Compute energy for interpolation in ambient space

Author: Atcold

10-autoencoder.ipynb

@@ -14,7 +14,9 @@
     "from torch.utils.data import DataLoader\n",
     "from torchvision import transforms\n",
     "from torchvision.datasets import MNIST\n",
-    "from matplotlib import pyplot as plt"
+    "from matplotlib import pyplot as plt\n",
+    "from res.plot_lib import set_default\n",
+    "set_default(figsize=(15, 4))"
    ]
   },
   {
@@ -145,9 +147,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "scrolled": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "# Train standard or denoising autoencoder (AE)\n",
@@ -176,9 +176,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "scrolled": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "# Visualise a few kernels of the encoder\n",
@@ -247,9 +245,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "scrolled": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "# Train standard or denoising autoencoder (AE)\n",
@@ -336,6 +332,15 @@
     "    display_images(TELEA, NS)"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Experimenting\n",
+    "\n",
+    "The section below needs some more love and refactorin"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -351,7 +356,8 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "y = (img[1:2] + img[15:16])/2\n",
+    "A, B = 0, 9\n",
+    "y = (img[[A]] + img[[B]])/2\n",
     "with torch.no_grad():\n",
     "    ỹ = model(y)\n",
     "plt.figure(figsize=(10,5))\n",
@@ -360,11 +366,28 @@
     "plt.subplot(122)\n",
     "plt.imshow(to_img(ỹ).squeeze())"
    ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "N = 16\n",
+    "samples = torch.Tensor(N, 28 * 28).to(device)\n",
+    "for i in range(N):\n",
+    "    samples[i] = i / (N - 1) * img[B].data + (1 - i / (N - 1) ) * img[A].data\n",
+    "with torch.no_grad():\n",
+    "    reconstructions = model(samples)[0]\n",
+    "\n",
+    "plt.title(f'{A = }, {B = }')\n",
+    "plt.plot(samples.sub(reconstructions).pow(2).sum(dim=(1)), '-o')"
+   ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "Python 3",
+   "display_name": "Python 3 (ipykernel)",
    "language": "python",
    "name": "python3"
   },
@@ -378,7 +401,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.8.3"
+   "version": "3.8.13"
   }
  },
  "nbformat": 4,

11-VAE.ipynb

@@ -12,7 +12,8 @@
     "from torch.utils.data import DataLoader\n",
     "from torchvision import transforms\n",
     "from torchvision.datasets import MNIST\n",
-    "from matplotlib import pyplot as plt"
+    "from matplotlib import pyplot as plt\n",
+    "from res.plot_lib import set_default"
    ]
   },
   {
@@ -23,7 +24,7 @@
    "source": [
     "# Displaying routine\n",
     "\n",
-    "def display_images(in_, out, n=1, label=None, count=False):\n",
+    "def display_images(in_, out, n=1, label='', count=False, energy=None):\n",
     "    for N in range(n):\n",
     "        if in_ is not None:\n",
     "            in_pic = in_.data.cpu().view(-1, 28, 28)\n",
@@ -39,7 +40,9 @@
     "            plt.subplot(1,4,i+1)\n",
     "            plt.imshow(out_pic[i+4*N])\n",
     "            plt.axis('off')\n",
-    "            if count: plt.title(str(4 * N + i), color='w')"
+    "            c = 4 * N + i\n",
+    "            if count: plt.title(str(c), color='w')\n",
+    "            if count and energy is not None: plt.title(f'{c}, e={energy[c].item():.2f}', color='w')\n"
    ]
   },
   {
@@ -166,9 +169,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "scrolled": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "# Training and testing the VAE\n",
@@ -238,8 +239,10 @@
    "outputs": [],
    "source": [
     "# Display last test batch\n",
-    "\n",
-    "display_images(None, y, 4, count=True)"
+    "with torch.no_grad():\n",
+    "    ỹ = model(y)[0].view(-1, 28, 28)\n",
+    "energy = y.squeeze().sub(ỹ).pow(2).sum(dim=(1,2))\n",
+    "display_images(None, y, 4, count=True, energy=energy)"
    ]
   },
   {
@@ -250,14 +253,18 @@
    "source": [
     "# Choose starting and ending point for the interpolation -> shows original and reconstructed\n",
     "\n",
-    "A, B = 1, 14\n",
+    "A, B = 0, 6\n",
     "sample = model.decoder(torch.stack((mu[A].data, mu[B].data), 0))\n",
     "display_images(None, torch.stack(((\n",
     "    y[A].data.view(-1),\n",
     "    y[B].data.view(-1),\n",
     "    sample.data[0],\n",
-    "    sample.data[1]\n",
-    ")), 0))"
+    "    sample.data[1],\n",
+    "    sample.data[0],\n",
+    "    sample.data[1],\n",
+    "    y[A].data.view(-1) - sample.data[0],\n",
+    "    y[B].data.view(-1) - sample.data[1]\n",
+    ")), 0), 2)"
    ]
   },
   {
@@ -269,13 +276,13 @@
     "# Perform an interpolation between input A and B, in N steps\n",
     "\n",
     "N = 16\n",
-    "code = torch.Tensor(N, 20).to(device)\n",
-    "sample = torch.Tensor(N, 28, 28).to(device)\n",
+    "# code = torch.Tensor(N, 20).to(device)\n",
+    "samples = torch.Tensor(N, 28, 28).to(device)\n",
     "for i in range(N):\n",
-    "    #code[i] = i / (N - 1) * mu[B].data + (1 - i / (N - 1) ) * mu[A].data\n",
-    "    sample[i] = i / (N - 1) * y[B].data + (1 - i / (N - 1) ) * y[A].data\n",
-    "# sample = model.decoder(code)\n",
-    "display_images(None, sample, N // 4, count=True)"
+    "    # code[i] = i / (N - 1) * mu[B].data + (1 - i / (N - 1) ) * mu[A].data\n",
+    "    samples[i] = i / (N - 1) * y[B].data + (1 - i / (N - 1) ) * y[A].data\n",
+    "# samples = model.decoder(code)\n",
+    "display_images(None, samples, N // 4, count=True)"
    ]
   },
   {
@@ -288,10 +295,27 @@
     "with torch.no_grad():\n",
     "    ỹ = model(ẏ)[0]\n",
     "plt.figure(figsize=(10,5))\n",
-    "plt.subplot(121)\n",
-    "plt.imshow((ẏ).view(28, 28))\n",
-    "plt.subplot(122)\n",
-    "plt.imshow((ỹ).view(28, 28))"
+    "plt.subplot(121), plt.imshow((ẏ).view(28, 28))\n",
+    "plt.subplot(122), plt.imshow((ỹ).view(28, 28))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "tags": []
+   },
+   "outputs": [],
+   "source": [
+    "N = 16\n",
+    "samples = torch.Tensor(N, 28, 28).to(device)\n",
+    "for i in range(N):\n",
+    "    samples[i] = i / (N - 1) * y[B].data + (1 - i / (N - 1) ) * y[A].data\n",
+    "with torch.no_grad():\n",
+    "    reconstructions = model(samples)[0].view(-1, 28, 28)\n",
+    "\n",
+    "plt.title(f'{A = }, {B = }')\n",
+    "plt.plot(samples.sub(reconstructions).pow(2).sum(dim=(1,2)), '-o')"
    ]
   },
   {
@@ -301,8 +325,7 @@
    "outputs": [],
    "source": [
     "import numpy as np\n",
-    "from sklearn.manifold import TSNE\n",
-    "from res.plot_lib import set_default"
+    "from sklearn.manifold import TSNE"
    ]
   },
   {
@@ -343,18 +366,11 @@
     "    a[i].axis('equal')\n",
     "f.colorbar(s, ax=a[:], ticks=np.arange(10), boundaries=np.arange(11) - .5)"
    ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": []
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "Python 3",
+   "display_name": "Python 3 (ipykernel)",
    "language": "python",
    "name": "python3"
   },
@@ -368,7 +384,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.8.3"
+   "version": "3.8.13"
   }
  },
  "nbformat": 4,