jni / lma-2021-bioimage-analysis-python Goto Github PK

During the workshop, Genevieve used the foreman analogy for the dask scheduler to explain why using tiny chunks are not optimal when trying to decide on chunk size. Chunks too big will crash our memory, but chunks too small will incur lots of overhead.

Feedback was great on the analogy so we should add it to the notebook! 😊

Reference from I2K:

https://github.com/jni/i2k-skimage-napari/blob/solved/lectures/3_dask.ipynb

This workshop was the first time the "images are numpy arrays" content was presentaed. Next time it runs, we should rearrange it so the exercises are interspersed through the content, rather than all at the end. I think that would be more engaging for participants.

We think it would be good to have another example in the segmentation tutorial.

Juan suggests demo-ing the napari cellpose plugin.

The dask tutorial includes this statement in the first couple of cells. I don't think this is strictly accurate, and it might be good to reword it.

A first life hack: in general, avoid using bare dask, and instead create a dask.distributed Client as in the cell below. What this buys you:

• memory management: the distributed scheduler that is automatically created with the default client will launch processes with maximum memory limits. If they exceed those limits, the scheduler will stop sending tasks to them and eventually kill them. In contrast, without distributed, you are subject to the same issues as bare Python. It is very easy to freeze your machine.

Dask documentation reference; https://docs.dask.org/en/latest/setup/single-distributed.html

Maybe we should add suggestions for where people can go for help after the workshop has finished.
Eg:

• https://forum.image.sc
• https://napari.zulipchat.com

Add dask-labextension to requirements / environment.yml
It's just nice to have, if you're using jupyter lab

https://github.com/dask/dask-labextension

Workshop participants put a whole bunch of great resources into the chat when we asked them about their favourite places to learn more about image analysis.

We should add that information here too (most likely to the last section in the README, which currently contains a link to forum.image.sc and the napari zulip channel).

Around Cell 19

I want us to add a link to this standup comedy act about spreadsheets to the first tutorial on numpy arrays. The video is 13 minutes long.

https://www.youtube.com/watch?v=UBX2QQHlQ_I