Run inference on large datasets

Once you are satisfied with your model’s performance, you will often want to use that model to run inference on a large number of videos. While the app allows you to upload videos through the browser and run inference, this is not the most efficient way to process a large number of videos (or even a small number of very large videos). To do so, we recommend using a provided script that is called from the command line.

Step 1: Upload data

Uploading large videos directly into the app is inefficient due to the way the Streamlit widget handles data in the browser. Instead, we recommend the following:

If using a local workstation:

Copy your videos to the directory Pose-app/data/<project_name>/videos_infer.

Note

It is good practice to make copies of your videos into the Lightning Pose project, and keep the originals saved elsewhere.

If using a Lightning Studio:

First, in the file browser on the left-hand side of the Studio verify if you have a directory named Pose-app/data/<project_name>/videos_infer. If not, right-click on the <project_name> directory and select New Folder.... Name the new folder videos_infer.

Next, you will need to create a temporary file in this directory so that it will sync with your Lightning Drive filesystem (empty directories are not synced).

Right click on videos_infer, select New File..., and name your file tmp.txt.

Now, click the Drive button on the right-hand menu in the Studio:

https://imgur.com/HQpoqth.png

You can then navigate to the directory this_studio > Pose-app > data > <project_name>. If you do not see videos_infer you can re-sync your Drive by clicking the sync button in the upper right corner:

https://imgur.com/qC3h2YI.png

Once videos_infer appears, click on this directory, then click the “Add data” button in the upper right corner:

https://imgur.com/MVOiyUb.png

Now you can drag and drop your videos to upload to Lightning, using a more efficient uploading mechanism than the Streamlit uploader available in the app.

If your videos are very large (>1GB) we recommend uploading one video at a time.

Step 2: Run inference from the command line

Now that the videos you wish to run inference on are all in one place, you can run a script that will loop over all videos and perform inference with a specific model. The bash script is provided in Pose-app/scripts. The bash script requires 5 positional arguments:

  1. absolute path to the model directory (usually in YY-MM-DD/HH-MM-SS format)

  2. absolute path to the directory of videos to be processed (note that every video in the directory will be processed; if you’d like to run inference on a subset of videos you may first save the new videos in a temporary directory, run this script (pointing it to the temporary directory), then move the videos into videos_infer)

  3. compute metrics flag; if true, compute metrics like temporal norm and pca errors (if these options were selected during project creation)

  4. label full video flag; if true, save a copy of the original video with model predictions overlaid (caution: if your original videos are large these will also take up lots of storage!)

  5. label video snippet flag: if true, create a 30 second video snippet with the highest motion energy with model predictions overlaid

Here is an example when running from a Lightning Studio (inside the Pose-app directory):

bash scripts/process_videos.sh /this_studio/Pose-app/data/<project_name>/2024-03-28/23-29-04_super-0 /this_studio/Pose-app/data/<project_name>/videos_infer true true true

This call will use the model located in 2024-03-28/23-29-04_super-0 to run inference on all videos in <project_name>/videos_infer, compute metrics on each video, and create both a full labeled video and a labeled video snippet.