csc-training/geoml

Course material for "Practical machine learning for spatial data"

geoml's Issues

2022 schedule and schedule comments

Day 1:
Intro	0:15
L: ML intro:	0:45	Mats
L: ML classif/reg/clustering	1:00	Mats
E1: Clustering	0:30	Kylli
LUNCH
L: Data preparation	0:35	Kylli/ Samantha
E2: Vector data prep	0:30	Samantha
L: Shallow modelling models	1:25	Mats
Day2
E4: Shallow modelling, regression	0:30	Samantha
E5: Shallow modelling, classification	1:15	Kylli
L: Intro to DL:	0:45	Markus
LUNCH
L: Fully connected neural networks	0:30	Markus
E6: Dense ML vector	1:00	Samantha
E7: Dense ML raster	1:00	Kylli
Day3:
L: CNN	1:10	Markus
L: Puhti overview (modules, partitions, job submission) + GPU + batch jobs	1:00	Samantha / Markus
E8: CNN data preparations	0:20	Kylli
LUNCH
E9: Keras CNN: Segmentation	2:10	Kylli
L: GIS tools and other links	0:20	Samantha

Comments:

• Vector data preparation could have more time.
• Some lectures could be split into shorter parts and mixed with exercises.
• GIS tools and other links was 10 min, could have been longer and possibly more organized

Update requirements.txt to include ALL needed packages

• check that all packages are includes
• also the ones for raster preparation

2022 feedback and issues to fix

Notes during course

• add to regression notebook some more info on metrics
• rm metrics_dict from lasso call

Deep regression:

• output layer, activation function, mention about activation none is equal to linear (from slides)
• some non-helpful comments leftovers; eg output layer; (could still be improved but main point fixed)
• make possible to train in multiple steps; split definition and fitting;
• make possible to plot all history
• plot description; could be better
• Neurons, Nodes, perceptrons; use same as in rest of material
• Add separate input layer
• Add image of Network

Deep classification:

• one comment missing in end of box 3 (or so?)
• gradient boost filename wrong → fails for last plotting

CNN:

• readme: first time HPC mention; replace with supercomputer?
• batch jobs: was 2 seconds, should be 2 minutes: 00:02:00 instead of 00:00:02
• mistake fixing in sbatch scripts: Rather than fixing multiple files (hard to follow), it could also be overwritten in the command line call (same for everyone)
• "if mistake" in CNN?
• make some commands from multiclass comments in second to last notebook
• ntask and srun combination, can it be left out completely?

Outro

• outro could be more ordered and link list more focused

Update README to reflect which parts need to be updated when working outside of the course environment

Mention

• hardcoded paths
• reservation
• project
• how to get the material to Puhti/own computer via Jupyter in web interface (git extension? / terminal)
• which module to use

Recurring TODOs:

Pre-course:

• Reserve Dogmi
• Fill course form
• Check that course project has enough BU
• Training accounts
• CPU reservation (?)
• GPU reservation
• Eventilla page
• Notebooks course request, as backup (?)
• Update/review lectures
• Update/review exercises
• Test Notebooks + exercises (?)
• Test Puhti exercises
• Test with training accounts
• Create Zoom room (Mats)
• Collaborative doc page
• Test Dogmi setup; computers, camera, microphones
• reserve microphones, camera, extra computer as needed
• daily: copy eduroam WIFI info from info screens to collaborative document
• pre-course practicalities email to participants
• book extra room for online help (if needed?)
• contact event support about coffeebreak timings and food request
• ...

Post-course:

• Material copies to neiet
• write down own feedback, share with all teachers and event-team
• Store (on event page without any names?) and share introductions only with participants, make sure no names and stuff stays public
• Wrap-up email to participants with links and hint to DL course (if space available) and other courses, mention importance of feedback (?)
• clean scratch 2002044
• ...