Some Questions about DLC

Usage & Issues
#1

Hello,

We are students in the Computational Biology department at Bar Ilan University, Israel.

We are conducting a project in Dr. Amiyaal Ilany’s lab, aimed at identifying the behavior of rock hyrax during social interaction by using artificial intelligence.

We are considering using Deep Lab Cut in our project.

We would appreciate if you could please help us with our following questions:

Firstly, we downloaded DLC to our PCs (Dell inspiration 14 core i7 and Lenovo Ideapad core i5), but each attempt to run a video of about half a minute took us several hours. We would appreciate if you could please tell us the magnitude of the time it should take to process a half-minute video.

Secondly, we tried to install the DLC through Microsoft’s cloud (Azure) but, after running the command “import deeplabcut”, we encountered the following error: “no module named deeplabcut”. We saw a similar topic in the forum but the solution offered there did not help us. Are you able to help us with this?

Due to the problems described above, we are interested in finding a more efficient platform for running DLC, and are debating between purchasing a new and more powerful computer, or purchasing a subscription to a cloud or external server. Are you able to recommend which of the two options would work better?

If your recommendation is to purchase a new computer, is there a particular computer model you recommend? Specifically, would the new Mac Mini with M1 processor work?

If your recommendation is to run on a server / cloud, which one do you recommend? (Google cloud, Azure, etc)

Thank you very much in advance for your help and guidance.

Kind regards,

Avital Abergel and Gitali Naim

#2

Hi,

Couple of questions first. What is the resolution of video, you were trying to analyze on laptops? With high resolution it is possible that analysis on a CPU would take forever - training the model properly would be even longer though.
Second, when installing DLC on Azure, did you use pip installed it some other way? I’m not familiar with Azure so I’m not really sure how it handles installing packages and dependencies for those packages.

If you want to use a cloud, Google Colab is probably the easiest to implement DLC workflow and all you need is a Google account with space on the Drive for your project and videos. If you’d rather use a desktop it all comes down to your budget and use case. Using GPU instead of CPU is far more efficient so with all due respect to Apple and their new chip it will still be worse than even a cheap GPU. If you plan to record highly detailed (for instance 4K, high bitrate) videos training and analysing data on a CPU is going to take forever and isn’t a viable option. For comfort, imo is better to have your own workstation or if you have easy access to a local cloud (As in, if your Uni has it’s own computational cloud). But using Colab isn’t so bad, you just have to be more hands on, since single session times out after 12 hours.

If you want to get your own workstation and don’t know much about building computers, you can message me and I can get you a PCPartPicker list for your budget.

First of all though, try out Colab using example notebook and see how DLC works for your use case before you start spending money.

#3

HI konard,
First of all, thank you very much for the answer and the recommendations.
For starters we tried to use Google Colab and it does seem like a pretty good solution, however, there seem to be very specific models. We are working with rock hyrax. Is it possible to define a new animal through Google Collab or does it only use existing models?
In other word, is it possible to set up a new animal in Google Collab or do we have to download the software and perform the processing manually?
In addition, I wanted to know if it is possible that uploading a specific video on one computer takes about twenty minutes and on another computer the same video takes many hours and even gets stuck and does not upload at all?
Finally, I will note that the resolution of the videos we work with is relatively high (around 1.44 MP).
Thanks alot!

#4

Well, you label the data yourself and train your own model, there is no restrictions as to what animal you want to get pose estimation from. Pretrained models from the zoo are not required to run DLC. I’m not sure what you mean by uploading a video. Uploading to google drive?

With DLC you use Colab only to train and analyze your data. First steps: creating project, extracting outlier frames, labeling happens on your computer. Then you upload the whole project to Colab and continue with next steps. I highly reccomend to just use the documentation DeepLabCut/UseOverviewGuide.md at 37d678605563e7d6a19caf851b9614f29f2ce7d2 · DeepLabCut/DeepLabCut · GitHub and follow it step by step. Whole workflow is explained very easily and thoroughly and the GUI is kinda self-explanatory cause every step is ordered just the way you want to do them.