add further exercises

Author: Kucharssim

exercises/bayesflow-diffusion.ipynb

@@ -432,6 +432,15 @@
     "f=plt.xlim(-5, 5)\n",
     "f=plt.xlim(-5, 5)"
    ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Further exercise\n",
+    "\n",
+    "The accuracy of the approximation is generally driven by how expressive the velocity network is, how well it is trained, but also by the accuracy of the integrator. Play around with the network complexity of the flow matching model, simulation budget, and the steps taken during the ODE solver to see how will it affect your ability to generate the datasaurus distribution."
+   ]
   }
  ],
  "metadata": {

exercises/bayesflow-normal.ipynb

@@ -263,7 +263,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 7,
+   "execution_count": null,
    "metadata": {},
    "outputs": [],
    "source": [
@@ -279,14 +279,15 @@
     "\n",
     "    def __getitem__(self, index):\n",
     "        data, _ = make_swiss_roll(self.batch_size, noise=1)\n",
-    "        data = data[:,[0, -1]]\n",
-    "        condition = np.random.choice([1, -1], size=(batch_size, 2))\n",
-    "        data = condition * data\n",
+    "        data=data[:,[0, -1]]\n",
+    "        \n",
+    "        condition=np.random.choice([1, -1], size=(batch_size, 2))\n",
+    "        data=condition * data\n",
     "\n",
-    "        base= make_ring(data.shape[0])\n",
+    "        base=make_ring(data.shape[0])\n",
     "        \n",
-    "        t = np.random.uniform(low=0, high=1, size=data.shape[0])\n",
-    "        t = np.repeat(t[:,np.newaxis], repeats=data.shape[1], axis=1)\n",
+    "        t=np.random.uniform(low=0, high=1, size=data.shape[0])\n",
+    "        t=np.repeat(t[:,np.newaxis], repeats=data.shape[1], axis=1)\n",
     "\n",
     "        target = data - base\n",
     "        return dict(x_0=base, x_1=data, t=t, condition=condition), target"
@@ -438,6 +439,16 @@
     "        axs[i,j].set_title(\"x scale: {}, y scale {}\".format(x_scale, y_scale))\n",
     "fig.tight_layout()"
    ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Further exercises\n",
+    "\n",
+    "1. Here we only scaled the data by 2 values (-1 or 1) in two directions ($x$ and $y$). However, there is nothing that can stop us from using different values. Try to replace the line `condition=np.random.choice([1, -1], size=(batch_size, 2))` with some other transformation (for example, generate values from a uniform distribution between -1 and 1). What do you think the network will learn? Try it for yourself.\n",
+    "2. The swiss roll distribution generates 3D data. In this exercise, we only used 2 of the axes and neglected the third one. Try to change to model so that you can actually reproduce the swiss role in 3D (note: you will also need to change the base distribution to be 3D)."
+   ]
   }
  ],
  "metadata": {

exercises/flow-matching-datasaurus.ipynb

@@ -704,7 +704,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.11.10"
+   "version": "3.11.11"
   }
  },
  "nbformat": 4,