detecting_loanwords/detecting_loanwords_keras.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Detecting loadwords with Keras"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### In this notebook you will:\n",
    "1. Train a neural network using Keras.\n",
    "1. Optimize hyperparameters for the Neural Net.\n",
    "1. Record model training statistics for viewing with tensor board.\n",
    "1. Save the best model using checkpoints. \n",
    "1. Compare the performance between the Neural Net and the Decision Tree classifier."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Imports"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[]"
      ]
     },
     "execution_count": 53,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import logging\n",
    "import random\n",
    "import pickle\n",
    "\n",
    "from keras.models import Sequential\n",
    "from keras.layers import Dense, Dropout\n",
    "from keras.wrappers.scikit_learn import KerasClassifier\n",
    "from sklearn.preprocessing import StandardScaler\n",
    "from keras.callbacks import EarlyStopping, TensorBoard, ModelCheckpoint\n",
    "from keras.models import load_model\n",
    "from keras import backend as K\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.utils.extmath import density\n",
    "from cltk.stem.latin.j_v import JVReplacer\n",
    "from cltk.prosody.latin.scansion_constants import ScansionConstants\n",
    "from cltk.prosody.latin.string_utils import remove_punctuation_dict\n",
    "from cltk.tokenize.word import WordTokenizer\n",
    "from cltk.corpus.readers import get_corpus_reader\n",
    "from cltk.utils.featurization import word_to_features\n",
    "from cltk.utils.file_operations import md5\n",
    "\n",
    "plt.style.use('fivethirtyeight')\n",
    "import logging\n",
    "logging.disable(logging.WARNING)\n",
    "K.tensorflow_backend._get_available_gpus()\n",
    "logging.disable(logging.NOTSET)\n",
    "K.tensorflow_backend._get_available_gpus()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Fix random seed values for reproducibility and tuning\n",
    "see: https://keras.io/getting-started/faq/#how-can-i-obtain-reproducible-results-using-keras-during-development"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# The below is necessary for starting Numpy generated random numbers\n",
    "# in a well-defined initial state.\n",
    "np.random.seed(12)\n",
    "\n",
    "# The below is necessary for starting core Python generated random numbers\n",
    "# in a well-defined state.\n",
    "random.seed(12) "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Add parent directory to path so we can access our common code"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import sys\n",
    "import inspect\n",
    "currentdir = os.path.dirname(os.path.abspath(inspect.getfile(inspect.currentframe())))\n",
    "parentdir = os.path.dirname(currentdir)\n",
    "sys.path.insert(0, parentdir) "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "from mlyoucanuse.featurize_text_fun import word_to_features"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Turn on logging, primarily so that library methods may report warnings, if any"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "LOG = logging.getLogger('make_model')\n",
    "LOG.addHandler(logging.NullHandler())\n",
    "logging.basicConfig(level=logging.INFO)\n",
    "# tf.logging.set_verbosity(tf.logging.ERROR) # silence a deprecation warning; todo revisit"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load the saved X,y matrix \n",
    "as created in the `detect_loanwords_problems_solutions` notebook"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([1., 1., 1., ..., 0., 0., 0.], dtype=float32)"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "all_X = None\n",
    "with open('all_X.pkl', 'rb') as pickle_file:\n",
    "    all_X = pickle.load(pickle_file)\n",
    "all_y = None\n",
    "with open('all_y', 'rb') as pickle_file:\n",
    "    all_y = pickle.load(pickle_file)\n",
    "all_X.astype(np.float32)\n",
    "all_y.astype(np.float32)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Process the saved matrix"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "X_train, X_test, y_train, y_test = train_test_split(all_X, all_y, test_size=0.3, random_state=42)\n",
    "# Partition equal sizes of test and validation sets\n",
    "X_test, X_validation, y_test, y_validation = train_test_split(X_test, y_test, test_size=0.5, random_state=42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "scaler = StandardScaler()\n",
    "# try:  all_X.astype(np.float64)\n",
    "# Fit scaler only on training slice of data; \n",
    "# As with all the transformations, it is important to fit the scalers to the training data only,\n",
    "# not to the full dataset (including the test set).\n",
    "scaler.fit(X_train.astype(np.float32))\n",
    "# TODO pickle this scaler for applying the same transforms to input\n",
    "with open('scaler.pkl', 'wb') as fwriter:\n",
    "    pickle.dump(scaler, fwriter)    \n",
    "# Fit all 3 data sets\n",
    "X_train = scaler.transform(X_train.astype(np.float32))\n",
    "X_test = scaler.transform(X_test.astype(np.float32))\n",
    "X_validation = scaler.transform(X_validation.astype(np.float32))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Create Keras callbacks for Checkpointing, Tensor board, and Early Stopping\n",
    "Early stopping is promising, but it did not work as well as expected for this data set. Elsewhere and with large batches, it can be invaluable."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "early_stopping = EarlyStopping(monitor='val_loss',\n",
    "                               min_delta=0,\n",
    "                               patience=3,\n",
    "                               verbose=0,\n",
    "                               mode='auto',\n",
    "                               baseline=None,\n",
    "                               restore_best_weights=False)\n",
    "tensor_board = TensorBoard(log_dir='./logs',\n",
    "                           histogram_freq=0,                           \n",
    "                           write_graph=True,\n",
    "                           write_grads=False,\n",
    "                           write_images=False,\n",
    "                           embeddings_freq=0,\n",
    "                           embeddings_layer_names=None,\n",
    "                           embeddings_metadata=None,\n",
    "                           embeddings_data=None,\n",
    "                           update_freq='epoch')\n",
    "# checkpoint\n",
    "filepath=\"weights.best.hdf5\"\n",
    "checkpoint = ModelCheckpoint(filepath, monitor='val_acc', verbose=1, save_best_only=True, mode='max')\n",
    "callbacks_list = [checkpoint, tensor_board]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Create a function to plot the history of the model training and validation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "def plot_history(history, filename=None):\n",
    "    \"\"\"Plot Keras metrics.\"\"\"\n",
    "    # Note, the key types can be reported differently, hence we use defaults on lookup\n",
    "    acc = history.history.get('acc', history.history.get('accuracy'))\n",
    "    val_acc = history.history.get('val_acc', history.history.get('val_accuracy')) \n",
    "    loss = history.history['loss']\n",
    "    val_loss = history.history['val_loss']\n",
    "    x = range(1, len(acc) + 1)\n",
    "    plt.figure(figsize=(12, 5))\n",
    "    plt.subplot(1, 2, 1)\n",
    "    plt.plot(x, acc, 'b', label='Training acc')\n",
    "    plt.plot(x, val_acc, 'r', label='Validation acc')\n",
    "    plt.title('Training and validation accuracy')\n",
    "    plt.xlabel('Epochs')\n",
    "    plt.legend()\n",
    "    plt.subplot(1, 2, 2)\n",
    "    plt.plot(x, loss, 'b', label='Training loss')\n",
    "    plt.plot(x, val_loss, 'r', label='Validation loss')\n",
    "    plt.title('Training and validation loss')\n",
    "    plt.xlabel('Epochs')\n",
    "    plt.legend()\n",
    "    if filename:\n",
    "        plt.savefig(filename)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Create Keras Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"Loanword detector\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "Dense_layer_1 (Dense)        (None, 256)               7424      \n",
      "_________________________________________________________________\n",
      "Dense_layer_2 (Dense)        (None, 256)               65792     \n",
      "_________________________________________________________________\n",
      "Dense_layer_3 (Dense)        (None, 1)                 257       \n",
      "=================================================================\n",
      "Total params: 73,473\n",
      "Trainable params: 73,473\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "num_features = all_X.shape[1]   \n",
    "\n",
    "model = Sequential(name='Loanword detector')\n",
    "model.add(Dense(256, name='Dense_layer_1',\n",
    "                input_dim=num_features, activation='relu'))\n",
    "model.add(Dense(256, name='Dense_layer_2',\n",
    "                activation='relu'))\n",
    "model.add(Dense(1, name='Dense_layer_3', activation='sigmoid'))\n",
    "model.compile(loss='binary_crossentropy', \n",
    "              optimizer= 'adam',\n",
    "              metrics=['accuracy'])  \n",
    "model.summary()\n",
    "# Also experiemented with Kernel regularizers and dropout but neither added improvements"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 153563 samples, validate on 32907 samples\n",
      "Epoch 1/40\n",
      "153563/153563 [==============================] - 4s 24us/step - loss: 0.4373 - accuracy: 0.8197 - val_loss: 0.4063 - val_accuracy: 0.8288\n",
      "Epoch 2/40\n",
      "153563/153563 [==============================] - 4s 23us/step - loss: 0.3877 - accuracy: 0.8324 - val_loss: 0.3777 - val_accuracy: 0.8384\n",
      "Epoch 3/40\n",
      "153563/153563 [==============================] - 3s 23us/step - loss: 0.3627 - accuracy: 0.8443 - val_loss: 0.3607 - val_accuracy: 0.8446\n",
      "Epoch 4/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.3436 - accuracy: 0.8519 - val_loss: 0.3336 - val_accuracy: 0.8581\n",
      "Epoch 5/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.3275 - accuracy: 0.8587 - val_loss: 0.3307 - val_accuracy: 0.8588\n",
      "Epoch 6/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.3138 - accuracy: 0.8655 - val_loss: 0.3093 - val_accuracy: 0.8657\n",
      "Epoch 7/40\n",
      "153563/153563 [==============================] - 3s 23us/step - loss: 0.3024 - accuracy: 0.8711 - val_loss: 0.3005 - val_accuracy: 0.8732\n",
      "Epoch 8/40\n",
      "153563/153563 [==============================] - 3s 23us/step - loss: 0.2925 - accuracy: 0.8759 - val_loss: 0.2986 - val_accuracy: 0.8749\n",
      "Epoch 9/40\n",
      "153563/153563 [==============================] - 4s 23us/step - loss: 0.2873 - accuracy: 0.8796 - val_loss: 0.2920 - val_accuracy: 0.8750\n",
      "Epoch 10/40\n",
      "153563/153563 [==============================] - 3s 23us/step - loss: 0.2749 - accuracy: 0.8844 - val_loss: 0.2889 - val_accuracy: 0.8785\n",
      "Epoch 11/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2670 - accuracy: 0.8878 - val_loss: 0.2787 - val_accuracy: 0.8824\n",
      "Epoch 12/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2608 - accuracy: 0.8906 - val_loss: 0.2776 - val_accuracy: 0.8841\n",
      "Epoch 13/40\n",
      "153563/153563 [==============================] - 3s 23us/step - loss: 0.2572 - accuracy: 0.8934 - val_loss: 0.2717 - val_accuracy: 0.8869\n",
      "Epoch 14/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2491 - accuracy: 0.8956 - val_loss: 0.2649 - val_accuracy: 0.8899\n",
      "Epoch 15/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2421 - accuracy: 0.8992 - val_loss: 0.2652 - val_accuracy: 0.8921\n",
      "Epoch 16/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2378 - accuracy: 0.9015 - val_loss: 0.2563 - val_accuracy: 0.8950\n",
      "Epoch 17/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2315 - accuracy: 0.9039 - val_loss: 0.2621 - val_accuracy: 0.8923\n",
      "Epoch 18/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2317 - accuracy: 0.9064 - val_loss: 0.2537 - val_accuracy: 0.8954\n",
      "Epoch 19/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2219 - accuracy: 0.9083 - val_loss: 0.2502 - val_accuracy: 0.8985\n",
      "Epoch 20/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2177 - accuracy: 0.9104 - val_loss: 0.2510 - val_accuracy: 0.8983\n",
      "Epoch 21/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2143 - accuracy: 0.9115 - val_loss: 0.2422 - val_accuracy: 0.9012\n",
      "Epoch 22/40\n",
      "153563/153563 [==============================] - 3s 23us/step - loss: 0.2104 - accuracy: 0.9133 - val_loss: 0.2517 - val_accuracy: 0.9003\n",
      "Epoch 23/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2069 - accuracy: 0.9145 - val_loss: 0.2463 - val_accuracy: 0.9028\n",
      "Epoch 24/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.2062 - accuracy: 0.9157 - val_loss: 0.2432 - val_accuracy: 0.9014\n",
      "Epoch 25/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1990 - accuracy: 0.9180 - val_loss: 0.2357 - val_accuracy: 0.9045\n",
      "Epoch 26/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1961 - accuracy: 0.9194 - val_loss: 0.2454 - val_accuracy: 0.9067\n",
      "Epoch 27/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1941 - accuracy: 0.9202 - val_loss: 0.2347 - val_accuracy: 0.9079\n",
      "Epoch 28/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1908 - accuracy: 0.9215 - val_loss: 0.2433 - val_accuracy: 0.9045\n",
      "Epoch 29/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1905 - accuracy: 0.9233 - val_loss: 0.2417 - val_accuracy: 0.9059\n",
      "Epoch 30/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1866 - accuracy: 0.9237 - val_loss: 0.2356 - val_accuracy: 0.9075\n",
      "Epoch 31/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1826 - accuracy: 0.9252 - val_loss: 0.2358 - val_accuracy: 0.9091\n",
      "Epoch 32/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1805 - accuracy: 0.9262 - val_loss: 0.2370 - val_accuracy: 0.9070\n",
      "Epoch 33/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1767 - accuracy: 0.9277 - val_loss: 0.2381 - val_accuracy: 0.9083\n",
      "Epoch 34/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1762 - accuracy: 0.9274 - val_loss: 0.2306 - val_accuracy: 0.9107\n",
      "Epoch 35/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1729 - accuracy: 0.9296 - val_loss: 0.2312 - val_accuracy: 0.9090\n",
      "Epoch 36/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1803 - accuracy: 0.9302 - val_loss: 0.2299 - val_accuracy: 0.9093\n",
      "Epoch 37/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1681 - accuracy: 0.9313 - val_loss: 0.2297 - val_accuracy: 0.9124\n",
      "Epoch 38/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1668 - accuracy: 0.9324 - val_loss: 0.2311 - val_accuracy: 0.9109\n",
      "Epoch 39/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1655 - accuracy: 0.9323 - val_loss: 0.2336 - val_accuracy: 0.9074\n",
      "Epoch 40/40\n",
      "153563/153563 [==============================] - 3s 22us/step - loss: 0.1631 - accuracy: 0.9334 - val_loss: 0.2322 - val_accuracy: 0.9107\n"
     ]
    }
   ],
   "source": [
    " history = model.fit(X_train, y_train,\n",
    "                     epochs=40,\n",
    "                     verbose=True,\n",
    "                     validation_data=(X_test, y_test),\n",
    "                     batch_size=32,\n",
    "                     callbacks=callbacks_list)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "153563/153563 [==============================] - 1s 7us/step\n",
      "Training Accuracy: 0.9373 Loss: 0.1521\n",
      "32907/32907 [==============================] - 0s 7us/step\n",
      "Testing Accuracy: 0.9107 Loss: 0.2322\n",
      "32907/32907 [==============================] - 0s 7us/step\n",
      "Unseen Accuracy: 0.9120 Loss: 0.4213\n"
     ]
    }
   ],
   "source": [
    "loss, accuracy  = model.evaluate(X_train, y_train, verbose=True)\n",
    "print(f\"Training Accuracy: {accuracy:.4f} Loss: {loss:.4f}\")\n",
    "loss, accuracy = model.evaluate(X_test, y_test, verbose=True)\n",
    "print(f\"Testing Accuracy: {accuracy:.4f} Loss: {loss:.4f}\")\n",
    "loss, accuracy = model.evaluate(X_validation, y_validation, verbose=True)\n",
    "print(f\"Unseen Accuracy: {accuracy:.4f} Loss: {loss:.4f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Sample training runs for tuning\n",
    "blindly running grid search on random parameter variations isn't usually very helpful. \n",
    "Here we show some obvious variations to narrow the parameter ranges and tune manually."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 128, 64\n",
    "\n",
    "# 153563/153563 [==============================] - 2s 16us/step\n",
    "# Training Accuracy: 0.9110 Loss: 0.2175\n",
    "# 32907/32907 [==============================] - 1s 17us/step\n",
    "# Testing Accuracy: 0.8991 Loss: 0.2483\n",
    "# 32907/32907 [==============================] - 1s 17us/step\n",
    "# Unseen Accuracy: 0.8974 Loss: 0.4771\n",
    "\n",
    "#128, 128\n",
    "# 153563/153563 [==============================] - 2s 16us/step\n",
    "# Training Accuracy: 0.9110 Loss: 0.2175\n",
    "# 32907/32907 [==============================] - 1s 17us/step\n",
    "# Testing Accuracy: 0.8991 Loss: 0.2483\n",
    "# 32907/32907 [==============================] - 1s 17us/step\n",
    "# Unseen Accuracy: 0.8974 Loss: 0.4771\n",
    "\n",
    "# 256, 256\n",
    "# 153563/153563 [==============================] - 2s 15us/step\n",
    "# Training Accuracy: 0.9397 Loss: 0.1494\n",
    "# 32907/32907 [==============================] - 1s 16us/step\n",
    "# Testing Accuracy: 0.9162 Loss: 0.2208\n",
    "# 32907/32907 [==============================] - 0s 15us/step\n",
    "# Unseen Accuracy: 0.9139 Loss: 0.5172\n",
    "\n",
    "# 256, 256 w/regularization\n",
    "# 153563/153563 [==============================] - 3s 18us/step\n",
    "# Training Accuracy: 0.8370 Loss: 0.3993\n",
    "# 32907/32907 [==============================] - 1s 18us/step\n",
    "# Testing Accuracy: 0.8358 Loss: 0.3997\n",
    "# 32907/32907 [==============================] - 1s 17us/step\n",
    "# Unseen Accuracy: 0.8365 Loss: 0.3970\n",
    "\n",
    "# 256,256 with droput\n",
    "# 153563/153563 [==============================] - 2s 16us/step\n",
    "# Training Accuracy: 0.9246 Loss: 0.1851\n",
    "# 32907/32907 [==============================] - 1s 16us/step\n",
    "# Testing Accuracy: 0.9094 Loss: 0.2239\n",
    "# 32907/32907 [==============================] - 1s 16us/step\n",
    "# Unseen Accuracy: 0.9086 Loss: 0.2397\n",
    "\n",
    "#256, 128 w/dropout\n",
    "# 153563/153563 [==============================] - 3s 18us/step\n",
    "# Training Accuracy: 0.9123 Loss: 0.2111\n",
    "# 32907/32907 [==============================] - 1s 18us/step\n",
    "# Testing Accuracy: 0.8999 Loss: 0.2410\n",
    "# 32907/32907 [==============================] - 1s 18us/step\n",
    "# Unseen Accuracy: 0.9019 Loss: 0.2815\n",
    "\n",
    "#256, 128\n",
    "# 153563/153563 [==============================] - 3s 17us/step\n",
    "# Training Accuracy: 0.9280 Loss: 0.1736\n",
    "# 32907/32907 [==============================] - 1s 17us/step\n",
    "# Testing Accuracy: 0.9079 Loss: 0.2321\n",
    "# 32907/32907 [==============================] - 1s 16us/step\n",
    "# Unseen Accuracy: 0.9080 Loss: 0.3344\n",
    "\n",
    "# 256,256 using a different seed\n",
    "# 153563/153563 [==============================] - 1s 7us/step\n",
    "# Training Accuracy: 0.9373 Loss: 0.1521\n",
    "# 32907/32907 [==============================] - 0s 7us/step\n",
    "# Testing Accuracy: 0.9107 Loss: 0.2322\n",
    "# 32907/32907 [==============================] - 0s 7us/step\n",
    "# Unseen Accuracy: 0.9120 Loss: 0.4213\n",
    "\n",
    "#500, 500 using the 256 seed above\n",
    "# 153563/153563 [==============================] - 1s 9us/step\n",
    "# Training Accuracy: 0.9453 Loss: 0.1326\n",
    "# 32907/32907 [==============================] - 0s 10us/step\n",
    "# Testing Accuracy: 0.9178 Loss: 0.2263\n",
    "# 32907/32907 [==============================] - 0s 9us/step\n",
    "# Unseen Accuracy: 0.9150 Loss: 0.2291\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 864x360 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_history(history, 'loanword_detectory_val_acc.256x256nn.png')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Tips for parameter tuning\n",
    "* Set a seed for random on the Python and library levels, as done early in this notebook\n",
    "* Be sure to recreate and compile your model fresh with each training session;\n",
    "    * If you rerun the .fit() method, you’ll start off with the computed weights from the previous training.\n",
    "* Study the history plots of each run, notice where training and validation characteristically starts to diverge more and more.\n",
    "    * Your training and validation graphs may have local minima, so you might not be able to use early stopping automatically.\n",
    "* Keep your validation and test (unseen) scores close; if they diverge, you're overfitting on the training and validation data."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Fit final model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [],
   "source": [
    "all_X_scaled = scaler.fit_transform(all_X.astype(np.float32))\n",
    "\n",
    "model = Sequential(name='Loanword detector')\n",
    "model.add(Dense(256, name='Dense_layer_1',\n",
    "                input_dim=num_features, activation='relu'))\n",
    "model.add(Dense(256, name='Dense_layer_2',\n",
    "                activation='relu'))\n",
    "model.add(Dense(1, name='Dense_layer_3', activation='sigmoid'))\n",
    "model.compile(loss='binary_crossentropy', \n",
    "              optimizer= 'adam',\n",
    "              metrics=['accuracy'])  \n",
    "model.fit(all_X_scaled, all_y,\n",
    "                     epochs=26,\n",
    "                     verbose=False,                     \n",
    "                     batch_size=32)\n",
    "model.save('loanword_detector.greek_latin.keras.h5')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load the model and call it a run time (don't use pickle)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "mymodel = load_model('loanword_detector.greek_latin.keras.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [],
   "source": [
    "the_char_idx = {' ': 0, 'A': 1, 'B': 2, 'C': 3, 'D': 4, 'E': 5, 'F': 6, 'G': 7, 'H': 8, 'I': 9, 'K': 10, 'L': 11, 'M': 12, 'N': 13, 'O': 14, 'P': 15, 'Q': 16, 'R': 17, 'S': 18, 'T': 19, 'U': 20, 'V': 21, 'X': 22, 'Y': 23, 'Z': 24, 'a': 25, 'b': 26, 'c': 27, 'd': 28, 'e': 29, 'f': 30, 'g': 31, 'h': 32, 'i': 33, 'k': 34, 'l': 35, 'm': 36, 'n': 37, 'o': 38, 'p': 39, 'q': 40, 'r': 41, 's': 42, 't': 43, 'u': 44, 'v': 45, 'x': 46, 'y': 47, 'z': 48, 'Å': 49, 'Ē': 50, 'ē': 51, 'Ō': 52, 'ō': 53}\n",
    "\n",
    "def encode_word(word, char_idx, max_len=28):\n",
    "    vals =np.zeros(max_len)\n",
    "    for idx, car in enumerate(word):\n",
    "        vals[idx] = char_idx.get(car, 0) / len(char_idx) # total_uniq_char_count\n",
    "    return vals"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "def is_greek(word):\n",
    "    x = [encode_word(word, the_char_idx, max_len=28)]\n",
    "    # Use the fitted transformer, load from a pickle if necessary\n",
    "    x=np.array(x).astype(np.float32)\n",
    "    x = scaler.transform(x) \n",
    "    return mymodel.predict(x)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0.1272271]], dtype=float32)"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "is_greek('arxomai') "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0.15047942]], dtype=float32)"
      ]
     },
     "execution_count": 44,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "is_greek('veritas')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0.9263797]], dtype=float32)"
      ]
     },
     "execution_count": 45,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "is_greek('thuon')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "-rw-r--r-- 1 todd todd 889K Mar 29 14:13 loanword_detector.greek_latin.keras.h5\r\n"
     ]
    }
   ],
   "source": [
    "! ls -lh  loanword_detector.greek_latin.keras.h5"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Comparing the scikit learn classifier vs. the Keras Neural Net\n",
    "* RandomForest Classifier\n",
    "    * 626 MB pickle size\n",
    "    * accuracy: 96%\n",
    "* Keras Neural Net\n",
    "    * 889 kb h5 pickle size\n",
    "    * accuracy: 92.88%\n",
    "\n",
    "These results are typical for a relatively small dataset. Decision tree classifiers do very well, but can be prone to overfitting; and on the downside, the decision tree model is somewhat large. The neural net is small and is not overfitting. It is more compact and quite performant, and should a worthwhile consideration.\n",
    "\n",
    "Both models might benefit from training with more data, and this is an area for future investigation.\n",
    "\n",
    "## When to use which model?\n",
    "* trying to win a Kaggle competition, choose the highest scoring one, even if it overfits ;-)\n",
    "* need to deploy at scale? Consider sacrificing a few percentage points for a smaller model size for:\n",
    "    * smaller memory footprint\n",
    "    * more throughput"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# That's all for now folks!"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}