Automate head merging

.github/workflows/continuous_integration.yml

@@ -18,7 +18,7 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@v3
       - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
       - name: Unit Tests

README.md

@@ -2,7 +2,7 @@
 
 OxonFair is an expressive toolkit designed to enforce a wide-range of fairness definitions and to customize binary classifier behavior.
 The toolkit is designed to overcome a range of shortcomings in existing fairness toolkits for high-capacity models that overfit to the training data.
-It is designed and works for computer vision and NLP problems alongside tabular data.
+Unlike other toolkits it is designed and works for computer vision and NLP problems alongside tabular data.
 
 Check out the colab demo [here](https://colab.research.google.com/drive/1CfcS3AX7M2MO1wW33wU1LDiY5DwtyyxH?usp=sharing) or read the [preprint](https://arxiv.org/abs/2407.13710).
 
@@ -115,6 +115,16 @@ OxonFair is build from the ground up to avoid these issues. It is a postprocessi
 
 That said, we make several additional design decisions which we believe make for a better experience for data scientists:
 
+#### Direct support for pytorch including NLP and Computer Vision
+
+See [here](./examples/pytorch_minimal_demo.ipynb) for an example.
+In brief, the steps are:
+
+1. Train your network.
+2. Call `DeepFairPredictor` with the network output on validation data.
+3. Call `fit` to enforce fairness. Use `evaluate_groups` and `plot_frontier` to explore trade-offs.
+4. Use `merge_heads_pytorch` to generate a fair network.  
+
 #### Fine-grained control of behavior
 
 ##### Wide Choice of performance measure

examples/high-dim_fairlearn_comparision.ipynb

@@ -1451,7 +1451,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.10.13"
+   "version": "3.10.14"
   }
  },
  "nbformat": 4,

examples/multi_group_fairlearn_comparision.ipynb

@@ -28,7 +28,7 @@
      "output_type": "stream",
      "text": [
       "Training time of xgboost without fairness\n",
-      "0.8215227920000006\n"
+      "0.6205382919999991\n"
      ]
     }
    ],
@@ -142,8 +142,8 @@
     "         sensitive_features=train_data['groups'])\n",
     "    stop=time.perf_counter()\n",
     "    score=evaluate_fairness(test_data['target'], fair.predict(test_data['data']), test_data['groups'], metrics={'a':gm.accuracy,'b':gm.demographic_parity},verbose=False)\n",
-    "    fairl[0,i]=score[0]['a']#['updated']\n",
-    "    fairl[1,i]=score[0]['b']#['updated']\n",
+    "    fairl[0,i]=score['a']#['updated']\n",
+    "    fairl[1,i]=score['b']#['updated']\n",
     "    fairl[2,i]=stop-start\n",
     "    train_data['groups'][train_data['groups']==name]=' Other'\n",
     "    test_data['groups'][test_data['groups']==name]=' Other'"
@@ -224,71 +224,71 @@
        "      <td>OxonFair</td>\n",
        "      <td>0.868619</td>\n",
        "      <td>0.019085</td>\n",
-       "      <td>44.246316</td>\n",
+       "      <td>41.136526</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>0</th>\n",
        "      <td>FairLearn</td>\n",
        "      <td>0.865001</td>\n",
        "      <td>0.037732</td>\n",
-       "      <td>46.683562</td>\n",
+       "      <td>46.391113</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>1</th>\n",
        "      <td>OxonFair</td>\n",
        "      <td>0.867663</td>\n",
        "      <td>0.012305</td>\n",
-       "      <td>0.816369</td>\n",
+       "      <td>0.804408</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>1</th>\n",
        "      <td>FairLearn</td>\n",
        "      <td>0.866639</td>\n",
        "      <td>0.016227</td>\n",
-       "      <td>28.097114</td>\n",
+       "      <td>27.712763</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>2</th>\n",
        "      <td>OxonFair</td>\n",
        "      <td>0.867936</td>\n",
        "      <td>0.021317</td>\n",
-       "      <td>0.067256</td>\n",
+       "      <td>0.071277</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>2</th>\n",
        "      <td>FairLearn</td>\n",
        "      <td>0.865070</td>\n",
        "      <td>0.007104</td>\n",
-       "      <td>24.851400</td>\n",
+       "      <td>25.013719</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>3</th>\n",
        "      <td>OxonFair</td>\n",
        "      <td>0.868823</td>\n",
        "      <td>0.003093</td>\n",
-       "      <td>0.053665</td>\n",
+       "      <td>0.049922</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>3</th>\n",
        "      <td>FairLearn</td>\n",
        "      <td>0.869165</td>\n",
        "      <td>0.002346</td>\n",
-       "      <td>20.321553</td>\n",
+       "      <td>20.080540</td>\n",
        "    </tr>\n",
        "  </tbody>\n",
        "</table>\n",
        "</div>"
       ],
       "text/plain": [
        "        Name  Accuracy  Demographic Parity       Time\n",
-       "0   OxonFair  0.868619            0.019085  44.246316\n",
-       "0  FairLearn  0.865001            0.037732  46.683562\n",
-       "1   OxonFair  0.867663            0.012305   0.816369\n",
-       "1  FairLearn  0.866639            0.016227  28.097114\n",
-       "2   OxonFair  0.867936            0.021317   0.067256\n",
-       "2  FairLearn  0.865070            0.007104  24.851400\n",
-       "3   OxonFair  0.868823            0.003093   0.053665\n",
-       "3  FairLearn  0.869165            0.002346  20.321553"
+       "0   OxonFair  0.868619            0.019085  41.136526\n",
+       "0  FairLearn  0.865001            0.037732  46.391113\n",
+       "1   OxonFair  0.867663            0.012305   0.804408\n",
+       "1  FairLearn  0.866639            0.016227  27.712763\n",
+       "2   OxonFair  0.867936            0.021317   0.071277\n",
+       "2  FairLearn  0.865070            0.007104  25.013719\n",
+       "3   OxonFair  0.868823            0.003093   0.049922\n",
+       "3  FairLearn  0.869165            0.002346  20.080540"
       ]
      },
      "execution_count": 7,
@@ -299,44 +299,6 @@
    "source": [
     "results"
    ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "\\begin{tabular}{llrrr}\n",
-      "\\toprule\n",
-      " & Name & Accuracy & Demographic Parity & Time \\\\\n",
-      "\\midrule\n",
-      "0 & OxonFair & 0.868619 & 0.019085 & 44.246316 \\\\\n",
-      "0 & FairLearn & 0.865001 & 0.037732 & 46.683562 \\\\\n",
-      "1 & OxonFair & 0.867663 & 0.012305 & 0.816369 \\\\\n",
-      "1 & FairLearn & 0.866639 & 0.016227 & 28.097114 \\\\\n",
-      "2 & OxonFair & 0.867936 & 0.021317 & 0.067256 \\\\\n",
-      "2 & FairLearn & 0.865070 & 0.007104 & 24.851400 \\\\\n",
-      "3 & OxonFair & 0.868823 & 0.003093 & 0.053665 \\\\\n",
-      "3 & FairLearn & 0.869165 & 0.002346 & 20.321553 \\\\\n",
-      "\\bottomrule\n",
-      "\\end{tabular}\n",
-      "\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(results.to_latex())"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": []
   }
  ],
  "metadata": {