diff --git a/jupyter/truncated_svd.ipynb b/jupyter/truncated_svd.ipynb
index f1ee0d26944751ae05c8ff25830480642b810267..7331d3c54b228f57364c1d27a3c17fbf9b94ecc3 100644
--- a/jupyter/truncated_svd.ipynb
+++ b/jupyter/truncated_svd.ipynb
@@ -147,6 +147,7 @@
     "\n",
     "\n",
     "# See Faster-RCNN: https://github.com/rbgirshick/py-faster-rcnn/blob/master/tools/compress_net.py\n",
+    "# Replaced numpy operations with pytorch operations (so that we can leverage the GPU).\n",
     "def truncated_svd(W, l):\n",
     "    \"\"\"Compress the weight matrix W of an inner product (fully connected) layer\n",
     "    using truncated SVD.\n",
@@ -157,15 +158,14 @@
     "    Ul, L: matrices such that W \\approx Ul*L\n",
     "    \"\"\"\n",
     "\n",
-    "    U, s, V = np.linalg.svd(W, full_matrices=False)\n",
+    "    U, s, V = torch.svd(W, some=True)\n",
     "\n",
     "    Ul = U[:, :l]\n",
     "    sl = s[:l]\n",
+    "    V = V.t()\n",
     "    Vl = V[:l, :]\n",
     "\n",
-    "    SV = np.dot(np.diag(sl), Vl)\n",
-    "    SV = torch.from_numpy(SV).cuda()\n",
-    "    Ul = torch.from_numpy(Ul).cuda()\n",
+    "    SV = torch.mm(torch.diag(sl), Vl)\n",
     "    return Ul, SV\n",
     "\n",
     "\n",
@@ -174,7 +174,7 @@
     "        super(TruncatedSVD,self).__init__()\n",
     "        self.replaced_gemm = replaced_gemm\n",
     "        print(\"W = {}\".format(gemm_weights.shape))\n",
-    "        self.U, self.SV = truncated_svd(gemm_weights.cpu(), int(0.4 * gemm_weights.size(0)))\n",
+    "        self.U, self.SV = truncated_svd(gemm_weights.data, int(0.4 * gemm_weights.size(0)))\n",
     "        print(\"U = {}\".format(self.U.shape))\n",
     "        \n",
     "        self.fc_u = nn.Linear(self.U.size(1), self.U.size(0)).cuda()\n",
@@ -214,7 +214,7 @@
     "import time\n",
     "import torchnet.meter as tnt\n",
     "t0 = time.time()\n",
-    "test_loader = imagenet_load_data(\"../../datasets/imagenet\", \n",
+    "test_loader = imagenet_load_data(\"../../data.imagenet\", \n",
     "                                 batch_size=64, \n",
     "                                 num_workers=4,\n",
     "                                 shuffle=False)\n",
@@ -228,20 +228,11 @@
     "        outputs = resnet50(inputs)\n",
     "        classerr.add(outputs.data, target)\n",
     "        if (validation_step+1) % 100 == 0:\n",
-    "            print((validation_step+1) * 512)\n",
+    "            print(\"progress: %d images\" % ((validation_step+1) * 512))\n",
     "        \n",
-    "print(classerr.value(1), classerr.value(5))\n",
+    "print(\"Top1: %.2f   Top5: %.2f\" % (classerr.value(1), classerr.value(5)))\n",
     "t2 = time.time()\n",
-    "print(t2-t0)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    ""
+    "print(\"Duration: \", t2-t0)"
    ]
   }
  ],
@@ -254,16 +245,16 @@
   "language_info": {
    "codemirror_mode": {
     "name": "ipython",
-    "version": 3.0
+    "version": 3
    },
    "file_extension": ".py",
    "mimetype": "text/x-python",
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.5.2"
+   "version": "3.6.7"
   }
  },
  "nbformat": 4,
- "nbformat_minor": 0
-}
\ No newline at end of file
+ "nbformat_minor": 1
+}