MIR_Submission.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Tabla bol recognition\n",
    "by Błażej Kotowski (blazej.kotowski@upf.edu) and Ninad Puranik (ninad.puranik01@estudiant.upf.edu)\n",
    "\n",
    "### Task definition\n",
    "The tabla is a pair of drums known as ‘tabla’ and ‘bayan’ commonly used in accompaniment\n",
    "and solo performances in Indian Music. A variety of sounds can be produced using different\n",
    "kinds of strokes on the tabla-bayan. A mnemonic syllable or bol is associated with each of\n",
    "these strokes. We address the problem of recognizing the bol performed in an audio\n",
    "recording of a tabla solo performance with a harmonium accompaniment. We use a\n",
    "supervised learning approach to classify the ‘bols’ using the timbral features of audio and the\n",
    "manual annotations from the compmusic tabla-solo dataset.\n",
    "\n",
    "### Dataset\n",
    "Tabla Solo dataset: https://compmusic.upf.edu/tabla-solo-dataset\n",
    "\n",
    "\n",
    "# Implementation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import essentia.standard as estd\n",
    "import essentia as ess\n",
    "import scipy\n",
    "import pandas as pd\n",
    "import csv\n",
    "import re\n",
    "import os\n",
    "import IPython.display as ipd\n",
    "import tempfile\n",
    "import random\n",
    "from sklearn.neighbors import KNeighborsClassifier\n",
    "from sklearn import svm\n",
    "from sklearn.tree import DecisionTreeClassifier\n",
    "from sklearn.metrics import confusion_matrix, classification_report\n",
    "from sklearn.feature_selection import SelectKBest, chi2\n",
    "from sklearn.model_selection import train_test_split, KFold, GridSearchCV\n",
    "from sklearn.preprocessing import MinMaxScaler\n",
    "import seaborn as sns\n",
    "import matplotlib.pyplot as plt\n",
    "import sys\n",
    "sns.set()\n",
    "JUP_STDOUT = sys.stdout"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Dataset definition\n",
    "In this cell we are setting dataset related parameters. These include:\n",
    "- location of the dataset,\n",
    "- results location,\n",
    "- bols available in the dataset\n",
    "\n",
    "They should be modified according to the local machine configuration and used dataset."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Set base directory where the dataset is located\n",
    "dataset_root = \"./tabla_solo_1.0\"\n",
    "\n",
    "# Define the rest of paths\n",
    "filelist_path = os.path.join(dataset_root, 'filelist.txt')\n",
    "audio_files_path = os.path.join(dataset_root, \"wav\")\n",
    "annotation_files_path = os.path.join(dataset_root, \"onsMap\")\n",
    "\n",
    "# Filename for features to be saved and loaded from\n",
    "features_filename = \"extracted_features.csv\"\n",
    "\n",
    "# Path for classifier outputs \n",
    "results_path = \"./results\"\n",
    "\n",
    "# List of all possible bols\n",
    "BOL_NAMES = ['DA', 'KI', 'GE', 'TA', 'NA', 'DIN', 'KDA', 'TIT', 'DHA', \n",
    "        'DHE', 'DHET', 'DHI', 'DHIN', 'RE', 'TE', 'TII', 'TIN', 'TRA']\n",
    "\n",
    "# Collect list of datafiles\n",
    "filenames = [line.rstrip('\\n') for line in open(filelist_path)]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Feature extraction\n",
    "### Steps\n",
    "1. First we are using Essentia's Extractor class to extract all low-level features using frameSize of 1024 and hopSize of 512. Then we aggregate them using PoolAggregator and transform vectors of length N to N single-value fields.\n",
    "2. Next we perform MFCCs extraction using Essentia's built-in extractor with analysis parameters of well-estabilished HTK toolkit. \n",
    "3. We calculate the MFCC derivatives of first end second order (velocity and acceleration)\n",
    "4. We then merge all the features into single Pool.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "def flatten_pool(pool):\n",
    "    '''\n",
    "    For each enumerable feature in a pool, create separate field for each value\n",
    "    '''\n",
    "    for feature in pool.descriptorNames():\n",
    "        if hasattr(pool[feature], '__iter__'):\n",
    "            values_list = pool[feature]\n",
    "            pool.remove(feature)\n",
    "            for i, elem in enumerate(values_list):\n",
    "                pool.set(\"{}.{}\".format(feature,i), elem)\n",
    "    return pool\n",
    "\n",
    "\n",
    "def extract_mfccs(audio_frame):\n",
    "    '''\n",
    "    Extracts mean mfcss from given audio frame. Mimicks behavior of HTK.\n",
    "    \n",
    "    Arguments:\n",
    "        - audio_frame:[float] - audio frame\n",
    "    Returns:\n",
    "        - result:dictionary – a dictionary of calculated mean mfccs\n",
    "    '''\n",
    "    # HTK-like parameters for MFCC\n",
    "    mfcc_extractor = estd.MFCC(\n",
    "        type = 'magnitude',\n",
    "        warpingFormula = 'htkMel',\n",
    "        weighting = 'linear',\n",
    "        highFrequencyBound = 8000,\n",
    "        numberBands = 26,\n",
    "        numberCoefficients = 13,\n",
    "        normalize = 'unit_max',\n",
    "        dctType = 3,\n",
    "        logType = 'log',\n",
    "        liftering = 22\n",
    "    )\n",
    "    \n",
    "    # HTK-like Parameters for FrameGenerator\n",
    "    frame_generator = estd.FrameGenerator(\n",
    "        audio_frame,\n",
    "        frameSize = 1102,\n",
    "        hopSize = 441,\n",
    "        startFromZero = True,\n",
    "        validFrameThresholdRatio = 1\n",
    "    )\n",
    "    \n",
    "    # HTK-like parameters for windowing\n",
    "    window = estd.Windowing(\n",
    "        type = 'hamming',\n",
    "        size = 1102,\n",
    "        zeroPadding = 946,\n",
    "        normalized = False\n",
    "    )\n",
    "    \n",
    "    spectrum = estd.Spectrum()\n",
    "    \n",
    "    features = ess.Pool()\n",
    "    mfccs = []\n",
    "    for frame in frame_generator:\n",
    "        spec = spectrum(window(frame))\n",
    "        _, mfcc_coeffs = mfcc_extractor(spec)\n",
    "        mfccs.append(mfcc_coeffs)\n",
    "\n",
    "    mfccs = np.array(mfccs).T\n",
    "    \n",
    "    for band_id, coeffs in enumerate(mfccs):\n",
    "        for coeff in coeffs:\n",
    "            features.add(\"htk_mfcc.%d\" % band_id, coeff)    \n",
    "\n",
    "    # Pool aggregation\n",
    "    aggregated = estd.PoolAggregator(defaultStats=['mean','var','dmean','dvar','dmean2','dvar2'])(features)\n",
    "    return aggregated\n",
    "    return features\n",
    "    \n",
    "    \n",
    "def extract_features(audio_frame):\n",
    "    '''\n",
    "    Extracts multiple features using Essentia's Extractor\n",
    "    \n",
    "    Arguments:\n",
    "        - audio_frame:[float] - audio frame\n",
    "    Returns:\n",
    "        - results:pool – Essentia's pool datastructure with aggregated features\n",
    "    '''\n",
    "    extractor = estd.Extractor(highLevel=False, \n",
    "                    midLevel=False,\n",
    "                    relativeIoi=False,\n",
    "                    rhythm=False,\n",
    "                    tuning=False,\n",
    "                    lowLevelFrameSize=1024,\n",
    "                    lowLevelHopSize = 512,\n",
    "                   )\n",
    "    \n",
    "    pool = extractor(audio_frame)\n",
    "    \n",
    "    pool_aggregator = estd.PoolAggregator(defaultStats=['mean','var','dmean','dvar','dmean2','dvar2'])\n",
    "    return flatten_pool(pool_aggregator(pool))\n",
    "\n",
    "\n",
    "def analyse_file(fname, slice_length=100, skip_length=0):\n",
    "    '''\n",
    "    Analyses the file at give filepath\n",
    "    \n",
    "    Arguments:\n",
    "        - fname:path - path to the filename\n",
    "        - slice_length:float - length of a bol slice in ms\n",
    "        - skip_start:float - length of bol beginning to skip in ms\n",
    "    Returns:\n",
    "        - result:[pool] - list of Essentia's pools with extracted features for each bol\n",
    "    '''\n",
    "    fs = 44100\n",
    "    \n",
    "    # Calculate parameters\n",
    "    frame_size = int(slice_length*fs/1000)\n",
    "    skip_samples = int(skip_length*fs/1000)\n",
    "    \n",
    "    # Load audio\n",
    "    audio_file = os.path.join(audio_files_path, fname+'.wav')\n",
    "    audio = estd.MonoLoader(filename=audio_file)()\n",
    "\n",
    "    # Load annotation\n",
    "    annotation_file = os.path.join(annotation_files_path, fname+'.csv')\n",
    "    df = pd.read_csv(annotation_file, header=None)\n",
    "    \n",
    "    # Extract features\n",
    "    bols = []\n",
    "    bols_number = {}\n",
    "    for item in df.itertuples(index=False):\n",
    "        timestamp = item[0]\n",
    "        \n",
    "        # 'Class'\n",
    "        bol_name = item[1].strip()\n",
    "        bol_id = BOL_NAMES.index(bol_name)\n",
    "        if bol_id not in bols_number:\n",
    "            bols_number[bol_id] = 0\n",
    "        bols_number[bol_id] += 1\n",
    "\n",
    "        start_frame = int(np.round(timestamp*fs)) + skip_samples\n",
    "        end_frame = start_frame + frame_size\n",
    "        audio_slice = audio[start_frame : end_frame]\n",
    "        \n",
    "        extractor_pool = extract_features(audio_slice)\n",
    "        mfcc_pool = extract_mfccs(audio_slice)\n",
    "        extractor_pool.merge(mfcc_pool)\n",
    "        extractor_pool.set('bol_id', bol_id)\n",
    "        \n",
    "        bols.append(extractor_pool)\n",
    "        \n",
    "    return bols\n",
    "        "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Perform extraction\n",
    "We are separating bols from audio using annotation data. We cut arbitrary selected 100ms of audio, starting from the annotated onset of a given bol. Such audio slice we treat as a representation of annotated bol from which we extract features. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 96,
   "metadata": {
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "# Extract all features\n",
    "slice_length = 100 # ms\n",
    "skip_length = 0 # ms\n",
    "pools = []\n",
    "for fname in filenames:\n",
    "    pools += analyse_file(fname, slice_length, skip_length)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Export features to a file\n",
    "We are saving features to CSV file in order to avoid the necessity of repeating time-consuming and computationally expensive operation of extracting the features. There are some frames that caused problems during analysis, however the number was neglectable in the scope of our task. Essentia doesn't provide any method of generating CSV from pool, so we had to implement the method ourselves."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 95,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      ".................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................."
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "...................................................Saved features to extracted_features_huge.csv. There was a problem with 1 out of 8245 frames.\n"
     ]
    }
   ],
   "source": [
    "# Save features to CSV file\n",
    "attributes = pools[0].descriptorNames()\n",
    "problematic_frames = 0\n",
    "with open(features_filename, \"w\") as file:\n",
    "    file.write(\",\".join(attributes) + \"\\n\")\n",
    "    attributes.remove('bol_id')\n",
    "    for pool in pools:\n",
    "        print(\".\", end=\"\")\n",
    "        try:\n",
    "            bol_name = BOL_NAMES[int(pool['bol_id'])]\n",
    "            vals = [bol_name]\n",
    "            vals += ([pool[attribute] for attribute in attributes])\n",
    "            file.write(\",\".join([str(item) for item in vals])+\"\\n\")\n",
    "        except KeyError:\n",
    "            problematic_frames += 1\n",
    "            \n",
    "print(\"Saved features to %s. There was a problem with %d out of %d frames.\" % (features_filename, problematic_frames, len(pools)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load features\n",
    "We are loading the extracted features from the file. We decided to use Pandas for further analysis."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Convert 'NA' syllable back to 'NA' (Pandas converts NA values to NaN)\n",
    "df = pd.read_csv(features_filename, converters={'bol_id': lambda x: BOL_NAMES.index(x)})"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Generate representation plot\n",
    "We are plotting the number of samples for each bol in the given dataset."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python3.7/site-packages/pandas/core/ops.py:1167: FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison\n",
      "  result = method(y)\n"
     ]
    },
    {
     "ename": "TypeError",
     "evalue": "invalid type comparison",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mTypeError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-16-33e7173a5abe>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[1;32m      2\u001b[0m \u001b[0mbols_number\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m{\u001b[0m\u001b[0;34m}\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      3\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mname\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mBOL_NAMES\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 4\u001b[0;31m     \u001b[0mbols_number\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mname\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mlen\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mdf\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mdf\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m'bol_id'\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;34m==\u001b[0m \u001b[0mname\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m      5\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      6\u001b[0m \u001b[0mx_bols\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mnp\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0marange\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mlen\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mbols_number\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m/usr/local/lib/python3.7/site-packages/pandas/core/ops.py\u001b[0m in \u001b[0;36mwrapper\u001b[0;34m(self, other, axis)\u001b[0m\n\u001b[1;32m   1281\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1282\u001b[0m             \u001b[0;32mwith\u001b[0m \u001b[0mnp\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0merrstate\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mall\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;34m'ignore'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 1283\u001b[0;31m                 \u001b[0mres\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mna_op\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mvalues\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mother\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m   1284\u001b[0m             \u001b[0;32mif\u001b[0m \u001b[0mis_scalar\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mres\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1285\u001b[0m                 raise TypeError('Could not compare {typ} type with Series'\n",
      "\u001b[0;32m/usr/local/lib/python3.7/site-packages/pandas/core/ops.py\u001b[0m in \u001b[0;36mna_op\u001b[0;34m(x, y)\u001b[0m\n\u001b[1;32m   1167\u001b[0m                     \u001b[0mresult\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mmethod\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0my\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1168\u001b[0m                 \u001b[0;32mif\u001b[0m \u001b[0mresult\u001b[0m \u001b[0;32mis\u001b[0m \u001b[0mNotImplemented\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m-> 1169\u001b[0;31m                     \u001b[0;32mraise\u001b[0m \u001b[0mTypeError\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\"invalid type comparison\"\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m   1170\u001b[0m             \u001b[0;32melse\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1171\u001b[0m                 \u001b[0mresult\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mop\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0my\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mTypeError\u001b[0m: invalid type comparison"
     ]
    }
   ],
   "source": [
    "# For statistical purposes\n",
    "bols_number = {}\n",
    "for name in BOL_NAMES:\n",
    "    bols_number[name] = len(df[df['bol_id'] == name])\n",
    "    \n",
    "x_bols = np.arange(len(bols_number))\n",
    "y_bols = bols_number.values()\n",
    "\n",
    "fig = plt.figure(figsize=(20,10))\n",
    "\n",
    "barplot = plt.bar(x_bols, y_bols)\n",
    "plt.xticks(x_bols, BOL_NAMES, rotation=45, fontsize=14)\n",
    "plt.yticks([])\n",
    "for i, v in enumerate(y_bols):\n",
    "    plt.text(i-.25, v+30, str(v), fontsize=14)\n",
    "plt.show()\n",
    "\n",
    "fig.savefig(os.path.join(results_path, \"dataset_representation.pdf\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Classification\n",
    "We are providing a training and evaluation framework based of scikit-learn. We are then testing performance of different feature sets in combination with different classifiers.\n",
    "### Feature sets we are considering include:\n",
    "- All the features available,\n",
    "- Only MFCCs extracted by Essentia's Extractor class,\n",
    "- Only MFCCs extracted with custom, HTK-based extraction method, together with their derivatives,\n",
    "- All the features excluding MFCC extracted by Essentia's Extractor class.\n",
    "\n",
    "### Classifiers we are testing include:\n",
    "- K-Nearest Neighbour\n",
    "- Support Vector Machine with PolyKernel and gamma scaling\n",
    "- Decision Tree\n",
    "\n",
    "### Training / Testing split\n",
    "1. We are keeping 10% of our data just for testing.\n",
    "2. With the remaining 90% we traing our classifier and perform 10-fold cross validation.\n",
    "3. With a trained classifier perform final classification test and report using following metrics:\n",
    "    - confusion matrix\n",
    "    - precision\n",
    "    - recall\n",
    "    - f1-score\n",
    "    - support\n",
    "    - averages and weighted averages of all mentioned metrics\n",
    "\n",
    "## Training classifiers"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_confusion_matrix(confusion_matrix, class_names, figsize = (10,7), fontsize=11):\n",
    "    \"\"\"Prints a confusion matrix, as returned by sklearn.metrics.confusion_matrix, as a heatmap.\n",
    "    \n",
    "    Arguments\n",
    "    ---------\n",
    "    confusion_matrix: numpy.ndarray\n",
    "        The numpy.ndarray object returned from a call to sklearn.metrics.confusion_matrix. \n",
    "        Similarly constructed ndarrays can also be used.\n",
    "    class_names: list\n",
    "        An ordered list of class names, in the order they index the given confusion matrix.\n",
    "    figsize: tuple\n",
    "        A 2-long tuple, the first value determining the horizontal size of the ouputted figure,\n",
    "        the second determining the vertical size. Defaults to (10,7).\n",
    "    fontsize: int\n",
    "        Font size for axes labels. Defaults to 11.\n",
    "        \n",
    "    Returns\n",
    "    -------\n",
    "    matplotlib.figure.Figure\n",
    "        The resulting confusion matrix figure\n",
    "    \"\"\"\n",
    "    df_cm = pd.DataFrame(\n",
    "        confusion_matrix, index=class_names, columns=class_names, \n",
    "    )\n",
    "    fig = plt.figure(figsize=figsize)\n",
    "    try:\n",
    "        heatmap = sns.heatmap(df_cm, annot=True, fmt=\"d\")\n",
    "    except ValueError:\n",
    "        raise ValueError(\"Confusion matrix values must be integers.\")\n",
    "    heatmap.yaxis.set_ticklabels(heatmap.yaxis.get_ticklabels(), rotation=0, ha='right', fontsize=fontsize)\n",
    "    heatmap.xaxis.set_ticklabels(heatmap.xaxis.get_ticklabels(), rotation=45, ha='center', fontsize=fontsize)\n",
    "    plt.ylabel('True label')\n",
    "    plt.xlabel('Predicted label')\n",
    "    return fig"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "def report(clf, X, y, clf_name, log_to_file=False):\n",
    "    if log_to_file:\n",
    "        # Create results directory if doesn't exist\n",
    "        try:  \n",
    "            os.mkdir(results_path)\n",
    "        except:\n",
    "            # Directory already exists\n",
    "            pass\n",
    "        # Redirect stdout\n",
    "        textfile = open(os.path.join(results_path, clf_name+'.txt'), 'w')\n",
    "        sys.stdout = textfile\n",
    "    \n",
    "    print(\"Results for %s classifier\\n\\n\" % clf_name)\n",
    "        \n",
    "    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.1, random_state=random.randint(0,10000))\n",
    "    k_fold = KFold(10)\n",
    "    print(\"----------------------------------------------------\")\n",
    "    print(\"10-fold validation results\\n\")\n",
    "    for k, (train, test) in enumerate(k_fold.split(X_train, y_train)):\n",
    "        clf.fit(X_train[train], y_train[train])\n",
    "        acc = clf.score(X_train[test], y_train[test])\n",
    "        print(\"Fold no %d: %f accuracy.\" % (k+1, acc))\n",
    "        \n",
    "    y_predict = clf.predict(X_test)\n",
    "    unique_classes = np.unique(list(y_test) +  list(y_predict))\n",
    "    class_names = [BOL_NAMES[class_id] for class_id in sorted(unique_classes)]\n",
    "    clf_report = classification_report(y_test, y_predict, target_names=class_names)\n",
    "    \n",
    "    print(\"----------------------------------------------------\")\n",
    "    print(\"\\nClassification report\\n\")\n",
    "    print(clf_report)\n",
    "    \n",
    "    # Calculate and save confusion matrix figure\n",
    "    cm = confusion_matrix(y_test, y_predict)\n",
    "    print(\"----------------------------------------------------\")\n",
    "    print(cm)\n",
    "    cmfig = print_confusion_matrix(cm, class_names)\n",
    "\n",
    "    if log_to_file:\n",
    "        cmfig.savefig(os.path.join(results_path, clf_name+'--cm.pdf'))\n",
    "        textfile.close()\n",
    "    sys.stdout = JUP_STDOUT"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "# FEATURE SELECTION: Uncomment expected behavior\n",
    "\n",
    "all_columns = list(df.columns)\n",
    "all_columns.remove('bol_id')\n",
    "\n",
    "# All the columns\n",
    "# fset_id = 'all'\n",
    "# selected_columns = all_columns\n",
    "\n",
    "# Only automatically computed mfccs\n",
    "fset_id = 'lowLevel.mfcc'\n",
    "selected_columns = [column for column in all_columns if 'lowLevel.mfcc' in column]\n",
    "\n",
    "# Only htk MFCCs + derivatives\n",
    "# fset_id = 'htk_mfcc'\n",
    "# selected_columns = [column for column in all_columns if 'htk_mfcc' in column]\n",
    "\n",
    "# All features apart of automatically computed mfccs\n",
    "# fset_id = 'not-lowLevel.mfcc'\n",
    "# selected_columns = [column for column in all_columns if 'lowLevel.mfcc' not in column]\n",
    "\n",
    "# Get features matrix\n",
    "X = df[selected_columns].values\n",
    "# Extract classes\n",
    "y = df['bol_id'].values"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Data normalization"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "min_max_scaler = MinMaxScaler()\n",
    "X_norm = min_max_scaler.fit_transform(X)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Feature selction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [],
   "source": [
    "select_k_best = SelectKBest(chi2, k = 150)\n",
    "X_reduced = select_k_best.fit_transform(X_norm, y)\n",
    "\n",
    "# make a list of indices of the selected features and a list of tuples with the index and name of all\n",
    "# of the selected features\n",
    "indices = select_k_best.get_support(indices = True)\n",
    "descriptor_names = [(i, selected_columns[i]) for i in range(len(selected_columns))]\n",
    "descriptor_names = [(i, descriptor_names[i][1]) for i in range(len(indices))]\n",
    "descriptor_names = [c[1] for c in descriptor_names]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Grid search for best parameters of SVN"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "cs = np.logspace(-3, 2, 6)\n",
    "gammas = np.logspace(-3, 2, 6)\n",
    "kernels = ['linear', 'rbf', 'poly']\n",
    "param_grid = {'C': cs, 'gamma' : gammas, 'kernel': kernels}\n",
    "grid_search = GridSearchCV(svm.SVC(), param_grid, cv = 10)\n",
    "\n",
    "grid_search.fit(X_reduced, y)\n",
    "best_params = grid_search.best_params_"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Generating reports"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/usr/local/lib/python3.7/site-packages/sklearn/metrics/classification.py:1143: UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples.\n",
      "  'precision', 'predicted', average, warn_for)\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 720x504 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# K-Nearest Neighbour\n",
    "knn_clf = KNeighborsClassifier()\n",
    "report(knn_clf, X, y, 'knn--'+fset_id, log_to_file=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'lowLevel.mfcc'"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "fset_id"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Support Vector Machine\n",
    "svm_clf = svm.SVC(*best_params)\n",
    "report(svm_clf, X, y, 'svn--'+fset_id)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Descision tree\n",
    "tree = DecisionTreeClassifier()\n",
    "report(tree, X, y, 'tree--'+fset_id, log_to_file=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Conclusions\n",
    "- SVM is the best performing classifier for defined task.\n",
    "- Speech-oriented HTK parameters for MFCC extraction perform suprisingly well for tabla classification problem.\n",
    "- Adding temporal features (MFCC derivatives) may be very important for correct timbral classification.\n",
    "- Spectral features are those that matter the most in case of tabla bol recognition."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Future work\n",
    "### Feature extraction\n",
    "- Harmonium could be separated from the audio using Harmonic / Residual model. This could possibly raise the quality of extracted features.\n",
    "\n",
    "### Classification\n",
    "- Feature selection techniques could be further explored.\n",
    "\n",
    "### Transcription\n",
    "- We could use onset detection coupled with this trained model to do an end-to end tabla transcription."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# References\n",
    "- S. Gupta, A. Srinivasamurthy, M. Kumar, H. A. Murthy, X. Serra. Discovery of Syllabic Percussion Patterns in Tabla Solo Recordings. In Proc. of the 16th International Society for Music Information Retrieval Conference (ISMIR), 2015. http://compmusic.upf.edu/ismir-2015-tabla\n",
    "- P. Chordia. Segmentation and recognition of tabla strokes. In Proc. of the 6th International Conference on Music Information Retrieval (ISMIR), pages 107–114, London, UK, September 2005. ​http://ismir2005.ismir.net/proceedings/1137.pdf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}