First attempt at easymode

[cyychong]

Hi Mart,

Last time I posted a query on the Pom github page since I was testing out Pom and you suggested trying out easymode. Thought I would let you guys know how it went -- of course, this is a work in progress so I wasn't expecting anything.

So far I have only applied the membrane model and the nucleus model (which is my main organelle of interest). The installation went ok but I couldn't get tensorflow to work on my GPU so it has been using my CPU. It is probably some IT-ignorance on my part and I suspect when this goes live, it will work out nicely.

The membrane model worked reasonably well but I couldn't get the nucleus model to work. Below are two screenshots of the tomographic slice and the segmentation on the same slice.

This is a budding yeast tomogram that was imaged using the Volta phase plate + cryoCARE denoised at 10 angstrom/pixel -- basically I tried out easymode on my "easiest" test tomogram. Both the nucleus model + membrane model didn't work on my defocus tomogram at 6.8 angstrom/pixel but I haven't tested it on a larger pixel size (it takes 24 hours to segment one tomogram).

Hope this info helps a little. Thanks for developing easymode! Delighted that I don't have to wait too long for the day where I can type one command to get a segmentation.

Claris

[mgflast]

Hi Claris,

Thanks a lot for the update! Feedback like this is super useful. That segmentation looks horrible, sorry! 🥲

1. About not finding GPUs: this is most likely related to the CUDA / cuDNN versions in your environment. If you installed easymode/Ais with the default requirements, you would need to have CUDA 11.8 and cuDNN 8.6 installed. The one that is most commonly missing is cuDNN 8.6 - on our own cluster, when we use 'module load cuda/11.8', we still don't get cuDNN 8.6, so we had to install that manually. Instructions for how to install can be found here: https://docs.nvidia.com/deeplearning/cudnn/installation/latest/linux.html

2. For some features we have two different models available. By default we use the 3D UNet, but for organelles like the nucleus, mitochondria, and cytoplasm, the 2D networks currently work a lot better (we will train new 3D instances later). You can use these by adding the '--2d' flag to the 'easymode segment' call.
   2b. The 2D models rely on the most recent Ais version, so you should upgrade: Ais-cryoET==1.0.47. Also, I'm not sure which easymode version you are on but if you installed via pip it will be an old one, so best is to re-install from github as follows: pip install git+https://github.com/mgflast/easymode.git The nucleus model works at 50 A/px, so 24 hours is too long even on a CPU. I think you're using an old easymode version from when we didn't rescale tomograms on the fly yet.

One last thing: this week I added a new feature where users can 'report' model failures using easymode report. This command will upload a chosen tomogram to our server so that we can include it in future training runs. The idea is that if people use this, over time we'll be able to improve the networks and make them more suited for general use. Currently the maximum file size is 1.5 GB to prevent server costs getting out of hand - it looks like your tomogram may be too large but... I thought I would mention it.

Good luck! Let me know how it goes 😄