Part 2b - Instrumental vocal detection.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Instrumental, Genre and Mood Detection in Music with Deep Learning\n",
    "\n",
    "This tutorial shows how different Convolutional Neural Network architectures are used for:\n",
    "* Instrumental vs. Vocal Detection:  detecting whether a piece of music is instrumental or contains vocals\n",
    "* Genre Classification\n",
    "* Mood Recognition\n",
    "\n",
    "The data set used is the [MagnaTagATune Dataset](http://mirg.city.ac.uk/codeapps/the-magnatagatune-dataset), but a smaller subset of it, with only 1 sample excerpt of each of the original audio files.\n",
    "\n",
    "It consists of 5405 files, each 30 seconds long. \n",
    "\n",
    "The annotations for this dataset contain a multitude of tags, including some that hint at whether the file is instrumental or vocal. (see [Create 2 classes from a list of tags](#Create-2-classes-from-a-list-of-tags) below)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Requirements\n",
    "\n",
    "* Python 3.5\n",
    "* Keras >= 2.1.1\n",
    "* Tensorflow\n",
    "* scikit-learn >= 0.18\n",
    "* Pandas\n",
    "* Librosa\n",
    "* MatplotLib"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Table of Contents\n",
    "\n",
    "This tutorial contains:\n",
    "* Loading and Preprocessing of Audio files\n",
    "* Loading class files from CSV and using Label Encoder\n",
    "* Audio Preprocessing: Generating log Mel spectrograms\n",
    "* Standardization of Data\n",
    "* Convolutional Neural Networks\n",
    "* Train/Test set split\n",
    "\n",
    "* Instrumental vs. Vocal Detection\n",
    "* Genre Classification\n",
    "* Mood Recognition\n",
    "\n",
    "You can execute the following code blocks by pressing SHIFT+Enter consecutively."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Download Data\n",
    "\n",
    "The (subsampled) data set can be downloaded from [here](https://owncloud.tuwien.ac.at/index.php/s/hivOGXKoUQtacbo).\n",
    "\n",
    "Please unzip it.\n",
    "\n",
    "Set the path to the unpacked folder in the next box:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "\n",
    "DATA_PATH = '/home/tlidy/data/MagnaTagATune'\n",
    "AUDIO_PATH = os.path.join(DATA_PATH, 'audio')\n",
    "META_PATH = os.path.join(DATA_PATH, 'metadata')\n",
    "\n",
    "#NPZ_FILE = '/home/tlidy/data/mel_spectrogram_segments_96x1366.npz'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "# General Imports\n",
    "\n",
    "import argparse\n",
    "import csv\n",
    "import datetime\n",
    "import glob\n",
    "import math\n",
    "import sys\n",
    "import time\n",
    "import numpy as np\n",
    "import pandas as pd # Pandas for reading CSV files and easier Data handling in preparation\n",
    "\n",
    "# Deep Learning\n",
    "\n",
    "import keras\n",
    "from keras.models import Sequential, Model\n",
    "from keras.layers import Input, Convolution2D, MaxPooling2D, Dense, Dropout, Activation, Flatten, merge\n",
    "from keras.layers.normalization import BatchNormalization\n",
    "from keras import optimizers\n",
    "from keras import backend as K\n",
    "\n",
    "from sklearn import preprocessing\n",
    "from sklearn.metrics import accuracy_score\n",
    "\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.model_selection import StratifiedShuffleSplit"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# local imports for audio reading and processing\n",
    "import audio_spectrogram as rp\n",
    "from audiofile_read import audiofile_read"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load the Metadata\n",
    "\n",
    "The tab-separated file contains pairs of filename TAB class."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>clip_id</th>\n",
       "      <th>no voice</th>\n",
       "      <th>singer</th>\n",
       "      <th>duet</th>\n",
       "      <th>plucking</th>\n",
       "      <th>hard rock</th>\n",
       "      <th>world</th>\n",
       "      <th>bongos</th>\n",
       "      <th>harpsichord</th>\n",
       "      <th>female singing</th>\n",
       "      <th>...</th>\n",
       "      <th>female singer</th>\n",
       "      <th>rap</th>\n",
       "      <th>metal</th>\n",
       "      <th>hip hop</th>\n",
       "      <th>quick</th>\n",
       "      <th>water</th>\n",
       "      <th>baroque</th>\n",
       "      <th>women</th>\n",
       "      <th>fiddle</th>\n",
       "      <th>english</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>mp3_path</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-01-cyclops-262-291.mp3</th>\n",
       "      <td>1119</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-02-all_seeing_eye-175-204.mp3</th>\n",
       "      <td>6021</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-03-black-175-204.mp3</th>\n",
       "      <td>11847</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-04-confusion_says-88-117.mp3</th>\n",
       "      <td>17119</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-05-the_beholder-291-320.mp3</th>\n",
       "      <td>25118</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-06-cyead-378-407.mp3</th>\n",
       "      <td>26533</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-07-telekonology-117-146.mp3</th>\n",
       "      <td>33637</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>0/american_bach_soloists-j_s__bach__cantatas_volume_v-01-gleichwie_der_regen_und_schnee_vom_himmel_fallt_bwv_18_i_sinfonia-117-146.mp3</th>\n",
       "      <td>29</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>0/american_bach_soloists-j_s__bach__cantatas_volume_v-02-gleichwie_der_regen_und_schnee_vom_himmel_fallt_bwv_18_ii_recitative__gleichwie_der_regen_und_schnee-30-59.mp3</th>\n",
       "      <td>5864</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>0/american_bach_soloists-j_s__bach__cantatas_volume_v-03-gleichwie_der_regen_und_schnee_vom_himmel_fallt_bwv_18_iii_recitative_and_litany__mein_gott_hier_wird_mein_herze_sein-88-117.mp3</th>\n",
       "      <td>11264</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>10 rows × 189 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                    clip_id  no voice  singer  \\\n",
       "mp3_path                                                                        \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...     1119         0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...     6021         0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    11847         0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    17119         0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    25118         0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    26533         0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    33637         0       0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...       29         0       0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...     5864         0       0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...    11264         0       0   \n",
       "\n",
       "                                                    duet  plucking  hard rock  \\\n",
       "mp3_path                                                                        \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...     0         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...     0         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...     0         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...     0         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...     0         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...     0         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...     0         0          0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...     0         0          0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...     0         0          0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...     0         0          0   \n",
       "\n",
       "                                                    world  bongos  \\\n",
       "mp3_path                                                            \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...      0       0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...      0       0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...      0       0   \n",
       "\n",
       "                                                    harpsichord  \\\n",
       "mp3_path                                                          \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...            0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...            0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...            0   \n",
       "\n",
       "                                                    female singing   ...     \\\n",
       "mp3_path                                                             ...      \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0   ...      \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0   ...      \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0   ...      \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0   ...      \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0   ...      \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0   ...      \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0   ...      \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...               0   ...      \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...               0   ...      \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...               0   ...      \n",
       "\n",
       "                                                    female singer  rap  metal  \\\n",
       "mp3_path                                                                        \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0    0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0    0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0    0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0    0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0    0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0    0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0    0      0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...              0    0      0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...              0    0      0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...              0    0      0   \n",
       "\n",
       "                                                    hip hop  quick  water  \\\n",
       "mp3_path                                                                    \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0      0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...        0      0      0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...        0      0      0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...        0      0      0   \n",
       "\n",
       "                                                    baroque  women  fiddle  \\\n",
       "mp3_path                                                                     \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0      0       0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...        0      0       0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...        0      0       0   \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...        0      0       0   \n",
       "\n",
       "                                                    english  \n",
       "mp3_path                                                     \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...        0  \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...        0  \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...        0  \n",
       "0/american_bach_soloists-j_s__bach__cantatas_vo...        0  \n",
       "\n",
       "[10 rows x 189 columns]"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "csv_file = os.path.join(META_PATH,'annotations_final_subsample.csv')\n",
    "\n",
    "# we select the last column (-1) as the index column (= filename)\n",
    "metadata = pd.read_csv(csv_file, index_col=0, sep='\\t')\n",
    "metadata.head(10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>no voice</th>\n",
       "      <th>singer</th>\n",
       "      <th>duet</th>\n",
       "      <th>plucking</th>\n",
       "      <th>hard rock</th>\n",
       "      <th>world</th>\n",
       "      <th>bongos</th>\n",
       "      <th>harpsichord</th>\n",
       "      <th>female singing</th>\n",
       "      <th>clasical</th>\n",
       "      <th>...</th>\n",
       "      <th>female singer</th>\n",
       "      <th>rap</th>\n",
       "      <th>metal</th>\n",
       "      <th>hip hop</th>\n",
       "      <th>quick</th>\n",
       "      <th>water</th>\n",
       "      <th>baroque</th>\n",
       "      <th>women</th>\n",
       "      <th>fiddle</th>\n",
       "      <th>english</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>mp3_path</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-01-cyclops-262-291.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-02-all_seeing_eye-175-204.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-03-black-175-204.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-04-confusion_says-88-117.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-05-the_beholder-291-320.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>5 rows × 188 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                    no voice  singer  duet  \\\n",
       "mp3_path                                                                     \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0       0     0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0       0     0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0       0     0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0       0     0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0       0     0   \n",
       "\n",
       "                                                    plucking  hard rock  \\\n",
       "mp3_path                                                                  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0          0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...         0          0   \n",
       "\n",
       "                                                    world  bongos  \\\n",
       "mp3_path                                                            \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0   \n",
       "\n",
       "                                                    harpsichord  \\\n",
       "mp3_path                                                          \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...            0   \n",
       "\n",
       "                                                    female singing  clasical  \\\n",
       "mp3_path                                                                       \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0         0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0         0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0         0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0         0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...               0         0   \n",
       "\n",
       "                                                     ...     female singer  \\\n",
       "mp3_path                                             ...                     \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...   ...                 0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...   ...                 0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...   ...                 0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...   ...                 0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...   ...                 0   \n",
       "\n",
       "                                                    rap  metal  hip hop  \\\n",
       "mp3_path                                                                  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    0      0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    0      0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    0      0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    0      0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...    0      0        0   \n",
       "\n",
       "                                                    quick  water  baroque  \\\n",
       "mp3_path                                                                    \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0        0   \n",
       "\n",
       "                                                    women  fiddle  english  \n",
       "mp3_path                                                                    \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0        0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0       0        0  \n",
       "\n",
       "[5 rows x 188 columns]"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# remove the unneeded column \"clip_id\"\n",
    "cols = \"clip_id\"\n",
    "metadata.drop(cols,axis=1,inplace=True)\n",
    "\n",
    "metadata.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 1) Instrumental vs. Vocal Detection"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Create 2 classes from a list of tags\n",
    "\n",
    "There are plenty of \"tags\" in this data set which hint at wether a track is \"vocal\" or \"instrumental\". We group these tags and finally come up with 1 boolean column saying whether a track is \"vocal\" or \"instrumental\"."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "tags_vocal = ['singer', 'female singing', 'female opera', 'male vocal', 'vocals', 'men', 'female', 'female voice', 'voice', 'male voice', 'girl', 'chanting', 'talking', 'choral', 'male singer', 'man singing', 'male opera', 'chant', 'man', 'female vocal', 'male vocals', 'vocal', 'woman', 'woman singing', 'singing', 'female vocals', 'voices', 'choir', 'female singer', 'women', 'choir', 'women']\n",
    "\n",
    "tags_instrumental = ['instrumental', 'no voice', 'no voices', 'no vocals', 'no vocal', 'no singing', 'no singer']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>singer</th>\n",
       "      <th>female singing</th>\n",
       "      <th>female opera</th>\n",
       "      <th>male vocal</th>\n",
       "      <th>vocals</th>\n",
       "      <th>men</th>\n",
       "      <th>female</th>\n",
       "      <th>female voice</th>\n",
       "      <th>voice</th>\n",
       "      <th>male voice</th>\n",
       "      <th>...</th>\n",
       "      <th>woman</th>\n",
       "      <th>woman singing</th>\n",
       "      <th>singing</th>\n",
       "      <th>female vocals</th>\n",
       "      <th>voices</th>\n",
       "      <th>choir</th>\n",
       "      <th>female singer</th>\n",
       "      <th>women</th>\n",
       "      <th>choir</th>\n",
       "      <th>women</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>mp3_path</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-01-cyclops-262-291.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-02-all_seeing_eye-175-204.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-03-black-175-204.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-04-confusion_says-88-117.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>d/ambient_teknology-the_all_seeing_eye_project-05-the_beholder-291-320.mp3</th>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>...</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>5 rows × 32 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                    singer  female singing  \\\n",
       "mp3_path                                                                     \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0               0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0               0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0               0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0               0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0               0   \n",
       "\n",
       "                                                    female opera  male vocal  \\\n",
       "mp3_path                                                                       \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0           0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0           0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0           0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0           0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0           0   \n",
       "\n",
       "                                                    vocals  men  female  \\\n",
       "mp3_path                                                                  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0    0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0    0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0    0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0    0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...       0    0       0   \n",
       "\n",
       "                                                    female voice  voice  \\\n",
       "mp3_path                                                                  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0      0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...             0      0   \n",
       "\n",
       "                                                    male voice  ...    woman  \\\n",
       "mp3_path                                                        ...            \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...           0  ...        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...           0  ...        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...           0  ...        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...           0  ...        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...           0  ...        0   \n",
       "\n",
       "                                                    woman singing  singing  \\\n",
       "mp3_path                                                                     \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0        0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0        0   \n",
       "\n",
       "                                                    female vocals  voices  \\\n",
       "mp3_path                                                                    \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0       0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...              0       0   \n",
       "\n",
       "                                                    choir  female singer  \\\n",
       "mp3_path                                                                   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0              0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0              0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0              0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0              0   \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0              0   \n",
       "\n",
       "                                                    women  choir  women  \n",
       "mp3_path                                                                 \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0      0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0      0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0      0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0      0  \n",
       "d/ambient_teknology-the_all_seeing_eye_project-...      0      0      0  \n",
       "\n",
       "[5 rows x 32 columns]"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "metadata[tags_vocal].head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "mp3_path\n",
       "d/ambient_teknology-the_all_seeing_eye_project-01-cyclops-262-291.mp3           False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-02-all_seeing_eye-175-204.mp3    False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-03-black-175-204.mp3             False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-04-confusion_says-88-117.mp3     False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-05-the_beholder-291-320.mp3      False\n",
       "dtype: bool"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# set vocal to True of any of the tags_vocal are 1\n",
    "gt_vocal = metadata[tags_vocal].any(axis=1)\n",
    "gt_vocal.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "mp3_path\n",
       "d/ambient_teknology-the_all_seeing_eye_project-01-cyclops-262-291.mp3           False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-02-all_seeing_eye-175-204.mp3    False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-03-black-175-204.mp3             False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-04-confusion_says-88-117.mp3     False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-05-the_beholder-291-320.mp3       True\n",
       "dtype: bool"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# set instrumental to True of any of the tags_instrumental are 1\n",
    "gt_instrumental = metadata[tags_instrumental].any(axis=1)\n",
    "gt_instrumental.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<b>We can only use the tag if EITHER instrumental OR vocal is True.</b><br>\n",
    "If both of them are True or both of them are False, we cannot trust the groundtruth data. Ergo we have to remove these and retain only the others."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "mp3_path\n",
       "d/ambient_teknology-the_all_seeing_eye_project-01-cyclops-262-291.mp3           False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-02-all_seeing_eye-175-204.mp3    False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-03-black-175-204.mp3             False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-04-confusion_says-88-117.mp3     False\n",
       "d/ambient_teknology-the_all_seeing_eye_project-05-the_beholder-291-320.mp3       True\n",
       "dtype: bool"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "retain = np.logical_xor(gt_vocal,gt_instrumental)\n",
    "retain.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "From originally 3023 input examples, we can only retain 959 trusted ones in our groundtruth\n"
     ]
    }
   ],
   "source": [
    "n_orig = len(gt_vocal)\n",
    "\n",
    "n_retain = sum(retain)\n",
    "\n",
    "print(\"From originally\", n_orig, \"input examples, we can only retain\",n_retain, \"trusted ones in our groundtruth\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In the end we cut from gt_vocal only the exampls to retain. If they are True they are vocal, if they are False, they are instrumental:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "mp3_path\n",
       "d/ambient_teknology-the_all_seeing_eye_project-05-the_beholder-291-320.mp3                                                                                                                   False\n",
       "0/american_bach_soloists-j_s__bach__cantatas_volume_v-03-gleichwie_der_regen_und_schnee_vom_himmel_fallt_bwv_18_iii_recitative_and_litany__mein_gott_hier_wird_mein_herze_sein-88-117.mp3     True\n",
       "0/american_bach_soloists-j_s__bach__cantatas_volume_v-07-weinen_klagen_sorgen_zagen_bwv_12_ii_chorus__weinen_klagen_sorgen_zagen-262-291.mp3                                                  True\n",
       "0/american_bach_soloists-j_s__bach__cantatas_volume_v-08-weinen_klagen_sorgen_zagen_bwv_12_iii_recitative__wir_mussen_durch_viel_truebsal-0-29.mp3                                            True\n",
       "0/american_bach_soloists-j_s__bach__cantatas_volume_v-09-weinen_klagen_sorgen_zagen_bwv_12_iv_aria__kreuz_und_krone_sind_verbunden-59-88.mp3                                                  True\n",
       "0/american_bach_soloists-j_s__bach__cantatas_volume_v-11-weinen_klagen_sorgen_zagen_bwv_12_vi_aria__sei_getreu-88-117.mp3                                                                     True\n",
       "0/american_bach_soloists-j_s__bach__cantatas_volume_v-13-nun_komm_der_heiden_heiland_bwv_61_i_ouverture__nun_komm_der_heiden_heiland-59-88.mp3                                                True\n",
       "0/american_bach_soloists-j_s__bach__cantatas_volume_v-14-nun_komm_der_heiden_heiland_bwv_61_ii_recitative__der_heiland_ist_gekommen-30-59.mp3                                                 True\n",
       "0/american_bach_soloists-j_s__bach__cantatas_volume_v-15-nun_komm_der_heiden_heiland_bwv_61_iii_aria__komm_jesu_komm_zu_deiner_kirche-117-146.mp3                                             True\n",
       "dtype: bool"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gt_final = gt_vocal[retain]\n",
    "gt_final.head(9)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "671 vocal tracks\n"
     ]
    }
   ],
   "source": [
    "print(str(sum(gt_final)) + \" vocal tracks\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "288 instrumental tracks\n"
     ]
    }
   ],
   "source": [
    "print(str(sum(np.logical_not(gt_final))) + \" instrumental tracks\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<b>Create two lists: one with filenames and one with associated classes</b>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# index in list of strings\n",
    "filelist = gt_final.index.tolist()\n",
    "# convert boolean to int and store in other list\n",
    "classes = (gt_final * 1).tolist()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['d/ambient_teknology-the_all_seeing_eye_project-05-the_beholder-291-320.mp3',\n",
       " '0/american_bach_soloists-j_s__bach__cantatas_volume_v-03-gleichwie_der_regen_und_schnee_vom_himmel_fallt_bwv_18_iii_recitative_and_litany__mein_gott_hier_wird_mein_herze_sein-88-117.mp3',\n",
       " '0/american_bach_soloists-j_s__bach__cantatas_volume_v-07-weinen_klagen_sorgen_zagen_bwv_12_ii_chorus__weinen_klagen_sorgen_zagen-262-291.mp3',\n",
       " '0/american_bach_soloists-j_s__bach__cantatas_volume_v-08-weinen_klagen_sorgen_zagen_bwv_12_iii_recitative__wir_mussen_durch_viel_truebsal-0-29.mp3',\n",
       " '0/american_bach_soloists-j_s__bach__cantatas_volume_v-09-weinen_klagen_sorgen_zagen_bwv_12_iv_aria__kreuz_und_krone_sind_verbunden-59-88.mp3']"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "filelist[0:5]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[0, 1, 1, 1, 1]"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "classes[0:5]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 68,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([0, 1, 1, 1, 1])"
      ]
     },
     "execution_count": 68,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# convert to Numpy array as needed by Keras\n",
    "classes = np.array(classes)\n",
    "classes[0:5]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load the Audio Files"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "...................................................................................................."
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python3.5/dist-packages/ipykernel_launcher.py:43: RuntimeWarning: divide by zero encountered in log10\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "...................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................\n",
      "Read 959 audio files\n"
     ]
    }
   ],
   "source": [
    "list_spectrograms = [] # spectrograms are put into a list first\n",
    "\n",
    "# desired output parameters\n",
    "\n",
    "# small spectrograms\n",
    "#n_mel_bands = 40   # y axis\n",
    "#frames = 80        # x axis\n",
    "\n",
    "# large  spectrograms\n",
    "n_mel_bands = 96   # y axis\n",
    "frames = 683        # x axis\n",
    "\n",
    "# extra large  spectrograms\n",
    "#n_mel_bands = 96   # y axis\n",
    "#frames = 1366        # x axis\n",
    "\n",
    "# some FFT parameters\n",
    "fft_window_size=512\n",
    "fft_overlap = 0.5\n",
    "hop_size = int(fft_window_size*(1-fft_overlap))\n",
    "segment_size = fft_window_size + (frames-1) * hop_size # segment size for desired # frames\n",
    "\n",
    "for filename in filelist:\n",
    "    print(\".\", end='')\n",
    "    filepath = os.path.join(AUDIO_PATH, filename)\n",
    "    samplerate, samplewidth, wavedata = audiofile_read(filepath,verbose=False)\n",
    "    sample_length = wavedata.shape[0]\n",
    "\n",
    "    # make Mono (in case of multiple channels / stereo)\n",
    "    if wavedata.ndim > 1:\n",
    "        wavedata = np.mean(wavedata, 1)\n",
    "        \n",
    "    # take only a segment; choose start position:\n",
    "    #pos = 0 # beginning\n",
    "    pos = int(wavedata.shape[0]/2 - segment_size/2) # center minus half segment size\n",
    "    wav_segment = wavedata[pos:pos+segment_size]\n",
    "\n",
    "    # 1) FFT spectrogram \n",
    "    spectrogram = rp.calc_spectrogram(wav_segment,fft_window_size,fft_overlap)\n",
    "\n",
    "    # 2) Transform to perceptual Mel scale (uses librosa.filters.mel)\n",
    "    spectrogram = rp.transform2mel(spectrogram,samplerate,fft_window_size,n_mel_bands)\n",
    "        \n",
    "    # 3) Log 10 transform\n",
    "    spectrogram = np.log10(spectrogram)\n",
    "    \n",
    "    list_spectrograms.append(spectrogram)\n",
    "        \n",
    "print(\"\\nRead\", len(filelist), \"audio files\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(22050, 2)"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "samplerate, samplewidth"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([-0.00286865, -0.02416992,  0.00906372, ...,  0.04193115,\n",
       "        0.05844116,  0.06524658])"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "wav_segment"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "959"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(list_spectrograms)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(96, 683)"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "spectrogram.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "An audio segment is 7.94 seconds long\n"
     ]
    }
   ],
   "source": [
    "print(\"An audio segment is\", round(float(segment_size) / samplerate, 2), \"seconds long\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Show Waveform and Spectrogram"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'e/yongen-moonrise-08-nylon_heart-30-59.mp3'"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# you can skip this if you do not have matplotlib installed\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline \n",
    "\n",
    "# show 1 sec wave segment\n",
    "plt.plot(wav_segment)\n",
    "filename"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# show spectrogram\n",
    "fig = plt.imshow(spectrogram, origin='lower', aspect='auto')\n",
    "fig.set_cmap('jet')\n",
    "fig.axes.get_xaxis().set_visible(False)\n",
    "fig.axes.get_yaxis().set_visible(False)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Make 1 big array of list of spectrograms"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'float32'"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# convert the input data to the right data type used by Keras Deep Learning (GPU)\n",
    "dtype = keras.backend.floatx()\n",
    "dtype"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(959, 96, 683)"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# a list of many spectrograms is made into 1 big array with 3 dimensions\n",
    "data = np.array(list_spectrograms, dtype=dtype)\n",
    "data.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Replace Inf values\n",
    "\n",
    "As in our preprocessing some files generated an Inf value in the log10 computation, we replace those by 0:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "data[np.isinf(data)] = 0"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Standardization\n",
    "\n",
    "<b>Always standardize</b> the data before feeding it into the Neural Network!\n",
    "\n",
    "We use <b>Zero-mean Unit-variance standardization</b> (also known as Z-score normalization).\n",
    "Here, we use <b>attribute-wise standardization</b>, i.e. each pixel is standardized individually, as opposed to computing a single mean and single standard deviation of all values.\n",
    "\n",
    "('Flat' standardization would also be possible, but we have seen benefits of attribut-wise standardization in our experiments).\n",
    "\n",
    "We use the StandardScaler from the scikit-learn package for our purpose.\n",
    "As it works typically on vector data, we have to vectorize (i.e. reshape) our matrices first."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(959, 65568)"
      ]
     },
     "execution_count": 30,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# vectorize\n",
    "N, ydim, xdim = data.shape\n",
    "data = data.reshape(N, xdim*ydim)\n",
    "data.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [],
   "source": [
    "# standardize\n",
    "scaler = preprocessing.StandardScaler()\n",
    "data = scaler.fit_transform(data)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([ -4.355928 ,  -4.3532376,  -4.378992 , ..., -11.941039 ,\n",
       "        -11.947343 , -11.95338  ], dtype=float32),\n",
       " array([1.849471  , 1.8415686 , 1.8250653 , ..., 0.4442456 , 0.4620821 ,\n",
       "        0.44511256], dtype=float32))"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# show mean and standard deviation: two vectors with same length as data.shape[1]\n",
    "scaler.mean_, scaler.scale_"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Convolutional Neural Networks\n",
    "\n",
    "A Convolutional Neural Network (ConvNet or CNN) is a type of (deep) Neural Network that is well-suited for 2D axes data, such as images or spectrograms, as it is optimized for learning from spatial proximity. Its core elements are 2D filter kernels which essentially learn the weights of the Neural Network, and downscaling functions such as Max Pooling.\n",
    "\n",
    "A CNN can have one or more Convolution layers, each of them having an arbitrary number of N filters (which define the depth of the CNN layer), following typically by a pooling step, which aggregates neighboring pixels together and thus reduces the image resolution by retaining only the maximum values of neighboring pixels."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Preparing the Data\n",
    "\n",
    "### Adding the channel\n",
    "\n",
    "As CNNs were initially made for image data, we need to add a dimension for the color channel to the data. RGB images typically have a 3rd dimension with the color. \n",
    "\n",
    "<b>Spectrograms, however, are considered like greyscale images, as in the previous tutorial.\n",
    "Likewise we need to add an extra dimension for compatibility with the CNN implementation.</b>\n",
    "\n",
    "For greyscale images, we add the number 1 as the depth of the additional dimension of the input shape (for RGB color images, the number of channels is 3).\n",
    "\n",
    "<i>Note on Tensorflow vs. Theano:</i>\n",
    "\n",
    "In Theano, traditionally the color channel was the <b>first</b> dimension in the image shape. \n",
    "In Tensorflow, the color channel is the <b>last</b> dimension in the image shape. \n",
    "\n",
    "This can be configured in ~/.keras/keras.json: \"image_dim_ordering\": \"th\" or \"tf\" (for Theano or Tensorflow) *or* with \"image_data_format\" set to \"channels_first\" or \"channels_last\".\n",
    "\n",
    "Tensorflow is now the default image ordering for Kears (\"tf\" and/or \"channels_last\").\n",
    "To be on the safe side, we added the if statement below."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'channels_last'"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "keras.backend.image_data_format()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [],
   "source": [
    "n_channels = 1 # for grey-scale, 3 for RGB, but usually already present in the data\n",
    "\n",
    "if keras.backend.image_data_format() == 'channels_last':  # TENSORFLOW\n",
    "    # Tensorflow ordering (~/.keras/keras.json: \"image_dim_ordering\": \"tf\")\n",
    "    data = data.reshape(data.shape[0], ydim, xdim, n_channels)\n",
    "else: # THEANO\n",
    "    # Theano ordering (~/.keras/keras.json: \"image_dim_ordering\": \"th\")\n",
    "    data = data.reshape(data.shape[0], n_channels, ydim, xdim)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(959, 96, 683, 1)"
      ]
     },
     "execution_count": 35,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(96, 683, 1)"
      ]
     },
     "execution_count": 36,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# we store the new shape of the images in the 'input_shape' variable.\n",
    "# take all dimensions except the 0th one (which is the number of images)\n",
    "input_shape = data.shape[1:]  \n",
    "input_shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Train & Test Set Split\n",
    "\n",
    "We split the original full data set into two parts: Train Set (75%) and Test Set (25%).\n",
    "\n",
    "Note: \n",
    "For demo purposes we use only 1 split here. A better way to do it is to use **Cross-Validation**, doing the split multiple times, iterating training and testing over the splits and averaging the results."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [],
   "source": [
    "testset_size = 0.25 # % portion of whole data set to keep for testing, i.e. 75% is used for training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "TRAIN INDEX: [135 957 350 397 651 154 704 255 811 881 937 921 229 635 839 643   2 619\n",
      " 299 225 388 400 717 444 736 673 675 548 731 850 401 112 608  71 791  76\n",
      " 281 721 545 538 155 753 206  41 527 856 491 745 366 756 720 227 218 836\n",
      " 103 711  72 878  42 144 447 349 589 710 362 203 376 864 782 226 302 220\n",
      " 379 773 256 555 476 231  49  51 601 458 744 623 906 685 136 730 115 781\n",
      " 747 628 869 441 148 728 264 934  37 479 599 237 134 493 557 265 687  88\n",
      " 363 884 179 201 195 578 166 342 797 339 842 284 924 416 145 382  22  80\n",
      " 851 927  60 952 912 292 378 634 287 263 181 316 954  53 141 928 278 348\n",
      " 563 890 735 553  93 385 564 294 824 371 841 585 198 757 700 876   3 558\n",
      " 953 748 678 903 544 920 784  16 832 466 448 883 190 187 200 483 650 654\n",
      " 128 442  75 752 914 894 776 386 100  30 178   1 249 420 129 101 497 429\n",
      "  61 888 450 396 480 172 923 885 694 693 916 373 909 273 624 621 331  55\n",
      "  39 570 104 234 785 672  86 270 475  54  95 569 146  59 603 521 543 568\n",
      " 403 726  47 702 577  94  12 216 267 760 471 828 482 300 829 768 242 413\n",
      " 617 729 767 436  18 795 321 542 758 790 549 664  32 922 514 461 175 830\n",
      " 293 592 827 106 304 676 949 706 246  43 825 199 233 235 739 431 674 337\n",
      "  99 254 866 610 682 525 121 662 633 587 367 434 245 370 595 122 556 383\n",
      " 171 251 193 309 541 656 897   9 318 737 365 932 435 554 207  70 798 833\n",
      " 325   6 408 566 637 755 740 770 783 162 852 327 831 140 474 955 212 590\n",
      " 290 534 933  11   7 271 574  44 573 150 138 399 616 695 387 307 262 733\n",
      " 481  84 772 666 594 355  77 426 130 261  34 714 918 948 392 208 872 605\n",
      " 503  50 636 854 809 137 528  24 422 346 646 865 668 340  36 683  27 313\n",
      "  57 751 192 301 295  38 170 407 907 501 238 323 456 205 391 642 329 679\n",
      " 180 855 796  40 722 451  13 336 707 276 185 488 421 320 202  58 495 513\n",
      " 536 526 507 584 667 411 763 696 631 186 788  52 404 925 579 862 703 905\n",
      "  97 861 716 163 804 412 390 613 464 698  64 532 377 818 277 575  83 500\n",
      " 241 808 283 232 125 472 893 762 402 789 561 596 230 496 452 248 410  79\n",
      "  46 459 517 268 253 512 457 539 467 361 494 196  89 640  29  87 843 473\n",
      " 372  65 699  15  19 847 586 820 465 142 222 826 778 571 823 659 462 215\n",
      " 580 303  85 357 780  74 351 223 519 478 417 343 567 936 857 484 338  78\n",
      " 581 591 626 124 882 560 219 260 153 738 638 806 425 424 279   0  23 648\n",
      " 105 794 116 240 946 689  56  48 224 669 449 213 460 951 750 742 805 661\n",
      " 904 209 799 766 546 469 612 317 428 498 838 347 734 108 793 418 620  10\n",
      " 688 814 364 725 236 341 285 427 845 272 547 443 509 123 902 853 381 746\n",
      " 565  81 606 310 898 258 615 598 705 210 718 405 328 131 432 113  20 868\n",
      " 732 530 761 871 614 873 531 803 446 840 652 552 133  21 899 194 649 167\n",
      " 508 315 311 754 681 870 118 324 380 690  68 147  67 306  91 244 152 522\n",
      " 801 250 360 505 117  96 950 344  92 913 812 247  35 298 184 139 879 723\n",
      " 641 282 727 609 697  33 671 161 114 352 764 454 437 625 414 515 708 393\n",
      " 169 111 741 802 297 243 655 419 291 941 821 394 712 956 627 749 919]\n",
      "TEST INDEX: [800 158 816 468 197 680 529 463 516 958 354 947 769 438 665 333 330 917\n",
      " 891 415 588   8 151  45 597 204 239 813 374 938 880 499 433 489 844 691\n",
      " 834 440 647 867 107 322 817 692 901 319 931 191 848 375 485 368 334 520\n",
      " 774 518 940 353 819 858 406 143 157 622 860   5 792 645 359 266 510 132\n",
      " 572  17 211 653 535  98 176 455 849 775 945 164  63 670 188 686 228  62\n",
      " 896 398 286 660 877 630 308 779 120 477 332 926 743 280 326 787   4 430\n",
      " 506 859 944 930 658 214 409 356 314 895  25 629 550 551 423 771 713 701\n",
      " 274 892 611 943 663 709 395 504 453 109 910 942 177 684 576 156 369 777\n",
      "  90 846 445 822 810 149 639  14 523  73 604 759 540 389 252 492 889  28\n",
      " 335 312 677 863 345 533 908 384 600 874  66 168 110 607 939 524 875 935\n",
      " 837 275 439 470 502 583 288 183 217 486 182 119 305 644 911 487 886 189\n",
      " 126 358 724  69 296 715 618  82 165 582 657 807  31 257 929 559 786 562\n",
      " 160 815 490 900 289 269 173  26 719 915 632 593 835 537 259 511 127 765\n",
      " 887 221 602 174 159 102]\n"
     ]
    }
   ],
   "source": [
    "# Stratified Split retains the class balance in both sets\n",
    "\n",
    "splitter = StratifiedShuffleSplit(n_splits=1, test_size=testset_size, random_state=0)\n",
    "splits = splitter.split(data, classes)\n",
    "\n",
    "for train_index, test_index in splits:\n",
    "    print(\"TRAIN INDEX:\", train_index)\n",
    "    print(\"TEST INDEX:\", test_index)\n",
    "    train_set = data[train_index]\n",
    "    test_set = data[test_index]\n",
    "    train_classes = classes[train_index]\n",
    "    test_classes = classes[test_index]\n",
    "# Note: this for loop is only executed once if n_splits==1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(719, 96, 683, 1)\n",
      "(240, 96, 683, 1)\n"
     ]
    }
   ],
   "source": [
    "print(train_set.shape)\n",
    "print(test_set.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Class Counts: Class 0: 216 Class 1: 503\n"
     ]
    }
   ],
   "source": [
    "print(\"Class Counts: Class 0:\", sum(train_classes==0), \"Class 1:\", sum(train_classes))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Creating CNN Models in Keras"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## SKIP - Creating a Single Layer and a Two Layer CNN"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Try: (comment/uncomment code in the following code block)\n",
    "* 1 Layer\n",
    "* 2 Layer\n",
    "* more conv_filters\n",
    "* Dropout"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#np.random.seed(0) # make results repeatable\n",
    "\n",
    "model = Sequential()\n",
    "\n",
    "#conv_filters = 16   # number of convolution filters (= CNN depth)\n",
    "conv_filters = 32   # number of convolution filters (= CNN depth)\n",
    "\n",
    "# Layer 1\n",
    "model.add(Convolution2D(conv_filters, (3, 3), input_shape=input_shape))\n",
    "model.add(MaxPooling2D(pool_size=(2, 2))) \n",
    "model.add(Dropout(0.25)) \n",
    "\n",
    "# Layer 2\n",
    "model.add(Convolution2D(conv_filters, (3, 3)))\n",
    "model.add(MaxPooling2D(pool_size=(2, 2))) \n",
    "\n",
    "# After Convolution, we have a 16*x*y matrix output\n",
    "# In order to feed this to a Full(Dense) layer, we need to flatten all data\n",
    "# Note: Keras does automatic shape inference, i.e. it knows how many (flat) input units the next layer will need,\n",
    "# so no parameter is needed for the Flatten() layer.\n",
    "model.add(Flatten()) \n",
    "\n",
    "# Full layer\n",
    "model.add(Dense(256, activation='sigmoid')) \n",
    "\n",
    "# Output layer\n",
    "# For binary/2-class problems use ONE sigmoid unit, \n",
    "# for multi-class/multi-label problems use n output units and activation='softmax!'\n",
    "model.add(Dense(1,activation='sigmoid'))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "If you get OverflowError: Range exceeds valid bounds in the above box, check the correct Theano vs. Tensorflow ordering in the box before and your keras.json configuration file."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training the CNN"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 76,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Define a loss function \n",
    "loss = 'binary_crossentropy'  # 'categorical_crossentropy' for multi-class problems\n",
    "\n",
    "# Optimizer = Stochastic Gradient Descent\n",
    "optimizer = 'sgd' \n",
    "\n",
    "# Compiling the model\n",
    "model.compile(loss=loss, optimizer=optimizer, metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# TRAINING the model\n",
    "epochs = 15\n",
    "history = model.fit(train_set, train_classes, batch_size=32, epochs=epochs)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Accuracy goes up pretty quickly for 1 layer on Train set! Also on Test set?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.71875"
      ]
     },
     "execution_count": 40,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 2 layer\n",
    "accuracy_score(test_classes, test_pred)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.71875"
      ]
     },
     "execution_count": 41,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 2 layer + 32 convolution filters\n",
    "accuracy_score(test_classes, test_pred)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.71875"
      ]
     },
     "execution_count": 42,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 2 layer + 32 convolution filters + Dropout\n",
    "accuracy_score(test_classes, test_pred)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Compact CNN\n",
    "\n",
    "This is a 5 layer Convolutional Neural Network inspired and adapted from Keunwoo Choi (https://github.com/keunwoochoi/music-auto_tagging-keras)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [],
   "source": [
    "# additional imports for more fancy stuff\n",
    "\n",
    "from keras.layers.advanced_activations import ELU"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(959, 96, 683, 1)"
      ]
     },
     "execution_count": 48,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "#from keras import regularizers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [],
   "source": [
    "def CompactCNN(input_shape, nb_conv, nb_filters, n_mels, normalize, nb_hidden, dense_units, \n",
    "               output_shape, activation, dropout, multiple_segments=False, graph_model=False, input_tensor=None):\n",
    "    \n",
    "    melgram_input = Input(shape=input_shape)\n",
    "\n",
    "    if n_mels >= 256:\n",
    "        poolings = [(2, 4), (4, 4), (4, 5), (2, 4), (4, 4)]\n",
    "    elif n_mels >= 128:\n",
    "        poolings = [(2, 4), (4, 4), (2, 5), (2, 4), (4, 4)]\n",
    "    elif n_mels >= 96:\n",
    "        poolings = [(2, 4), (3, 4), (2, 5), (2, 4), (4, 4)]\n",
    "    elif n_mels >= 72:\n",
    "        poolings = [(2, 4), (3, 4), (2, 5), (2, 4), (3, 4)]\n",
    "    elif n_mels >= 64:\n",
    "        poolings = [(2, 4), (2, 4), (2, 5), (2, 4), (4, 4)]\n",
    "\n",
    "    # Determine input axis\n",
    "    if keras.backend.image_dim_ordering() == 'th':\n",
    "        channel_axis = 1\n",
    "        freq_axis = 2\n",
    "        time_axis = 3\n",
    "    else:\n",
    "        channel_axis = 3\n",
    "        freq_axis = 1\n",
    "        time_axis = 2\n",
    "            \n",
    "    # Input block\n",
    "    #x = BatchNormalization(axis=time_axis, name='bn_0_freq')(melgram_input)\n",
    "        \n",
    "    if normalize == 'batch':\n",
    "        x = BatchNormalization(axis=freq_axis, name='bn_0_freq')(melgram_input)\n",
    "    elif normalize in ('data_sample', 'time', 'freq', 'channel'):\n",
    "        x = Normalization2D(normalize, name='nomalization')(melgram_input)\n",
    "    elif normalize in ('no', 'False'):\n",
    "        x = melgram_input\n",
    "\n",
    "    # Conv block 1\n",
    "    x = Convolution2D(nb_filters[0], (3, 3), padding='same')(x)\n",
    "    x = BatchNormalization(axis=channel_axis, name='bn1')(x)\n",
    "    x = ELU()(x)\n",
    "    x = MaxPooling2D(pool_size=poolings[0], name='pool1')(x)\n",
    "        \n",
    "    # Conv block 2\n",
    "    x = Convolution2D(nb_filters[1], (3, 3), padding='same')(x)\n",
    "    x = BatchNormalization(axis=channel_axis, name='bn2')(x)\n",
    "    x = ELU()(x)\n",
    "    x = MaxPooling2D(pool_size=poolings[1], name='pool2')(x)\n",
    "        \n",
    "    # Conv block 3\n",
    "    x = Convolution2D(nb_filters[2], (3, 3), padding='same')(x)\n",
    "    x = BatchNormalization(axis=channel_axis, name='bn3')(x)\n",
    "    x = ELU()(x)\n",
    "    x = MaxPooling2D(pool_size=poolings[2], name='pool3')(x)\n",
    "    \n",
    "    # Conv block 4\n",
    "    if nb_conv > 3:        \n",
    "        x = Convolution2D(nb_filters[3], (3, 3), padding='same')(x)\n",
    "        x = BatchNormalization(axis=channel_axis, name='bn4')(x)\n",
    "        x = ELU()(x)   \n",
    "        x = MaxPooling2D(pool_size=poolings[3], name='pool4')(x)\n",
    "        \n",
    "    # Conv block 5\n",
    "    if nb_conv == 5:\n",
    "        x = Convolution2D(nb_filters[4], (3, 3), padding='same')(x)\n",
    "        x = BatchNormalization(axis=channel_axis, name='bn5')(x)\n",
    "        x = ELU()(x)\n",
    "        x = MaxPooling2D(pool_size=poolings[4], name='pool5')(x)\n",
    "\n",
    "    # Flatten the outout of the last Conv Layer\n",
    "    x = Flatten()(x)\n",
    "      \n",
    "    if nb_hidden == 1:\n",
    "        x = Dropout(dropout)(x)\n",
    "        x = Dense(dense_units, activation='relu')(x)\n",
    "    elif nb_hidden == 2:\n",
    "        x = Dropout(dropout)(x)\n",
    "        x = Dense(dense_units[0], activation='relu')(x)\n",
    "        x = Dropout(dropout)(x)\n",
    "        x = Dense(dense_units[1], activation='relu')(x) \n",
    "    else:\n",
    "        raise ValueError(\"More than 2 hidden units not supported at the moment.\")\n",
    "    \n",
    "    # Output Layer\n",
    "    x = Dense(output_shape, activation=activation, name = 'output')(x)\n",
    "    \n",
    "    # Create model\n",
    "    model = Model(melgram_input, x)\n",
    "    \n",
    "    return model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Set model parameters\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {},
   "outputs": [],
   "source": [
    "# number of Convolutional Layers\n",
    "nb_conv_layers = 4\n",
    "\n",
    "# number of Filters in each layer\n",
    "nb_filters = [64,64,64,128,128]\n",
    "\n",
    "# number of hidden layers at the end of the model\n",
    "nb_hidden = 1 # 2\n",
    "\n",
    "# how many neurons in each hidden layer\n",
    "dense_units = 128 #[128,56]\n",
    "\n",
    "# which activation function to use for OUTPUT\n",
    "activation = 'sigmoid'\n",
    "\n",
    "# which type of normalization\n",
    "normalization = 'batch'\n",
    "\n",
    "# droupout\n",
    "dropout = 0.2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = CompactCNN(input_shape, nb_conv = nb_conv_layers, nb_filters= nb_filters, n_mels = 96, \n",
    "                           normalize=normalization, \n",
    "                           nb_hidden = nb_hidden, dense_units = dense_units, \n",
    "                           output_shape = 1, activation = activation, \n",
    "                           dropout = dropout)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_2 (InputLayer)         (None, 96, 683, 1)        0         \n",
      "_________________________________________________________________\n",
      "bn_0_freq (BatchNormalizatio (None, 96, 683, 1)        384       \n",
      "_________________________________________________________________\n",
      "conv2d_6 (Conv2D)            (None, 96, 683, 64)       640       \n",
      "_________________________________________________________________\n",
      "bn1 (BatchNormalization)     (None, 96, 683, 64)       256       \n",
      "_________________________________________________________________\n",
      "elu_6 (ELU)                  (None, 96, 683, 64)       0         \n",
      "_________________________________________________________________\n",
      "pool1 (MaxPooling2D)         (None, 48, 170, 64)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_7 (Conv2D)            (None, 48, 170, 64)       36928     \n",
      "_________________________________________________________________\n",
      "bn2 (BatchNormalization)     (None, 48, 170, 64)       256       \n",
      "_________________________________________________________________\n",
      "elu_7 (ELU)                  (None, 48, 170, 64)       0         \n",
      "_________________________________________________________________\n",
      "pool2 (MaxPooling2D)         (None, 16, 42, 64)        0         \n",
      "_________________________________________________________________\n",
      "conv2d_8 (Conv2D)            (None, 16, 42, 64)        36928     \n",
      "_________________________________________________________________\n",
      "bn3 (BatchNormalization)     (None, 16, 42, 64)        256       \n",
      "_________________________________________________________________\n",
      "elu_8 (ELU)                  (None, 16, 42, 64)        0         \n",
      "_________________________________________________________________\n",
      "pool3 (MaxPooling2D)         (None, 8, 8, 64)          0         \n",
      "_________________________________________________________________\n",
      "conv2d_9 (Conv2D)            (None, 8, 8, 128)         73856     \n",
      "_________________________________________________________________\n",
      "bn4 (BatchNormalization)     (None, 8, 8, 128)         512       \n",
      "_________________________________________________________________\n",
      "elu_9 (ELU)                  (None, 8, 8, 128)         0         \n",
      "_________________________________________________________________\n",
      "pool4 (MaxPooling2D)         (None, 4, 2, 128)         0         \n",
      "_________________________________________________________________\n",
      "flatten_1 (Flatten)          (None, 1024)              0         \n",
      "_________________________________________________________________\n",
      "dropout_1 (Dropout)          (None, 1024)              0         \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 128)               131200    \n",
      "_________________________________________________________________\n",
      "output (Dense)               (None, 1)                 129       \n",
      "=================================================================\n",
      "Total params: 281,345\n",
      "Trainable params: 280,513\n",
      "Non-trainable params: 832\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training Setup"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Loss\n",
    "loss = 'binary_crossentropy' "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Optimizers\n",
    "sgd = optimizers.SGD(momentum=0.9, nesterov=True)\n",
    "rmsprop = optimizers.RMSprop(lr=0.001, rho=0.9, epsilon=1e-08, decay=0.01)#lr=0.001 decay = 0.03\n",
    "adagrad = optimizers.Adagrad(lr=0.01, epsilon=1e-08, decay=0.0)\n",
    "\n",
    "# We use mostly ADAM\n",
    "adam = optimizers.Adam(lr=0.003, beta_1=0.9, beta_2=0.999, epsilon=1e-07, decay=0.01)\n",
    "nadam = optimizers.Nadam(lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=1e-07, schedule_decay=0.004)\n",
    "\n",
    "# choose\n",
    "optimizer = adam"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Metrics\n",
    "\n",
    "def precision(y_true, y_pred):\n",
    "    true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))\n",
    "    predicted_positives = K.sum(K.round(K.clip(y_pred, 0, 1)))\n",
    "    precision = true_positives / (predicted_positives + K.epsilon())\n",
    "    return precision\n",
    "\n",
    "def recall(y_true, y_pred):\n",
    "    true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)))\n",
    "    possible_positives = K.sum(K.round(K.clip(y_true, 0, 1)))\n",
    "    recall = true_positives / (possible_positives + K.epsilon())\n",
    "    return recall\n",
    "\n",
    "metrics = ['accuracy', precision, recall]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 55,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Other\n",
    "batch_size = 32 \n",
    "\n",
    "epochs = 40\n",
    "\n",
    "validation_split=0.1 \n",
    "\n",
    "#n_folds = 5\n",
    "random_seed = 0\n",
    "\n",
    "callbacks = None"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Tensorboard (optional)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.callbacks import TensorBoard"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {},
   "outputs": [],
   "source": [
    "home_dir = os.getenv(\"HOME\")\n",
    "\n",
    "tb_logdir = os.path.join(home_dir, \"tensorboard\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {},
   "outputs": [],
   "source": [
    "# OPTIONAL\n",
    "# new tensorboard callback at each training\n",
    "# tensorboard_run_id = \"Vocal_magna_2seg_adam_compact_128fbis_128h\"\n",
    "# tb_logdir = \"%s/%s_fold%d %s\" %(tb_logdir, tensorboard_run_id, fold, strftime(\"%Y-%m-%d %H:%M:%S\", localtime()))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 59,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Execute the following in a terminal:\n",
      "\n",
      "tensorboard --logdir=/home/tlidy/tensorboard\n"
     ]
    }
   ],
   "source": [
    "print(\"Execute the following in a terminal:\\n\")\n",
    "print(\"tensorboard --logdir=\" + tb_logdir)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 60,
   "metadata": {},
   "outputs": [],
   "source": [
    "# initialize TensorBoard in Python\n",
    "tensorboard = TensorBoard(log_dir = tb_logdir)\n",
    "\n",
    "# + add to callbacks\n",
    "callbacks = [tensorboard]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Then open Tensorboard in browser:\n",
    "\n",
    "http://localhost:6006"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 61,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "binary_crossentropy\n",
      "<keras.optimizers.Adam object at 0x7f34f081d208>\n",
      "['accuracy', <function precision at 0x7f34f0b3f2f0>, <function recall at 0x7f34f0b3fae8>]\n",
      "Batch size: 32 Epochs: 40\n"
     ]
    }
   ],
   "source": [
    "# Summary of Training options\n",
    "\n",
    "print(loss)\n",
    "print(optimizer)\n",
    "print(metrics)\n",
    "print(\"Batch size:\", batch_size, \"Epochs:\", epochs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "metadata": {},
   "outputs": [],
   "source": [
    "# COMPILE MODEL\n",
    "\n",
    "model.compile(loss=loss, metrics=metrics, optimizer=optimizer)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 67,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 647 samples, validate on 72 samples\n",
      "Epoch 1/40\n",
      "647/647 [==============================] - 5s 8ms/step - loss: 0.0555 - acc: 0.9768 - precision: 0.9874 - recall: 0.9830 - val_loss: 1.0835 - val_acc: 0.8611 - val_precision: 0.8443 - val_recall: 1.0000\n",
      "Epoch 2/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0894 - acc: 0.9614 - precision: 0.9691 - recall: 0.9801 - val_loss: 0.8851 - val_acc: 0.8333 - val_precision: 0.8598 - val_recall: 0.9251\n",
      "Epoch 3/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0339 - acc: 0.9907 - precision: 0.9931 - recall: 0.9946 - val_loss: 1.1578 - val_acc: 0.8750 - val_precision: 0.8697 - val_recall: 0.9815\n",
      "Epoch 4/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0303 - acc: 0.9907 - precision: 0.9945 - recall: 0.9916 - val_loss: 0.8963 - val_acc: 0.8472 - val_precision: 0.8636 - val_recall: 0.9436\n",
      "Epoch 5/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0171 - acc: 0.9938 - precision: 0.9965 - recall: 0.9959 - val_loss: 0.6863 - val_acc: 0.8750 - val_precision: 0.9052 - val_recall: 0.9251\n",
      "Epoch 6/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0212 - acc: 0.9969 - precision: 0.9978 - recall: 0.9979 - val_loss: 0.7754 - val_acc: 0.8472 - val_precision: 0.9181 - val_recall: 0.8680\n",
      "Epoch 7/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0322 - acc: 0.9876 - precision: 0.9927 - recall: 0.9898 - val_loss: 0.6466 - val_acc: 0.8750 - val_precision: 0.9088 - val_recall: 0.9251\n",
      "Epoch 8/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0079 - acc: 0.9985 - precision: 1.0000 - recall: 0.9978 - val_loss: 0.6795 - val_acc: 0.8472 - val_precision: 0.8741 - val_recall: 0.9251\n",
      "Epoch 9/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0137 - acc: 0.9954 - precision: 0.9923 - recall: 1.0000 - val_loss: 0.6583 - val_acc: 0.8472 - val_precision: 0.8874 - val_recall: 0.9066\n",
      "Epoch 10/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0174 - acc: 0.9907 - precision: 0.9939 - recall: 0.9907 - val_loss: 0.7998 - val_acc: 0.8056 - val_precision: 0.9329 - val_recall: 0.7915\n",
      "Epoch 11/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0060 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.8224 - val_acc: 0.8056 - val_precision: 0.9333 - val_recall: 0.7923\n",
      "Epoch 12/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0194 - acc: 0.9954 - precision: 0.9946 - recall: 0.9978 - val_loss: 1.2215 - val_acc: 0.7500 - val_precision: 0.9500 - val_recall: 0.6973\n",
      "Epoch 13/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0435 - acc: 0.9815 - precision: 0.9957 - recall: 0.9791 - val_loss: 1.0149 - val_acc: 0.7917 - val_precision: 0.9122 - val_recall: 0.7923\n",
      "Epoch 14/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0202 - acc: 0.9938 - precision: 0.9927 - recall: 0.9978 - val_loss: 1.0249 - val_acc: 0.7500 - val_precision: 0.9255 - val_recall: 0.7158\n",
      "Epoch 15/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0132 - acc: 0.9985 - precision: 1.0000 - recall: 0.9980 - val_loss: 0.6672 - val_acc: 0.8611 - val_precision: 0.9198 - val_recall: 0.8873\n",
      "Epoch 16/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0079 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.6116 - val_acc: 0.8611 - val_precision: 0.9205 - val_recall: 0.8881\n",
      "Epoch 17/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0083 - acc: 0.9985 - precision: 0.9969 - recall: 1.0000 - val_loss: 0.7015 - val_acc: 0.8333 - val_precision: 0.9333 - val_recall: 0.8293\n",
      "Epoch 18/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0047 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.6665 - val_acc: 0.8333 - val_precision: 0.9354 - val_recall: 0.8301\n",
      "Epoch 19/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0030 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.6601 - val_acc: 0.8611 - val_precision: 0.9198 - val_recall: 0.8873\n",
      "Epoch 20/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0040 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.6819 - val_acc: 0.8611 - val_precision: 0.9374 - val_recall: 0.8680\n",
      "Epoch 21/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0115 - acc: 0.9969 - precision: 1.0000 - recall: 0.9960 - val_loss: 0.6998 - val_acc: 0.8750 - val_precision: 0.9059 - val_recall: 0.9251\n",
      "Epoch 22/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0055 - acc: 0.9969 - precision: 0.9951 - recall: 1.0000 - val_loss: 0.6873 - val_acc: 0.8889 - val_precision: 0.9222 - val_recall: 0.9251\n",
      "Epoch 23/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0041 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.7059 - val_acc: 0.8889 - val_precision: 0.9222 - val_recall: 0.9251\n",
      "Epoch 24/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0034 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.7022 - val_acc: 0.8889 - val_precision: 0.9222 - val_recall: 0.9251\n",
      "Epoch 25/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0032 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.7069 - val_acc: 0.8750 - val_precision: 0.9214 - val_recall: 0.9066\n",
      "Epoch 26/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0021 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.7271 - val_acc: 0.8611 - val_precision: 0.9205 - val_recall: 0.8881\n",
      "Epoch 27/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0054 - acc: 0.9985 - precision: 1.0000 - recall: 0.9980 - val_loss: 0.7469 - val_acc: 0.8611 - val_precision: 0.9205 - val_recall: 0.8881\n",
      "Epoch 28/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0023 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.7885 - val_acc: 0.8472 - val_precision: 0.9181 - val_recall: 0.8680\n",
      "Epoch 29/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0035 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.7864 - val_acc: 0.8611 - val_precision: 0.9374 - val_recall: 0.8680\n",
      "Epoch 30/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0034 - acc: 0.9985 - precision: 1.0000 - recall: 0.9979 - val_loss: 0.6829 - val_acc: 0.8750 - val_precision: 0.9214 - val_recall: 0.9066\n",
      "Epoch 31/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0019 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.6818 - val_acc: 0.8889 - val_precision: 0.9222 - val_recall: 0.9251\n",
      "Epoch 32/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0027 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.7535 - val_acc: 0.8611 - val_precision: 0.9374 - val_recall: 0.8680\n",
      "Epoch 33/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0032 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.7137 - val_acc: 0.8750 - val_precision: 0.9214 - val_recall: 0.9066\n",
      "Epoch 34/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0054 - acc: 0.9969 - precision: 0.9953 - recall: 1.0000 - val_loss: 0.7606 - val_acc: 0.8750 - val_precision: 0.9383 - val_recall: 0.8873\n",
      "Epoch 35/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0085 - acc: 0.9954 - precision: 0.9982 - recall: 0.9956 - val_loss: 0.7127 - val_acc: 0.8889 - val_precision: 0.9222 - val_recall: 0.9251\n",
      "Epoch 36/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0081 - acc: 0.9985 - precision: 1.0000 - recall: 0.9981 - val_loss: 1.0106 - val_acc: 0.8056 - val_precision: 0.9333 - val_recall: 0.7923\n",
      "Epoch 37/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0078 - acc: 0.9969 - precision: 0.9971 - recall: 0.9983 - val_loss: 0.6810 - val_acc: 0.8889 - val_precision: 0.9391 - val_recall: 0.9066\n",
      "Epoch 38/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0027 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.6472 - val_acc: 0.8750 - val_precision: 0.9214 - val_recall: 0.9066\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 39/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0047 - acc: 0.9985 - precision: 0.9969 - recall: 1.0000 - val_loss: 0.6763 - val_acc: 0.8611 - val_precision: 0.9037 - val_recall: 0.9066\n",
      "Epoch 40/40\n",
      "647/647 [==============================] - 5s 7ms/step - loss: 0.0024 - acc: 1.0000 - precision: 1.0000 - recall: 1.0000 - val_loss: 0.7388 - val_acc: 0.8889 - val_precision: 0.9222 - val_recall: 0.9251\n"
     ]
    }
   ],
   "source": [
    "# START TRAINING\n",
    "\n",
    "# past_epochs is only for the case that we execute this code box multiple times (so that Tensorboard is displaying properly)\n",
    "past_epochs = 0\n",
    "\n",
    "history = model.fit(train_set, train_classes, \n",
    "                     validation_split=validation_split,\n",
    "                     #validation_data=(X_test,y_test), \n",
    "                     epochs=epochs, \n",
    "                     initial_epoch=past_epochs,\n",
    "                     batch_size=batch_size, \n",
    "                     callbacks=callbacks\n",
    "                     )\n",
    "\n",
    "past_epochs += epochs"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Verifying Accuracy on Test Set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "# always execute this, and then a box of accuracy_score below to print the result\n",
    "test_pred = model.predict_classes(test_set)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.71875"
      ]
     },
     "execution_count": 39,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 1 layer\n",
    "accuracy_score(test_classes, test_pred)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Additional Parameters & Techniques\n",
    "\n",
    "Try out more parameters and techniques: (comment/uncomment code blocks below)\n",
    "* Adding ReLU activation\n",
    "* Adding Batch normalization\n",
    "* Adding Dropout"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = Sequential()\n",
    "\n",
    "conv_filters = 16   # number of convolution filters (= CNN depth)\n",
    "filter_size = (3,3)\n",
    "pool_size = (2,2)\n",
    "\n",
    "# Layer 1\n",
    "model.add(Convolution2D(conv_filters, filter_size, padding='valid', input_shape=input_shape))\n",
    "#model.add(BatchNormalization())\n",
    "#model.add(Activation('relu')) \n",
    "model.add(MaxPooling2D(pool_size=pool_size)) \n",
    "#model.add(Dropout(0.3))\n",
    "\n",
    "# Layer 2\n",
    "model.add(Convolution2D(conv_filters, filter_size, padding='valid', input_shape=input_shape))\n",
    "#model.add(BatchNormalization())\n",
    "#model.add(Activation('relu')) \n",
    "model.add(MaxPooling2D(pool_size=pool_size)) \n",
    "#model.add(Dropout(0.1))\n",
    "\n",
    "# In order to feed this to a Full(Dense) layer, we need to flatten all data\n",
    "model.add(Flatten()) \n",
    "\n",
    "# Full layer\n",
    "model.add(Dense(256))  \n",
    "#model.add(Activation('relu'))\n",
    "#model.add(Dropout(0.1))\n",
    "\n",
    "# Output layer\n",
    "# For binary/2-class problems use ONE sigmoid unit, \n",
    "# for multi-class/multi-label problems use n output units and activation='softmax!'\n",
    "model.add(Dense(1,activation='sigmoid'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {},
   "outputs": [],
   "source": [
    "# MIREX 2015 model\n",
    "model = Sequential()\n",
    "\n",
    "conv_filters = 15   # number of convolution filters (= CNN depth)\n",
    "\n",
    "# Layer 1\n",
    "model.add(Convolution2D(conv_filters, (12, 8), padding='valid', input_shape=input_shape))\n",
    "#model.add(BatchNormalization())\n",
    "#model.add(Activation('relu')) \n",
    "model.add(Activation('sigmoid')) \n",
    "model.add(MaxPooling2D(pool_size=(2, 1))) \n",
    "#model.add(Dropout(0.3))\n",
    "\n",
    "\n",
    "# In order to feed this to a Full(Dense) layer, we need to flatten all data\n",
    "model.add(Flatten()) \n",
    "\n",
    "# Full layer\n",
    "model.add(Dense(200, activation='sigmoid'))  \n",
    "#model.add(Activation('relu'))\n",
    "model.add(Dropout(0.5))\n",
    "\n",
    "# Output layer\n",
    "# For binary/2-class problems use ONE sigmoid unit, \n",
    "# for multi-class/multi-label problems use n output units and activation='softmax!'\n",
    "model.add(Dense(1,activation='sigmoid'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Compiling and training the model\n",
    "model.compile(loss=loss, optimizer=optimizer, metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.7729 - acc: 0.4688\n",
      "Epoch 2/15\n",
      "96/96 [==============================] - 0s 1ms/step - loss: 0.7734 - acc: 0.4583\n",
      "Epoch 3/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.7648 - acc: 0.5417\n",
      "Epoch 4/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.7545 - acc: 0.4583\n",
      "Epoch 5/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.6950 - acc: 0.5417\n",
      "Epoch 6/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.6950 - acc: 0.5833\n",
      "Epoch 7/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.6554 - acc: 0.5833\n",
      "Epoch 8/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.6468 - acc: 0.6146\n",
      "Epoch 9/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.6807 - acc: 0.6250\n",
      "Epoch 10/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.6628 - acc: 0.6042\n",
      "Epoch 11/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.7091 - acc: 0.6146\n",
      "Epoch 12/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.6760 - acc: 0.6042\n",
      "Epoch 13/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.6157 - acc: 0.6667\n",
      "Epoch 14/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.6132 - acc: 0.6771\n",
      "Epoch 15/15\n",
      "96/96 [==============================] - 0s 2ms/step - loss: 0.5837 - acc: 0.7396\n"
     ]
    }
   ],
   "source": [
    "epochs = 15\n",
    "history = model.fit(train_set, train_classes, batch_size=32, epochs=epochs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.71875"
      ]
     },
     "execution_count": 47,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Verifying Accuracy on Test Set\n",
    "\n",
    "test_pred = model.predict_classes(test_set)\n",
    "accuracy_score(test_classes, test_pred)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true
   },
   "source": [
    "## Parallel CNNs\n",
    "\n",
    "It has been discovered, that CNNs for music work best, when they have one filter that is detecting frequencies in the vertical axis, and nother filter that is focused on the time axis, i.e. detecting rhythm. Consequently, this is realized in a parallel CNN, where 2 layers are not stacked after each other, but first run independently in parallel with their output being merged later.\n",
    "\n",
    "To create parallel CNNs we need a \"graph-based\" model. In Keras 1.x this is realized via the functional API of the Model() class.\n",
    "We use it to create two CNN layers that run in parallel to each other and are merged subsequently.\n",
    "In the functional API, you pass the name of the previous layer in (brackets) after defining the next layer."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [],
   "source": [
    "from keras.layers.merge import concatenate"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Input only specifies the input shape\n",
    "input = Input(input_shape)\n",
    "\n",
    "# CNN layers\n",
    "# specify desired number of filters\n",
    "n_filters = 16 \n",
    "# The functional API allows to specify the predecessor in (brackets) after the new Layer function call\n",
    "conv_layer1 = Convolution2D(16, (10, 2))(input)  # a vertical filter\n",
    "conv_layer2 = Convolution2D(25, (2, 10))(input)  # a horizontal filter\n",
    "\n",
    "# possibly add Activation('relu') here\n",
    "\n",
    "# Pooling layers\n",
    "maxpool1 = MaxPooling2D(pool_size=(1,2))(conv_layer1) # horizontal pooling\n",
    "maxpool2 = MaxPooling2D(pool_size=(2,1))(conv_layer2) # vertical pooling\n",
    "\n",
    "# we have to flatten the Pooling output in order to be concatenated\n",
    "flat1 = Flatten()(maxpool1)\n",
    "flat2 = Flatten()(maxpool2)\n",
    "\n",
    "# Merge the 2\n",
    "merged = concatenate([flat1, flat2])\n",
    "\n",
    "full = Dense(256, activation='relu')(merged)\n",
    "output_layer = Dense(1, activation='sigmoid')(full)\n",
    "\n",
    "# finally create the model\n",
    "model = Model(inputs=input, outputs=output_layer)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "__________________________________________________________________________________________________\n",
      "Layer (type)                    Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      "input_1 (InputLayer)            (None, 40, 80, 1)    0                                            \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_6 (Conv2D)               (None, 31, 79, 16)   336         input_1[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_7 (Conv2D)               (None, 39, 71, 25)   525         input_1[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "max_pooling2d_6 (MaxPooling2D)  (None, 31, 39, 16)   0           conv2d_6[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "max_pooling2d_7 (MaxPooling2D)  (None, 19, 71, 25)   0           conv2d_7[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "flatten_4 (Flatten)             (None, 19344)        0           max_pooling2d_6[0][0]            \n",
      "__________________________________________________________________________________________________\n",
      "flatten_5 (Flatten)             (None, 33725)        0           max_pooling2d_7[0][0]            \n",
      "__________________________________________________________________________________________________\n",
      "concatenate_1 (Concatenate)     (None, 53069)        0           flatten_4[0][0]                  \n",
      "                                                                 flatten_5[0][0]                  \n",
      "__________________________________________________________________________________________________\n",
      "dense_7 (Dense)                 (None, 256)          13585920    concatenate_1[0][0]              \n",
      "__________________________________________________________________________________________________\n",
      "dense_8 (Dense)                 (None, 1)            257         dense_7[0][0]                    \n",
      "==================================================================================================\n",
      "Total params: 13,587,038\n",
      "Trainable params: 13,587,038\n",
      "Non-trainable params: 0\n",
      "__________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Define a loss function \n",
    "loss = 'binary_crossentropy'  # 'categorical_crossentropy' for multi-class problems\n",
    "\n",
    "# Optimizer = Stochastic Gradient Descent\n",
    "optimizer = 'sgd' \n",
    "\n",
    "# Compiling the model\n",
    "model.compile(loss=loss, optimizer=optimizer, metrics=['accuracy'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 52,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/15\n",
      "96/96 [==============================] - 0s 5ms/step - loss: 0.5242 - acc: 0.7188\n",
      "Epoch 2/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.2932 - acc: 0.8958\n",
      "Epoch 3/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.2043 - acc: 0.9375\n",
      "Epoch 4/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.1615 - acc: 0.9688\n",
      "Epoch 5/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.1361 - acc: 0.9896\n",
      "Epoch 6/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.1095 - acc: 0.9896\n",
      "Epoch 7/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.0953 - acc: 0.9896\n",
      "Epoch 8/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.0815 - acc: 1.0000\n",
      "Epoch 9/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.0726 - acc: 1.0000\n",
      "Epoch 10/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.0641 - acc: 1.0000\n",
      "Epoch 11/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.0574 - acc: 1.0000\n",
      "Epoch 12/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.0516 - acc: 1.0000\n",
      "Epoch 13/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.0468 - acc: 1.0000\n",
      "Epoch 14/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.0429 - acc: 1.0000\n",
      "Epoch 15/15\n",
      "96/96 [==============================] - 0s 3ms/step - loss: 0.0403 - acc: 1.0000\n"
     ]
    }
   ],
   "source": [
    "# TRAINING the model\n",
    "epochs = 15\n",
    "history = model.fit(train_set, train_classes, batch_size=32, epochs=epochs)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Verifying Accuracy on Test Set\n",
    "\n",
    "Note: The functional API, i.e. Model() does not have a convenience method `.predict_classes()`. We therefore do 'raw' predictions with `predict()`, which returns values between 0 and 1, and then round to the nearest value (0 or 1)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([1.8116090e-03, 9.9629503e-01, 8.9173570e-02, 1.3832313e-02,\n",
       "       1.5312824e-02, 6.2414613e-02, 9.4741523e-01, 2.8874350e-01,\n",
       "       6.6524655e-02, 7.7041835e-01, 5.9007388e-03, 8.8411587e-01,\n",
       "       7.7622104e-03, 5.5290854e-01, 7.2940448e-03, 9.7821641e-01,\n",
       "       1.9546174e-03, 1.2710236e-01, 2.5445417e-02, 9.2131579e-01,\n",
       "       5.5363268e-01, 7.3306054e-02, 9.8920244e-01, 8.6430952e-02,\n",
       "       8.9812309e-01, 6.6113585e-05, 2.7863439e-03, 9.9372864e-01,\n",
       "       2.8969401e-02, 9.8893112e-01, 7.3728962e-03, 6.7004067e-01],\n",
       "      dtype=float32)"
      ]
     },
     "execution_count": 53,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test_pred = model.predict(test_set)\n",
    "test_pred[0:35,0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.78125"
      ]
     },
     "execution_count": 54,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test_pred = np.round(test_pred)\n",
    "accuracy_score(test_classes, test_pred)"
   ]
  }
 ],
 "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": "3.5.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}