Staging

.github/actions/setup-rendering-deps/action.yml

@@ -18,15 +18,12 @@ inputs:
 runs:
   using: 'composite'
   steps:
-    - name: Cache APT packages
-      id: cache-apt
+    - name: Cache fonts
+      id: cache-fonts
       uses: actions/cache@v3
       with:
-        path: |
-          /var/cache/apt/archives
-          /usr/share/fonts/truetype/humor-sans
-        key: apt-packages-${{ runner.os }}-fonts-backend-graphviz-v2
-        restore-keys: apt-packages-${{ runner.os }}-
+        path: /usr/share/fonts/truetype/humor-sans
+        key: fonts-${{ runner.os }}-humor-sans-v1
 
     - name: Install XKCD fonts
       if: ${{ inputs.skip-fonts != 'true' }}

.github/workflows/notebook-pr.yaml

@@ -148,33 +148,23 @@ jobs:
         run: |
           nb="${{ matrix.notebook }}"
           dir=$(dirname "$nb")
+          nb_name=$(basename "$nb" .ipynb)
           echo "dir=$dir" >> $GITHUB_OUTPUT
-          # Create a safe artifact name from the notebook path
-          safe_name=$(echo "$nb" | tr '/' '_' | tr ' ' '_')
-          echo "artifact_name=$safe_name" >> $GITHUB_OUTPUT
+          echo "nb_name=$nb_name" >> $GITHUB_OUTPUT
+          # Use ___ as delimiter (won't appear in paths) so we can restore later
+          safe_dir=$(echo "$dir" | sed 's|/|___|g')
+          echo "artifact_name=${safe_dir}___${nb_name}" >> $GITHUB_OUTPUT
 
-      - name: Upload processed notebook
+      - name: Upload processed tutorial directory
         uses: actions/upload-artifact@v4
         with:
-          name: notebook-${{ steps.get-dir.outputs.artifact_name }}
-          path: ${{ matrix.notebook }}
-          retention-days: 1
-
-      - name: Upload static files
-        uses: actions/upload-artifact@v4
-        if: always()
-        with:
-          name: static-${{ steps.get-dir.outputs.artifact_name }}
-          path: ${{ steps.get-dir.outputs.dir }}/static/
-          if-no-files-found: ignore
-          retention-days: 1
-
-      - name: Upload solutions
-        uses: actions/upload-artifact@v4
-        if: always()
-        with:
-          name: solutions-${{ steps.get-dir.outputs.artifact_name }}
-          path: ${{ steps.get-dir.outputs.dir }}/solutions/
+          name: tutorial-${{ steps.get-dir.outputs.artifact_name }}
+          path: |
+            ${{ matrix.notebook }}
+            ${{ steps.get-dir.outputs.dir }}/static/${{ steps.get-dir.outputs.nb_name }}*
+            ${{ steps.get-dir.outputs.dir }}/solutions/${{ steps.get-dir.outputs.nb_name }}*
+            ${{ steps.get-dir.outputs.dir }}/student/${{ steps.get-dir.outputs.nb_name }}*
+            ${{ steps.get-dir.outputs.dir }}/instructor/${{ steps.get-dir.outputs.nb_name }}*
           if-no-files-found: ignore
           retention-days: 1
 
@@ -212,53 +202,26 @@ jobs:
         run: |
           echo "Restoring processed files from artifacts..."
 
-          # Process notebook artifacts
-          for dir in artifacts/notebook-*; do
+          # Artifact name format: tutorial-tutorials___W1D5_Microcircuits___W1D5_Tutorial2
+          # The path is encoded with ___ as delimiter. Last segment is notebook name, rest is directory.
+          # upload-artifact strips common path prefixes, so we need to reconstruct the target directory.
+          for dir in artifacts/tutorial-*; do
             if [ -d "$dir" ]; then
-              echo "Processing $dir"
-              cp -v "$dir"/*.ipynb tutorials/ 2>/dev/null || true
-              # Find the actual notebook and copy to correct location
-              find "$dir" -name "*.ipynb" -exec sh -c '
-                for f; do
-                  # Extract original path from artifact structure
-                  rel_path=$(basename "$f")
-                  # Find where this notebook should go based on its name
-                  original=$(find tutorials -name "$rel_path" -type f 2>/dev/null | head -1)
-                  if [ -n "$original" ]; then
-                    cp -v "$f" "$original"
-                  fi
-                done
-              ' sh {} +
+              # Extract artifact name and parse it to get target directory
+              artifact_name=$(basename "$dir" | sed 's/^tutorial-//')
+              # Convert ___ to / to get full path, then extract directory
+              full_path=$(echo "$artifact_name" | sed 's|___|/|g')
+              tutorial_dir=$(dirname "$full_path")
+              echo "Restoring from artifact $(basename "$dir") to: $tutorial_dir"
+
+              # Copy all files to the correct tutorial directory
+              # The artifact contains files with stripped prefixes (e.g., W1D5_Tutorial2.ipynb, student/, solutions/)
+              mkdir -p "$tutorial_dir"
+              cp -rv "$dir"/* "$tutorial_dir/" 2>/dev/null || true
             fi
           done
 
-          # Process static artifacts
-          for dir in artifacts/static-*; do
-            if [ -d "$dir" ]; then
-              echo "Processing static: $dir"
-              # Extract tutorial path from artifact name
-              artifact_name=$(basename "$dir" | sed 's/^static-//')
-              # Convert back: tutorials_W1D1_xxx.ipynb -> tutorials/W1D1_xxx
-              tutorial_dir=$(echo "$artifact_name" | sed 's/_/\//g' | sed 's/\.ipynb$//' | xargs dirname)
-              if [ -d "$tutorial_dir" ]; then
-                mkdir -p "$tutorial_dir/static"
-                cp -rv "$dir"/* "$tutorial_dir/static/" 2>/dev/null || true
-              fi
-            fi
-          done
-
-          # Process solutions artifacts
-          for dir in artifacts/solutions-*; do
-            if [ -d "$dir" ]; then
-              echo "Processing solutions: $dir"
-              artifact_name=$(basename "$dir" | sed 's/^solutions-//')
-              tutorial_dir=$(echo "$artifact_name" | sed 's/_/\//g' | sed 's/\.ipynb$//' | xargs dirname)
-              if [ -d "$tutorial_dir" ]; then
-                mkdir -p "$tutorial_dir/solutions"
-                cp -rv "$dir"/* "$tutorial_dir/solutions/" 2>/dev/null || true
-              fi
-            fi
-          done
+          echo "Restore complete."
 
       - name: Verify exercises
         env:
@@ -285,13 +248,18 @@ jobs:
           python ci/find_unreferenced_content.py > to_remove.txt
           if [ -s to_remove.txt ]; then git rm --pathspec-from-file=to_remove.txt; fi
 
+      - name: Clean up artifacts directory
+        run: rm -rf artifacts/
+
       - name: Commit post-processed files
         run: |
           git config --local user.email "action@github.com"
           git config --local user.name "GitHub Action"
           git add '**/*.ipynb'
           git add '**/static/*.png'
           git add '**/solutions/*.py'
+          git add '**/student/*.ipynb'
+          git add '**/instructor/*.ipynb'
           git add '**/README.md'
           git diff-index --quiet HEAD || git commit -m "Process tutorial notebooks"
 

requirements.txt

@@ -32,4 +32,4 @@ git+https://github.com/neuromatch/GNS-Modeling#egg=gns
 git+https://github.com/neuromatch/pyBPL#egg=pybpl
 git+https://github.com/neuromatch/MotorNet#egg=motornet
 git+https://github.com/ctn-waterloo/sspspace@neuromatch#egg=sspspace
-git+https://github.com/mitchellostrow/DSA#egg=DSA
+git+https://github.com/mitchellostrow/DSA#egg=dsa-metric

tutorials/W1D1_Generalization/W1D1_Tutorial1.ipynb

@@ -1563,7 +1563,7 @@
     "# Section 3: Dissecting TrOCR\n",
     "\n",
     "TrOCR is a model that performs printed optical character recognition and handwriting transcription using the transformer model. But what's inside of it?\n",
-    "It's important to note here that the original transformer model consisted of an encoder step, following by a decoder step. Taken together, this was the initial Transformer model of Vaswani et al. However, subsequent research into transformers led researchers to find applications of the encoding step specifically (encoding models like BERT) and also specific applications of the decoder step (autoregressive models like GPT). This meant that the terminology then changed to be *encoder transformers* and *decoder/causal/autoregressive transformers*. TrOCR is an example of the original transformer setup (both an encoder step and decoder step joined together). The image below outlines this setup. This also matches the transformer architecture given in the video above."
+    "It's important to note here that the original transformer model consisted of an encoder step, followed by a decoder step. Taken together, this was the initial Transformer model of Vaswani et al. However, subsequent research into transformers led researchers to find applications of the encoding step specifically (encoding models like BERT) and also specific applications of the decoder step (autoregressive models like GPT). This meant that the terminology then changed to be *encoder transformers* and *decoder/causal/autoregressive transformers*. TrOCR is an example of the original transformer setup (both an encoder step and decoder step joined together). The image below outlines this setup. This also matches the transformer architecture given in the video above."
    ]
   },
   {
@@ -2600,7 +2600,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.22"
+   "version": "3.10.19"
   }
  },
  "nbformat": 4,

tutorials/W1D1_Generalization/W1D1_Tutorial3.ipynb

@@ -187,7 +187,7 @@
     "logging.getLogger('matplotlib.font_manager').disabled = True\n",
     "\n",
     "%matplotlib inline\n",
-    "%config InlineBackend.figure_format = 'retina' # perfrom high definition rendering for images and plots\n",
+    "%config InlineBackend.figure_format = 'retina' # perform high-definition rendering for images and plots\n",
     "plt.style.use(\"https://raw.githubusercontent.com/NeuromatchAcademy/course-content/main/nma.mplstyle\")"
    ]
   },
@@ -620,7 +620,7 @@
     "\n",
     "Let's put ourselves in the mindset of a cognitive scientist studying handwriting. We're interested in how people learn to recognize new characters. Indeed, humans display low **sample complexity** when learning new visual concepts: they seem to grasp new concepts with very few presentations, generalizing effortlessly. In AI, learning from $k$ labeled examples is known as $k$-shot learning; one-shot and few-shot learning refer to learning from one or a few labeled examples.\n",
     "\n",
-    "A good dataset to investigate one-shot learning is the Omniglot dataset. Omniglot has sometimes been described as *MNIST, transposed*. Instead of **thousands** of examples from **10** digit classes, Omniglot consists of **20** instances from **1623** character classes. These character classes are sourced from 50 alphabets, both natural (e.g. Cherokee or Greek) and constructed (e.g. the alien alphabet from the TV show Futurama). \n",
+    "A good dataset to investigate one-shot learning is the Omniglot dataset. Omniglot has sometimes been described as *MNIST, transposed*. Instead of **thousands** of examples from **10** digit classes (many examples, few classes), Omniglot consists of **20** instances from **1623** character classes (few examples, many classes). These character classes are sourced from 50 alphabets, both natural (e.g. Cherokee or Greek) and constructed (e.g. the alien alphabet from the TV show Futurama). \n",
     "\n",
     "![Sample characters from the Omniglot dataset](https://github.com/brendenlake/omniglot/raw/master/omniglot_grid.jpg)\n",
     "\n",
@@ -992,7 +992,7 @@
    "name": "python3"
   },
   "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
+   "display_name": "Python 3",
    "language": "python",
    "name": "python3"
   },
@@ -1006,7 +1006,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.22"
+   "version": "3.10.19"
   }
  },
  "nbformat": 4,

tutorials/W1D1_Generalization/instructor/W1D1_Tutorial1.ipynb

@@ -1569,7 +1569,7 @@
     "# Section 3: Dissecting TrOCR\n",
     "\n",
     "TrOCR is a model that performs printed optical character recognition and handwriting transcription using the transformer model. But what's inside of it?\n",
-    "It's important to note here that the original transformer model consisted of an encoder step, following by a decoder step. Taken together, this was the initial Transformer model of Vaswani et al. However, subsequent research into transformers led researchers to find applications of the encoding step specifically (encoding models like BERT) and also specific applications of the decoder step (autoregressive models like GPT). This meant that the terminology then changed to be *encoder transformers* and *decoder/causal/autoregressive transformers*. TrOCR is an example of the original transformer setup (both an encoder step and decoder step joined together). The image below outlines this setup. This also matches the transformer architecture given in the video above."
+    "It's important to note here that the original transformer model consisted of an encoder step, followed by a decoder step. Taken together, this was the initial Transformer model of Vaswani et al. However, subsequent research into transformers led researchers to find applications of the encoding step specifically (encoding models like BERT) and also specific applications of the decoder step (autoregressive models like GPT). This meant that the terminology then changed to be *encoder transformers* and *decoder/causal/autoregressive transformers*. TrOCR is an example of the original transformer setup (both an encoder step and decoder step joined together). The image below outlines this setup. This also matches the transformer architecture given in the video above."
    ]
   },
   {
@@ -2608,7 +2608,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.22"
+   "version": "3.10.19"
   }
  },
  "nbformat": 4,

tutorials/W1D1_Generalization/instructor/W1D1_Tutorial3.ipynb

@@ -187,7 +187,7 @@
     "logging.getLogger('matplotlib.font_manager').disabled = True\n",
     "\n",
     "%matplotlib inline\n",
-    "%config InlineBackend.figure_format = 'retina' # perfrom high definition rendering for images and plots\n",
+    "%config InlineBackend.figure_format = 'retina' # perform high-definition rendering for images and plots\n",
     "plt.style.use(\"https://raw.githubusercontent.com/NeuromatchAcademy/course-content/main/nma.mplstyle\")"
    ]
   },
@@ -620,7 +620,7 @@
     "\n",
     "Let's put ourselves in the mindset of a cognitive scientist studying handwriting. We're interested in how people learn to recognize new characters. Indeed, humans display low **sample complexity** when learning new visual concepts: they seem to grasp new concepts with very few presentations, generalizing effortlessly. In AI, learning from $k$ labeled examples is known as $k$-shot learning; one-shot and few-shot learning refer to learning from one or a few labeled examples.\n",
     "\n",
-    "A good dataset to investigate one-shot learning is the Omniglot dataset. Omniglot has sometimes been described as *MNIST, transposed*. Instead of **thousands** of examples from **10** digit classes, Omniglot consists of **20** instances from **1623** character classes. These character classes are sourced from 50 alphabets, both natural (e.g. Cherokee or Greek) and constructed (e.g. the alien alphabet from the TV show Futurama). \n",
+    "A good dataset to investigate one-shot learning is the Omniglot dataset. Omniglot has sometimes been described as *MNIST, transposed*. Instead of **thousands** of examples from **10** digit classes (many examples, few classes), Omniglot consists of **20** instances from **1623** character classes (few examples, many classes). These character classes are sourced from 50 alphabets, both natural (e.g. Cherokee or Greek) and constructed (e.g. the alien alphabet from the TV show Futurama). \n",
     "\n",
     "![Sample characters from the Omniglot dataset](https://github.com/brendenlake/omniglot/raw/master/omniglot_grid.jpg)\n",
     "\n",
@@ -992,7 +992,7 @@
    "name": "python3"
   },
   "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
+   "display_name": "Python 3",
    "language": "python",
    "name": "python3"
   },
@@ -1006,7 +1006,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.22"
+   "version": "3.10.19"
   }
  },
  "nbformat": 4,

tutorials/W1D1_Generalization/student/W1D1_Tutorial1.ipynb

@@ -1460,7 +1460,7 @@
     "# Section 3: Dissecting TrOCR\n",
     "\n",
     "TrOCR is a model that performs printed optical character recognition and handwriting transcription using the transformer model. But what's inside of it?\n",
-    "It's important to note here that the original transformer model consisted of an encoder step, following by a decoder step. Taken together, this was the initial Transformer model of Vaswani et al. However, subsequent research into transformers led researchers to find applications of the encoding step specifically (encoding models like BERT) and also specific applications of the decoder step (autoregressive models like GPT). This meant that the terminology then changed to be *encoder transformers* and *decoder/causal/autoregressive transformers*. TrOCR is an example of the original transformer setup (both an encoder step and decoder step joined together). The image below outlines this setup. This also matches the transformer architecture given in the video above."
+    "It's important to note here that the original transformer model consisted of an encoder step, followed by a decoder step. Taken together, this was the initial Transformer model of Vaswani et al. However, subsequent research into transformers led researchers to find applications of the encoding step specifically (encoding models like BERT) and also specific applications of the decoder step (autoregressive models like GPT). This meant that the terminology then changed to be *encoder transformers* and *decoder/causal/autoregressive transformers*. TrOCR is an example of the original transformer setup (both an encoder step and decoder step joined together). The image below outlines this setup. This also matches the transformer architecture given in the video above."
    ]
   },
   {
@@ -2474,7 +2474,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.22"
+   "version": "3.10.19"
   }
  },
  "nbformat": 4,