updating pytorch to zero_grad for leaf variables. Adjusting pytorch.rfft to pytorch.fft.rfft

boazcogan · boazcogan · commit 920868115142 · 2021-05-04T21:18:40.000-05:00
diff --git a/book/training/building_blocks.ipynb b/book/training/building_blocks.ipynb
@@ -1606,14 +1606,13 @@
     "fig, ax = plt.subplots(8, 8, figsize=(12, 10))\n",
     "ax = ax.flatten()\n",
     "\n",
-    "fourier_basis = torch.rfft(\n",
-    "    torch.eye(filter_length), \n",
-    "    1, onesided=True\n",
+    "fourier_basis = torch.fft.rfft(\n",
+    "    torch.eye(filter_length)\n",
     ")\n",
     "cutoff = 1 + filter_length // 2\n",
     "fourier_basis = torch.cat([\n",
-    "    fourier_basis[:, :cutoff, 0],\n",
-    "    fourier_basis[:, :cutoff, 1]\n",
+    "    torch.real(fourier_basis[:, :cutoff]),\n",
+    "    torch.imag(fourier_basis[:, :cutoff])\n",
     "], dim=1)\n",
     "fourier_basis = fourier_basis.float()\n",
     "\n",
@@ -1729,7 +1728,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.8.5"
+   "version": "3.8.8"
   }
  },
  "nbformat": 4,
diff --git a/book/training/gradient_descent.ipynb b/book/training/gradient_descent.ipynb
@@ -365,13 +365,14 @@
     "    possible_b.shape[0]\n",
     "))\n",
     "\n",
-    "for i, m_hat in enumerate(possible_m):\n",
-    "    for j, b_hat in enumerate(possible_b):\n",
-    "        line.layer.weight[0, 0] = m_hat\n",
-    "        line.layer.bias[0] = b_hat\n",
-    "        y_hat = line(to_tensor(x))\n",
-    "        _loss = (y_hat - to_tensor(y)).abs().mean()\n",
-    "        loss[i, j] = _loss\n",
+    "with torch.no_grad():\n",
+    "    for i, m_hat in enumerate(possible_m):\n",
+    "        for j, b_hat in enumerate(possible_b):\n",
+    "            line.layer.weight[0, 0] = m_hat\n",
+    "            line.layer.bias[0] = b_hat\n",
+    "            y_hat = line(to_tensor(x))\n",
+    "            _loss = (y_hat - to_tensor(y)).abs().mean()\n",
+    "            loss[i, j] = _loss\n",
     "\n",
     "plt.title('2D Visualization of Loss Landscape')\n",
     "plt.pcolormesh(possible_m, possible_b, loss, shading='auto')\n",
@@ -442,10 +443,11 @@
     }
    ],
    "source": [
-    "line.layer.weight[0, 0] = m_hat\n",
-    "line.layer.bias[0] = b_hat\n",
+    "with torch.no_grad():\n",
+    "    line.layer.weight[0, 0] = m_hat\n",
+    "    line.layer.bias[0] = b_hat\n",
     "\n",
-    "y_hat = line(to_tensor(x))\n",
+    "    y_hat = line(to_tensor(x))\n",
     "\n",
     "plt.title(\"Training data + network predictions\")\n",
     "plt.scatter(x, y, label='Training data')\n",
@@ -1014,7 +1016,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.8.5"
+   "version": "3.8.8"
   }
  },
  "nbformat": 4,
