{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# Recurrent Neural Networks" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import pandas as pd\n", "import numpy as np\n", "%matplotlib inline\n", "import matplotlib.pyplot as plt" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Time series forecasting" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from pandas.tseries.offsets import MonthEnd\n", "\n", "df = pd.read_csv('../data/cansim-0800020-eng-6674700030567901031.csv',\n", " skiprows=6, skipfooter=9,\n", " engine='python')\n", "\n", "df['Adjustments'] = pd.to_datetime(df['Adjustments']) + MonthEnd(1)\n", "df = df.set_index('Adjustments')\n", "df.head()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "split_date = pd.Timestamp('01-01-2011')" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "train = df.loc[:split_date, ['Unadjusted']]\n", "test = df.loc[split_date:, ['Unadjusted']]" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from sklearn.preprocessing import MinMaxScaler\n", "\n", "sc = MinMaxScaler()\n", "\n", "train_sc = sc.fit_transform(train)\n", "test_sc = sc.transform(test)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "train_sc_df = pd.DataFrame(train_sc, columns=['Scaled'], index=train.index)\n", "test_sc_df = pd.DataFrame(test_sc, columns=['Scaled'], index=test.index)\n", "\n", "for s in range(1, 13):\n", " train_sc_df['shift_{}'.format(s)] = train_sc_df['Scaled'].shift(s)\n", " test_sc_df['shift_{}'.format(s)] = test_sc_df['Scaled'].shift(s)\n", "\n", "X_train = train_sc_df.dropna().drop('Scaled', axis=1)\n", "y_train = train_sc_df.dropna()[['Scaled']]\n", "\n", "X_test = test_sc_df.dropna().drop('Scaled', axis=1)\n", "y_test = test_sc_df.dropna()[['Scaled']]\n", "\n", "X_train = X_train.values\n", "X_test= X_test.values\n", "\n", "y_train = y_train.values\n", "y_test = y_test.values" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "X_train.shape" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Exercise 1\n", "\n", "In the model above we reshaped the input shape to: `(num_samples, 1, 12)`, i.e. we treated a window of 12 months as a vector of 12 coordinates that we simultaneously passed to all the LSTM nodes. An alternative way to look at the problem is to reshape the input to `(num_samples, 12, 1)`. This means we consider each input window as a sequence of 12 values that we will pass in sequence to the LSTM. In principle this looks like a more accurate description of our situation. But does it yield better predictions? Let's check it.\n", "\n", "- Reshape `X_train` and `X_test` so that they represent a set of univariate sequences\n", "- retrain the same LSTM(6) model, you'll have to adapt the `input_shape`\n", "- check the performance of this new model, is it better at predicting the test data?" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "X_train_t = X_train.reshape(X_train.shape[0], 12, 1)\n", "X_test_t = X_test.reshape(X_test.shape[0], 12, 1)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "X_train_t.shape" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from tensorflow.keras.models import Sequential\n", "from tensorflow.keras.layers import LSTM, Dense\n", "import tensorflow.keras.backend as K\n", "from tensorflow.keras.callbacks import EarlyStopping" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "K.clear_session()\n", "model = Sequential()\n", "\n", "model.add(LSTM(6, input_shape=(12, 1)))\n", "\n", "model.add(Dense(1))\n", "\n", "model.compile(loss='mean_squared_error', optimizer='adam')" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "model.summary()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "early_stop = EarlyStopping(monitor='loss', patience=1, verbose=1)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "model.fit(X_train_t, y_train, epochs=600,\n", " batch_size=32, verbose=0)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "y_pred = model.predict(X_test_t)\n", "plt.plot(y_test)\n", "plt.plot(y_pred)" ] }, { "cell_type": "markdown", "metadata": { "collapsed": true }, "source": [ "## Exercise 2\n", "\n", "RNN models can be applied to images too. In general we can apply them to any data where there's a connnection between nearby units. Let's see how we can easily build a model that works with images.\n", "\n", "- Load the MNIST data, by now you should be able to do it blindfolded :)\n", "- reshape it so that an image looks like a long sequence of pixels\n", "- create a recurrent model and train it on the training data\n", "- how does it perform compared to a fully connected? How does it compare to Convolutional Neural Networks?\n", "\n", "(feel free to run this exercise on a cloud GPU if it's too slow on your laptop)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from tensorflow.keras.datasets import mnist\n", "from tensorflow.keras.utils import to_categorical" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "(X_train, y_train), (X_test, y_test) = mnist.load_data()\n", "X_train = X_train.astype('float32') / 255.0\n", "X_test = X_test.astype('float32') / 255.0\n", "y_train_cat = to_categorical(y_train, 10)\n", "y_test_cat = to_categorical(y_test, 10)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "X_train = X_train.reshape(X_train.shape[0], -1, 1)\n", "X_test = X_test.reshape(X_test.shape[0], -1, 1)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "print(X_train.shape)\n", "print(X_test.shape)\n", "print(y_train_cat.shape)\n", "print(y_test_cat.shape)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "# define the model\n", "K.clear_session()\n", "model = Sequential()\n", "model.add(LSTM(32, input_shape=X_train.shape[1:]))\n", "model.add(Dense(10, activation='softmax'))\n", "\n", "# compile the model\n", "model.compile(loss='categorical_crossentropy',\n", " optimizer='rmsprop',\n", " metrics=['accuracy'])\n", "\n", "model.fit(X_train, y_train_cat,\n", " batch_size=32,\n", " epochs=100,\n", " validation_split=0.3,\n", " shuffle=True,\n", " verbose=2,\n", " )\n", "\n", "model.evaluate(X_test, y_test_cat)" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] } ], "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": "3.7.10" } }, "nbformat": 4, "nbformat_minor": 2 }