Overhaul exercise for 2024 course

.github/workflows/build-notebooks.yaml

@@ -0,0 +1,32 @@
+name: Build Notebooks
+on:
+  push:
+
+jobs:
+  run:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: "3.10"
+
+      - name: Install dependencies
+        run: |
+          python -m pip install -U pip
+          python -m pip install jupytext nbconvert
+
+
+      - name: Build notebooks
+        run: |
+          jupytext --to ipynb --update-metadata '{"jupytext":{"cell_metadata_filter":"all"}}' solution.py
+
+          jupyter nbconvert solution.ipynb --TagRemovePreprocessor.enabled=True --TagRemovePreprocessor.remove_cell_tags solution --to notebook --output exercise.ipynb
+          jupyter nbconvert solution.ipynb --TagRemovePreprocessor.enabled=True --TagRemovePreprocessor.remove_cell_tags task --to notebook --output solution.ipynb
+
+      - uses: EndBug/add-and-commit@v9
+        with:
+          add: solution.ipynb exercise.ipynb

README.md

@@ -1,19 +1,39 @@
-# Exercise 9: Explainable AI and Knowledge Extraction
+# Exercise 8: Explainable AI and Knowledge Extraction
 
+## Overview
+The goal of this exercise is to learn how to probe what a pre-trained classifier has learned about the data it was trained on. 
+
+We will be working with a simple example which is a fun derivation on the MNIST dataset that you will have seen in previous exercises in this course. 
+Unlike regular MNIST, our dataset is classified not by number, but by color!
+
+![CMNIST](assets/cmnist.png)
+
+In this exercise, we will return to conventional, gradient-based attribution methods to see what they can tell us about what the classifier knows. 
+We will see that, even for such a simple problem, there is some information that these methods do not give us. 
+
+We will then train a generative adversarial network, or GAN, to try to create counterfactual images. 
+These images are modifications of the originals, which are able to fool the classifier into thinking they come from a different class!. 
+We will evaluate this GAN using our classifier; Is it really able to change an image's class in a meaningful way? 
+
+Finally, we will combine the two methods — attribution and counterfactual — to get a full explanation of what exactly it is that the classifier is doing. We will likely learn whether it can teach us anything, and whether we should trust it!
+
+If time permits, we will try to apply this all over again as a bonus exercise to a much more complex and more biologically relevant problem.
+
+![synister](assets/synister.png)
 ## Setup
 
-Before anything else, in the super-repository called `DL-MBL-2023`:
+Before anything else, in the super-repository called `DL-MBL-2024`:
 ```
 git pull
-git submodule update --init 09_knowledge_extraction
+git submodule update --init 08_knowledge_extraction
 ```
 
 Then, if you have any other exercises still running, please save your progress and shut down those kernels.
 This is a GPU-hungry exercise so you're going to need all the GPU memory you can get.
 
 Next, run the setup script. It might take a few minutes.
 ```
-cd 09_knowledge_extraction
+cd 08_knowledge_extraction
 source setup.sh
 ```
 This will:
@@ -28,10 +48,7 @@ jupyter lab
 ```
 ...and continue with the instructions in the notebook.
 
-## Overview
 
-In this exercise we will:
-1. Train a classifier to predict, from 2D EM images of synapses, which neurotransmitter is (mostly) used at that synapse
-2. Use a gradient-based attribution method to try to find out what parts of the images contribute to the prediction
-3. Train a CycleGAN to create counterfactual images
-4. Run a discriminative attribution from counterfactuals
+### Acknowledgments
+
+This notebook was written by Jan Funke and modified by Tri Nguyen and Diane Adjavon, using code from Nils Eckstein and a modified version of the [CycleGAN](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix) implementation.

create_environment.sh

@@ -0,0 +1,5 @@
+# Contains the steps that I used to create the environment, for memory
+mamba create -n 08_knowledge_extraction python=3.11 pytorch torchvision pytorch-cuda=12.1 -c conda-forge -c pytorch -c nvidia
+mamba activate 08_knowledge_extraction
+pip install -r requirements.txt
+mamba env export > environment.yaml