Merge pull request #187 from dataquestio/mike-dl2-gp-solution-file-fix

Update Mission784Solutions.ipynb

Mission784Solutions.ipynb

@@ -199,7 +199,7 @@
     },
     {
       "cell_type": "code",
-      "execution_count": 16,
+      "execution_count": 6,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/",
@@ -251,7 +251,7 @@
     },
     {
       "cell_type": "code",
-      "execution_count": 17,
+      "execution_count": 8,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/",
@@ -538,7 +538,6 @@
         "- Trying different combinations of data augmentation layers.\n",
         "- Increasing the number of filters, trying different sizes, different strides, and number of convolutional layers.\n",
         "- Adding more fully-connected layers with a different number of units.\n",
-        "- Adding batch normalization to all conv layers. Batch normalization before the max pooling layer resulted in a drop in performance so it was moved to after the max pooling layer.\n",
         "- Removing batch normalization layers after the fully-connected layers and replacing them with dropout layers. Before this change, the model's performance was a bit erratic.\n",
         "- Reducing the learning rate helped the model to converge.\n",
         "- Adding an early stopping callback that would stop the training if the validation loss didn't decrease for three consecutive epochs. Using the `restore_best_weights` argument ensured that the model would use weights from the epoch with the lowest validation loss only."
@@ -819,4 +818,4 @@
   },
   "nbformat": 4,
   "nbformat_minor": 0
-}
+}