In notebook 3, I remember the students struggling a bit with the model configuration task since there are so many possible parameters. Can we give them a list of the essential parameters that they need to consider and focus on?

Universal tasks

• Update environment setup to use mamba (See https://github.com/dlmbl/DL-MBL-2023/wiki/Exercise-Guidelines#environment)
• Add a red callout block to the beginning of each notebook that specifies which kernel to use

<div class="alert alert-danger">
Please use kernel <kernel name here>
</div>

• Check for any references to slack. We are going to use Element this year.

Hi Larissa @neptunes5thmoon,

Great notebooks! Very minor nitpicking feedback:

exercise1.ipynb

1. Small typo in acquisitions in Here the high SNR images are acquistions of Human U2OS cells taken
2. In the line here

"Since the image pairs were synthetically created in this example, they are already aligned perfectly.",

one could add an extra line saying that ...

"for real data, usually, low and high image pairs are not pixel-level aligned."

3. Small typo in CSBDeep in

"To train a denoising network we will use the CSB Deep Repo"

4. Small typo. Instead of:

"This part will no tasks to fill in in this exercise",

we can say:

"This part will have no tasks to fill in, but go through ..."

5. The data comes from the BBBC dataset as mentioned in the introduction, but then why do we say later that the data is Tribolium?

" First we download some example data, consisting of low-SNR and high-SNR 3D images of Tribolium"

6. Maybe since the dataset which is downloaded is 277 MiB and could take time to download, we could add a line to the order of:

"... Please wait for a few minutes for the data to be downloaded."

7. Could change:
   print("low image size =", y.shape)
   to
   print("low SNR image size =", y.shape)
8. Could explain the generated graph in:

"Visualize results Plot the test stack pair and the predicted restored stack (middle)."

9. Could change the italics formatting from:
   This is _not_ done in this notebook
   to This is <i>not</i> done in this notebook

exercise2.ipynb

1. Typo in microsecond:

"As you can see the last image, which is the average of 4 images with 5$\mu s$ scantime"

Timing

CARE notebook: ~90 min till we went over questions (cause it was lunchtime) ca. 10:30-12:00
N2N notebook: ~90 min (ca. 15:30-17:00) for students to work on, ca. 30 min to discuss in the group (3 students had finished the notebook at that point)
N2V notebook: ~30 min (ca.17:30-18:00), then discuss solutions in ca 5 min

Problems/Ideas for improvement:

• Small bug: logdir needs to exist for tensorboard to work
• some students had other notebooks from previous exercises running and then got cryptic OOM errors
• for N2N: students struggled with understanding that the SEM data comes as one tiff stack of several images and then should be resaved as single images for training
• N2N: clarify difference between model.predict and model.keras_model.predict (model.predict under the hood splits the image into patches to predict on, maybe more)
• N2V: in addition to the assert that the training/test data is there, check that the folders do not contain other tif files - or switch implementation to load_imgs
• useful addition would be a task to look at patch size vs receptive field size

Finished reviewing (looked at ben/denoisplit as per mattermost).
01 and 04 need some work I think but for the other ones only minor comments:

Installation

• end up in a different directory after running source setup.sh
• recommended is putting data on our s3 bucket and download from there
• dependencies of denoiSplit not installed

01_CARE

• missing colored boxes for checkpoints, exercises etc. in the exercise notebook
• still got TODOs in there (most importantly: missing scaffolding for some tasks)
• import for augment_batch and normalize is at the very beginning and defined by you. Maybe add a comment that makes clear that it's provided by you and add a cell with ?augment_batch and ?normalize at task 1, similarly for denormalize in task 3
• agree that calculation of means and stds should go in the function
• they already know about data loaders in principle
• the unet should come from here: https://github.com/dlmbl/dlmbl-unet
• IMO importing torch.nn and torch.optim would be better than importing with as nn
• I got NaNs in my loss when my lr was too high
• something with the training is not working for me - validation loss is always the same

02_Noise2Void

• config explanation: I think axes being "YX" warrants more explanation (e.g. they might get 2D but be thrown for a loop by it not being "XY" )
• getting some warnings during training
• Task2: tensorboard is already started
• in export: TypeError: CAREamist.export_to_bmz() got an unexpected keyword argument 'path' (parameter is called path_to_archive)
• Task 5: detail but data is already downloaded, so change text to "loaded in the following cell"

03_COSDD

• put explanation for dimensions before or in the task instead of after
• still has two TODOs (figure and pretrained model)
• in tensorboard metrics: 20 min in text instead of 10 min
• if you can add tensorboard logs for the pretrained models that'd be great
• Bonus
  • import is wrong
  • also still TODO: add pretrained
  • device is not defined

04_DenoiSplit

• sys path extension isn't right
• missing dependency albumentations
• I assume this isn't done? Very light on details of what's happening
• exercise version has answers to questions

05 Noise2Noise

• text: "regression exercises" -> restoration exercises
• Task1: typo vizualize -> visualize