Issue 47: 3D Viz Next Steps

[davramov]

I am starting this as a draft PR for the features outlined here: #47

This initial commit includes:

• a small 11MB scan for testing (examples/tomo_scan.h5)
• downsampling methods using Torch and OpenCV (orchestration/flows/bl832/downsample.py)
  • For small datasets (~11MB), Torch/OpenCV performs similarly (4 seconds)
  • For large datasets (~7GB), OpenCV works well, however, Torch appears to run into a memory issue (at least on my laptop) which prevents some of the images from downsampling correctly (exchange/data). To address this, I include a memory check, enabling Torch chunking. This slows down the downsampling (OpenCV: 80 seconds vs Torch: 123 seconds), but at least it works. I am keeping both versions here in case one works better in production.

[davramov]

Yesterday I was able to get Seij's Bluesky TIled React component installed in my app via npm, and connected it to my "Select Dataset" button. There are still a few things to fix before it works properly:

• Test switching between datasets
• Update Tiled when a new dataset is available
• Fix the UI so there isn't a huge wrapper around the widget
• Load the bluesky icon correctly
• Close the Tiled Browser widget when the user presses "Select"

[davramov]

While testing at the beamline, I noticed that I should make it easier to tell the app (Dockerfile) where Tiled should look for the data.

EDIT: Actually I just need better documentation. Setting DATA_PATH in .env to the full path on the filesystem where the Zarr files are located works as expected

[davramov]

A few steps left to make this perfect...

• currently the tiled server isn't picking up new zarrs in the hosted directory, and I have to restart the container each time a new scan is done. this is not ideal, it should update automatically
• set up the vis computer with a prefect agent + reconstruction job controller flow
  • Reconstruction on the new vis computer is very fast (2-3x faster than a full HPC node, 2-4 minutes vs 10 minutes)
  • tiff to zarr on the other hand is very slow (6x slower than HPC. 10 minutes vs nearly 60 minutes). I am not sure what exactly is causing this slowdown
  • best of both worlds: maybe we run automated reconstruction on the vis computer, then send it to an HPC facility for tiff to zarr, alternating between ALCF and NERSC to keep the throughput up

[davramov]

A few things I've noticed after using the vis app some more:

• Tiled browser widget

  • broken image
  • the "Select" hover-over popup button is not wired into the select function. this could be confusing

• itk-vtk-viewer levels of detail

  • I haven't seen it actually load L0 or L1. Not sure if it's hitting some built-in browser limit, but I expect the vis PC with its specs should be able to handle it.

• A tutorial or info popup could be helpful