/training_config.json --batch_size 4 --tracking.tracker none --verbosity json --no-empty-frames

Processed at ~24 FPS

Decent performance with some outliers

This produces: /data/groups/wheelenj/mosquitoes/20250616a01cueard_20250616_140456.24568709/000000.mp4.predictions.slp

To get the CSV, use sleap-convert --format analysis.csv -o 000000.predictions.analysis.csv 000000.mp4.predictions.slp

2026