fixed and optimized GPU-enabled codepath. Now accuracy is computed on…

… GPU and is transferred back to CPU only for printing. Execution became somewhat faster.
cmudeeplearning11785 · Oct 1, 2018 · 4dd0c48 · 4dd0c48
1 parent 2e0bf48
commit 4dd0c48
Showing 1 changed file with 4 additions and 6 deletions.
diff --git a/recitation-2/pytorch_example.ipynb b/recitation-2/pytorch_example.ipynb
@@ -163,18 +163,16 @@
     "        if i%100==0:\n",
     "            print(\"At iteration\",i)\n",
     "            # compute the accuracy of the prediction\n",
-    "            train_prediction = train_output.cpu().detach().argmax(dim=1)\n",
-    "            train_accuracy = (train_prediction.numpy()==train_labels.numpy()).mean() \n",
+    "            train_accuracy = (train_labels.eq(train_output.argmax(dim=1))).float().mean() \n",
     "            # Now for the validation set\n",
     "            val_output = net(val_data)\n",
     "            val_loss = criterion(val_output,val_labels)\n",
     "            # compute the accuracy of the prediction\n",
-    "            val_prediction = val_output.cpu().detach().argmax(dim=1)\n",
-    "            val_accuracy = (val_prediction.numpy()==val_labels.numpy()).mean() \n",
+    "            val_accuracy = (val_labels.eq(val_output.argmax(dim=1))).float().mean() \n",
     "            print(\"Training loss :\",train_loss.cpu().detach().numpy())\n",
-    "            print(\"Training accuracy :\",train_accuracy)\n",
+    "            print(\"Training accuracy :\",train_accuracy.cpu().detach().numpy())\n",
     "            print(\"Validation loss :\",val_loss.cpu().detach().numpy())\n",
-    "            print(\"Validation accuracy :\",val_accuracy)\n",
+    "            print(\"Validation accuracy :\",val_accuracy.cpu().detach().numpy())\n",
     "    \n",
     "    net = net.cpu()"
    ]