MGC-MusiCNN+Sklearn.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Music Genre Classification on the [GTZAN](http://marsyas.info/downloads/datasets.html) dataset using musically relevant feature extractors. \n",
    "\n",
    "We take a transfer learning approach for Music Genre Classification for the [GTZAN](http://marsyas.info/downloads/datasets.html) dataset. We develop a model using [musicnn](https://github.com/jordipons/musicnn) to extract timbral and temporal features appropriate for the task and then use sklearn classifiers to classify them to tags. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 1. Import necessary libraries and find dataset files\n",
    "\n",
    "Also set the correct path to the dataset files"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import glob\n",
    "import random\n",
    "\n",
    "import tqdm\n",
    "import musicnn\n",
    "import joblib\n",
    "import sklearn\n",
    "import h5py\n",
    "import joblib\n",
    "\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "# import xgboost as xgb\n",
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "\n",
    "from collections import Counter, defaultdict\n",
    "\n",
    "from sklearn.model_selection import train_test_split, KFold\n",
    "from sklearn.metrics import confusion_matrix, precision_score \n",
    "from sklearn.metrics import classification_report, average_precision_score, roc_auc_score\n",
    "from sklearn.preprocessing import label_binarize\n",
    "from sklearn.decomposition import PCA\n",
    "from sklearn.multiclass import OneVsRestClassifier\n",
    "from sklearn.svm import SVC\n",
    "from sklearn.pipeline import Pipeline\n",
    "from sklearn.metrics import accuracy_score\n",
    "from xgboost import XGBClassifier\n",
    "\n",
    "from musicnn.extractor import extractor\n",
    "\n",
    "# Set the correct path to the gtzan directory\n",
    "DATA_DIR=\"data/genres\"\n",
    "\n",
    "# Set the musical genres you want to check against\n",
    "TAGS = [\n",
    "    'blues', \n",
    "    'classical', \n",
    "    'country', \n",
    "    'disco', \n",
    "    'hiphop', \n",
    "    'jazz', \n",
    "    'metal', \n",
    "    'pop', \n",
    "    'reggae', \n",
    "    'rock'\n",
    "]\n",
    "\n",
    "# Dictionaries for accessing tags \n",
    "idx2tag = dict((n, TAGS[n]) for n in range(len(TAGS)))\n",
    "tag2idx = dict((TAGS[n], n) for n in range(len(TAGS)))\n",
    "\n",
    "# Audio files with accompanying tags.\n",
    "# Since GTZAN does not provide a split we used the split from https://github.com/jordipons/sklearn-audio-transfer-learning. \n",
    "# This split has been checked for faults and also having a base split helps evaluate the models.\n",
    "# Alternatively we can use trail_test_split from sklearn.model_selection\n",
    "\n",
    "with open('data/split/GTZAN/train_filtered.txt') as f:\n",
    "    TRAIN_DATA = [(fname.strip(), os.path.dirname(fname)) for fname in f.readlines()]\n",
    "with open('data/split/GTZAN/valid_filtered.txt') as f:\n",
    "    VAL_DATA = [(fname.strip(), os.path.dirname(fname)) for fname in f.readlines()]\n",
    "with open('data/split/GTZAN/test_filtered.txt') as f:\n",
    "    TEST_DATA = [(fname.strip(), os.path.dirname(fname)) for fname in f.readlines()]   \n",
    "    \n",
    "FULL_DATA = TRAIN_DATA + VAL_DATA + TEST_DATA"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 2. Extract features using [musicnn](https://github.com/jordipons/musicnn) and save them to disk.\n",
    "\n",
    "We begin by extracting features using [musicnn](https://github.com/jordipons/musicnn) and saving then to `data/gtzan.hdf5` **WARNING** This step takes around 22 minutes to complete."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Create a file to store our data (if it does not exist)\n",
    "\n",
    "if not os.path.exists(\"data/gtzan.hdf5\"):\n",
    "    with h5py.File(\"data/gtzan.hdf5\", \"w\") as f: \n",
    "\n",
    "        print(\"Extracting features for the songs.\")\n",
    "        for fname, tag in tqdm.tqdm_notebook(FULL_DATA):\n",
    "\n",
    "            # Files should be indexed without the full filesystem path.\n",
    "            basefname = os.path.splitext(os.path.basename(fname))[0]\n",
    "\n",
    "            # Extract the max-pool layer from the original model.\n",
    "            _, _, features           = extractor(os.path.join(DATA_DIR,fname), \n",
    "                                       model='MSD_musicnn', \n",
    "                                       #input_length=3, \n",
    "                                       input_overlap=1,\n",
    "                                       extract_features=True)\n",
    "\n",
    "            f.create_dataset(basefname, data=features['max_pool'])    \n",
    "\n",
    "    print(\"Done.\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 3. Load audio files, extracting features and saving to a `.hdf5` store\n",
    "\n",
    "Some helper functions easily access features in our dataset:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "def song_to_features(fname: str, dfile=f) -> np.ndarray:\n",
    "    \"\"\"\n",
    "        Takes a song filename and returns \n",
    "        the features from the stored dataset. \n",
    "        \n",
    "        :param fname: The song filename (must exist in the dataset)\n",
    "        :param dfile: an open h5py file\n",
    "        :return: An array with shape (n_frames, n_features)\n",
    "    \"\"\"\n",
    "    \n",
    "    # Files are indexed based on the basename (i.e. without the filesystem path)\n",
    "    basefname = os.path.splitext(os.path.basename(fname))[0]\n",
    "    \n",
    "    return np.array(f[basefname])\n",
    "\n",
    "def data_list_to_features_tags_fnames(l: list, tag2idx=tag2idx, dfile=f) -> (np.ndarray, np.ndarray):\n",
    "    \"\"\"\n",
    "        Takes a list of the form [(fname1, tag1), (fname2, tag2), ...] and returns \n",
    "        a matrix X of features and y with integers representing each tag. \n",
    "        \n",
    "        :param      l: A list of the form [(fname1, tag2), (fname2, tag2), ...]\n",
    "        :param tag2dx: A dictionary that returns a unique integer for each tag in l\n",
    "        :param  dfile: an open h5py file\n",
    "        :return: X is a features matrix of shape (n_total_frames, n_features), y\n",
    "        is an integer labels array of shape (n_total_frames, ), and fnames is a list\n",
    "        of the filenames it corresponds to.\n",
    "    \"\"\"\n",
    "    \n",
    "    X = []\n",
    "    y = []\n",
    "    fnames = []\n",
    "    \n",
    "    for fname, tag in l:\n",
    "        \n",
    "        features = song_to_features(fname, f)\n",
    "        X.append(\n",
    "            features\n",
    "        ) \n",
    "        \n",
    "        y.append(\n",
    "            np.ones((features.shape[0],))*tag2idx[tag]\n",
    "        )\n",
    "        \n",
    "        fnames += [fname for _ in range(features.shape[0])]\n",
    "        \n",
    "    return np.vstack(X), np.hstack(y).astype(np.int), fnames"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Load features and labels for our splits from the stored dataset:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "with h5py.File('data/gtzan.hdf5', 'r') as f:\n",
    "    X_train, y_train, fnames_train = data_list_to_features_tags_fnames(TRAIN_DATA)\n",
    "    X_val, y_val, fnames_val = data_list_to_features_tags_fnames(VAL_DATA)\n",
    "    X_test, y_test, fnames_test = data_list_to_features_tags_fnames(TEST_DATA)\n",
    "    \n",
    "    sample_features = song_to_features('blues.00001.wav')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Visualize a song's features:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Text(33.0, 0.5, 'feature #')"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "sns.heatmap(sample_features.T)\n",
    "plt.title('Features extracted from MSD_musicnn')\n",
    "plt.xlabel('frame #')\n",
    "plt.ylabel('feature #')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Helper functions for computing song-tag predictions from frame-tag predictions and plotting confusion matrices."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "from collections import Counter, defaultdict\n",
    "\n",
    "def frame_predictions_to_song_predictions(y_labels: list, \n",
    "                                          fname_list: list, \n",
    "                                         ) -> dict:\n",
    "    \"\"\"\n",
    "        Accepts a list of prediction labels `y_labels' and \n",
    "        the corresponding filename list `fname_list' and returns \n",
    "        a dict of frequencies for each label and each song.\n",
    "        \n",
    "        :param y_labels: list of tags (e.g. ['blues', 'blues', ...]) for each predicted frame\n",
    "        :param fname_list: list of the corresponding filenames (e.g. ['blues.0000.wav',...])\n",
    "        :returns: a dictionary of the form {'blues.0000.wav': 'blues', ...}\n",
    "        \n",
    "    \"\"\"\n",
    "    \n",
    "    \n",
    "    fname2tags = defaultdict(Counter)\n",
    "    fname2tag  = dict()\n",
    "    \n",
    "    \n",
    "    # First pass, get frequencies for each tag\n",
    "    for n, fname in enumerate(fname_list):\n",
    "        fname2tags[fname][y_labels[n]] += 1  \n",
    "        \n",
    "    # Second pass, get the most common tag for each song\n",
    "    for fname in fname2tags:\n",
    "        fname2tag[fname] = fname2tags[fname].most_common()[0][0]\n",
    "        \n",
    "    return fname2tag\n",
    "\n",
    "def song_prediction_dict_to_lists(y_dict_true: dict,\n",
    "                                  y_dict_pred: dict,\n",
    "                                  fnames=None,\n",
    "                                 ) -> (list,list,list):\n",
    "    \"\"\"\n",
    "        Accepts two dictionaries of the form. {'blues.0000.wav' : blues, 'rock.0000.wav'}\n",
    "        one holding the ground truth and one the predictions and returns two lists\n",
    "        `y_true', `y_pred' aligned with a third list of filenames `fnames'.\n",
    "        \n",
    "        :param y_dict_true: dictionary of ground truth \n",
    "        :param y_dict_pred: dictionary of predicted values\n",
    "        :param fnames: if not None, then it aligns predictions to that list.\n",
    "        :returns: two lists `y_pred', `y_true' of predictions as well as \n",
    "        a third list `fnames' containing the filenames for both.\n",
    "    \"\"\"\n",
    "    \n",
    "    if not fnames:\n",
    "        fnames = y_dict_true.keys()\n",
    "    y_true = [y_dict_true[p] for p in fnames]\n",
    "    y_pred = [y_dict_pred[p] for p in fnames]\n",
    "    \n",
    "    return y_true, y_pred, fnames\n",
    "\n",
    "def get_song_taggrams(y_prob: np.ndarray,\n",
    "                      fname_list: list) -> dict:\n",
    "    \"\"\"\n",
    "        Gets the song taggram from a y_prob array \n",
    "    \"\"\"\n",
    "    \n",
    "    fname2arrlists = defaultdict(list)\n",
    "    fname2taggram = dict()\n",
    "    \n",
    "    # First pass, gather the probabilities for each frame\n",
    "    # for each tag for each song.\n",
    "    for n, fname in enumerate(fname_list):\n",
    "        fname2arrlists[fname].append(y_prob[n, :])\n",
    "        \n",
    "    for fname in fname2arrlists:\n",
    "        fname2taggram[fname] = np.vstack(fname2arrlists[fname])\n",
    "        \n",
    "    return fname2taggram\n",
    "\n",
    "def taggrams_to_songs(taggrams: np.ndarray, idx2tag=idx2tag) -> dict:\n",
    "    \"\"\"\n",
    "        Accepts a taggrams dict and returns a dict that\n",
    "        maps each song to a tag\n",
    "    \"\"\"\n",
    "    \n",
    "    fname2song = dict()\n",
    "    \n",
    "    for fname in taggrams:\n",
    "        fname2song[fname] = idx2tag[np.argmax(taggrams[fname].sum(axis=0))]\n",
    "        \n",
    "    return fname2song\n",
    "\n",
    "def taggrams_to_proba(taggrams: np.ndarray, idx2tag=idx2tag) -> dict:\n",
    "    \"\"\"\n",
    "        Accepts a taggrams dict and returns a dict that\n",
    "        maps each song to a vector of probabilities. Used for\n",
    "        computing auc scores and average precision score.\n",
    "    \"\"\"\n",
    "    \n",
    "    fname2proba = dict()\n",
    "    \n",
    "    for fname in taggrams:\n",
    "        fname2proba[fname] = taggrams[fname].mean(axis=0)\n",
    "        \n",
    "    return fname2proba\n",
    "    \n",
    "\n",
    "\n",
    "def plot_confusion_matrix(y_true: list, y_pred: list) -> None:\n",
    "    \"\"\"\n",
    "        Plots a confusion matrix using seaborn\n",
    "        \n",
    "        :param y_true: the true labels\n",
    "        :param y_pred: the predicted labels\n",
    "        :param labels: \n",
    "    \"\"\"\n",
    "    \n",
    "    cf_mtx = confusion_matrix(y_true, y_pred)\n",
    "\n",
    "    # https://stackoverflow.com/questions/35572000/how-can-i-plot-a-confusion-matrix \n",
    "    cf_df = pd.DataFrame(cf_mtx, index=TAGS, columns=TAGS)\n",
    "    sns.heatmap(cf_df, annot=True)\n",
    "    \n",
    "def plot_taggram(fname:str, taggrams:np.ndarray, labels=TAGS):\n",
    "    \"\"\"\n",
    "        Plots a taggram \n",
    "    \"\"\"\n",
    "    sns.heatmap(taggrams[fname].T, yticklabels=labels)\n",
    "    plt.title('Tag-gram for {}'.format(fname))\n",
    "    plt.xlabel('feature #')\n",
    "    plt.ylabel('tag')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## *Training* and model selection using the *validation* split.\n",
    "\n",
    "Below we use the validation split to try different models. \n",
    "\n",
    "First try: musicnn (max_pool layer) + 128 PC + SVM with RBF kernel reports 77.24% accuracy\n",
    "in https://github.com/jordipons/sklearn-audio-transfer-learning. Let's try that on our validation set."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "              precision    recall  f1-score   support\n",
      "\n",
      "       blues       0.61      0.82      0.70        17\n",
      "   classical       1.00      0.95      0.98        21\n",
      "     country       0.43      0.62      0.51        16\n",
      "       disco       0.95      0.81      0.88        26\n",
      "      hiphop       1.00      0.60      0.75        30\n",
      "        jazz       0.94      0.70      0.80        23\n",
      "       metal       1.00      0.87      0.93        23\n",
      "         pop       0.38      0.45      0.42        11\n",
      "      reggae       0.59      0.77      0.67        13\n",
      "        rock       0.46      0.65      0.54        17\n",
      "\n",
      "    accuracy                           0.74       197\n",
      "   macro avg       0.74      0.72      0.72       197\n",
      "weighted avg       0.80      0.74      0.75       197\n",
      "\n",
      "CPU times: user 4min 42s, sys: 56.6 ms, total: 4min 42s\n",
      "Wall time: 4min 40s\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "clf = Pipeline(steps=[\n",
    "    ('pca', PCA(128, whiten=True)), \n",
    "    \n",
    "    # While SVC is multiclass by default, we need to output probabilities in\n",
    "    # order to be able to show taggrams, and compute average precision scores.\n",
    "    ('svm', OneVsRestClassifier(\n",
    "        SVC(kernel='rbf', C=1, gamma='scale', probability=True)\n",
    "            )\n",
    "    )\n",
    "])\n",
    "\n",
    "\n",
    "# Train classifier\n",
    "y_val_score = clf.fit(X_train, y_train).decision_function(X_val)\n",
    "\n",
    "# Predict validation \n",
    "y_pred = clf.predict(X_val)\n",
    "\n",
    "# Convert to labels\n",
    "y_pred_labels = [TAGS[p] for p in y_pred]\n",
    "y_val_labels = [TAGS[p] for p in y_val]\n",
    "\n",
    "# Convert frame-tag predictions to song-tag predictions \n",
    "s_pred_dict = frame_predictions_to_song_predictions(y_pred_labels, fnames_val)\n",
    "s_val_dict = frame_predictions_to_song_predictions(y_val_labels, fnames_val)\n",
    "s_val, s_pred, s_fnames = song_prediction_dict_to_lists(s_pred_dict, s_val_dict)\n",
    "\n",
    "plot_confusion_matrix(s_val, s_pred)\n",
    "print(classification_report(s_val, s_pred, ))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Results reporting in the *testing* split for the SVM classifier\n",
    "\n",
    "For MCG we usually report average precision, and ROC-AUC values, as well as show confusion matrices. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "              precision    recall  f1-score   support\n",
      "\n",
      "       blues       0.77      0.32      0.45       868\n",
      "   classical       1.00      1.00      1.00       868\n",
      "     country       0.96      0.87      0.91       840\n",
      "       disco       0.74      0.86      0.79       812\n",
      "      hiphop       0.65      0.89      0.75       761\n",
      "        jazz       0.57      1.00      0.73       756\n",
      "       metal       0.90      1.00      0.95       756\n",
      "         pop       0.78      0.70      0.74       840\n",
      "      reggae       0.86      0.69      0.77       728\n",
      "        rock       0.83      0.59      0.69       898\n",
      "\n",
      "    accuracy                           0.79      8127\n",
      "   macro avg       0.80      0.79      0.78      8127\n",
      "weighted avg       0.81      0.79      0.78      8127\n",
      "\n",
      "Average accuracy: 78.60\n",
      "Average precision from PR curve: 0.8822\n",
      "ROC-AUC: 0.9740\n",
      "CPU times: user 46.9 s, sys: 40 ms, total: 47 s\n",
      "Wall time: 45.9 s\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "%%time \n",
    "\n",
    "# Predict testing \n",
    "y_pred = clf.predict(X_test)\n",
    "\n",
    "# Convert to labels\n",
    "y_pred_labels = [TAGS[p] for p in y_pred]\n",
    "y_test_labels = [TAGS[p] for p in y_test]\n",
    "\n",
    "# Taggrams require probabilities\n",
    "y_proba = clf.predict_proba(X_test)\n",
    "\n",
    "# Extract all taggrams for the predicted songs\n",
    "taggrams_pred = get_song_taggrams(y_proba, fnames_test)\n",
    "s_pred_dict = taggrams_to_songs(taggrams_pred)\n",
    "\n",
    "# Convert frame-by-frame tagged labels to a dictionary that maps songs to tags\n",
    "# for both the testing set and the predicted values\n",
    "s_test_dict = frame_predictions_to_song_predictions(y_test_labels, fnames_test)\n",
    "s_test, s_pred, s_fnames = song_prediction_dict_to_lists(s_test_dict, s_pred_dict, fnames_test )\n",
    "\n",
    "# One-hot representation is used for computing average precision and auc scores \n",
    "s_test_onehot = np.zeros((len(s_test), len(TAGS)))\n",
    "for n in range(len(s_test)):\n",
    "    s_test_onehot[n,tag2idx[s_test[n]]] = 1.0\n",
    "    \n",
    "# Convert the predicted taggram to probabilities\n",
    "s_pred_proba_dict = taggrams_to_proba(taggrams_pred)\n",
    "\n",
    "s_pred_proba = []\n",
    "for fname in fnames_test:\n",
    "    s_pred_proba.append(s_pred_proba_dict[fname])\n",
    "s_pred_proba = np.vstack(s_pred_proba)\n",
    "\n",
    "plot_confusion_matrix(s_test, s_pred)\n",
    "print(classification_report(s_test, s_pred))\n",
    "\n",
    "print(\"Average accuracy: {:.2f}\".format(100*accuracy_score(s_test, s_pred)))\n",
    "print(\"Average precision from PR curve: {:.4f}\".format(average_precision_score(s_test_onehot, s_pred_proba)))\n",
    "print(\"ROC-AUC: {:.4f}\".format(roc_auc_score(s_test_onehot, s_pred_proba)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Second try: `XGBoost`\n",
    "\n",
    "`XGBoost` is an algorithm implementing ensembles of decision trees. It is supposed to be distributed and efficient to train.\n",
    "\n",
    "Training parameters: musicnn (max_pool layer) + 128 PC + xgboost"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "              precision    recall  f1-score   support\n",
      "\n",
      "       blues       0.52      0.80      0.63        15\n",
      "   classical       1.00      0.95      0.98        21\n",
      "     country       0.52      0.71      0.60        17\n",
      "       disco       0.91      0.74      0.82        27\n",
      "      hiphop       1.00      0.69      0.82        26\n",
      "        jazz       0.94      0.80      0.86        20\n",
      "       metal       1.00      0.83      0.91        24\n",
      "         pop       0.46      0.60      0.52        10\n",
      "      reggae       0.53      0.69      0.60        13\n",
      "        rock       0.54      0.54      0.54        24\n",
      "\n",
      "    accuracy                           0.74       197\n",
      "   macro avg       0.74      0.74      0.73       197\n",
      "weighted avg       0.79      0.74      0.75       197\n",
      "\n",
      "CPU times: user 3min 28s, sys: 243 ms, total: 3min 29s\n",
      "Wall time: 26.2 s\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "from xgboost import XGBClassifier\n",
    "\n",
    "clf = Pipeline(steps=[\n",
    "    ('pca', PCA(128, whiten=True)), \n",
    "    \n",
    "    # We use the default options here\n",
    "    ('xgb', OneVsRestClassifier(\n",
    "                XGBClassifier()\n",
    "            )\n",
    "    )\n",
    "])\n",
    "\n",
    "\n",
    "# Train classifier\n",
    "clf.fit(X_train, y_train)\n",
    "\n",
    "# Predict validation \n",
    "y_pred = clf.predict(X_val)\n",
    "\n",
    "# Convert to labels\n",
    "y_pred_labels = [TAGS[p] for p in y_pred]\n",
    "y_val_labels = [TAGS[p] for p in y_val]\n",
    "\n",
    "# Convert frame-tag predictions to song-tag predictions \n",
    "s_pred_dict = frame_predictions_to_song_predictions(y_pred_labels, fnames_val)\n",
    "s_val_dict = frame_predictions_to_song_predictions(y_val_labels, fnames_val)\n",
    "s_val, s_pred, s_fnames = song_prediction_dict_to_lists(s_pred_dict, s_val_dict)\n",
    "\n",
    "plot_confusion_matrix(s_val, s_pred)\n",
    "print(classification_report(s_val, s_pred, ))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Results reporting in the *testing* split\n",
    "\n",
    "Again: average precision, and ROC-AUC values, as well as show confusion matrices. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "              precision    recall  f1-score   support\n",
      "\n",
      "       blues       0.68      0.42      0.52       868\n",
      "   classical       0.97      1.00      0.98       868\n",
      "     country       0.87      0.87      0.87       840\n",
      "       disco       0.81      0.86      0.83       812\n",
      "      hiphop       0.63      0.89      0.74       761\n",
      "        jazz       0.63      0.96      0.76       756\n",
      "       metal       0.90      0.96      0.93       756\n",
      "         pop       0.81      0.70      0.75       840\n",
      "      reggae       0.84      0.62      0.71       728\n",
      "        rock       0.76      0.59      0.67       898\n",
      "\n",
      "    accuracy                           0.78      8127\n",
      "   macro avg       0.79      0.79      0.78      8127\n",
      "weighted avg       0.79      0.78      0.77      8127\n",
      "\n",
      "Average accuracy: 78.26\n",
      "Average precision from PR curve: 0.8861\n",
      "ROC-AUC: 0.9707\n",
      "CPU times: user 4.24 s, sys: 86.6 ms, total: 4.33 s\n",
      "Wall time: 569 ms\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 2 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "# Predict testing \n",
    "y_pred = clf.predict(X_test)\n",
    "\n",
    "# Convert to labels\n",
    "y_pred_labels = [TAGS[p] for p in y_pred]\n",
    "y_test_labels = [TAGS[p] for p in y_test]\n",
    "\n",
    "# Taggrams require probabilities\n",
    "y_proba = clf.predict_proba(X_test)\n",
    "\n",
    "# Extract all taggrams for the predicted songs\n",
    "taggrams_pred = get_song_taggrams(y_proba, fnames_test)\n",
    "s_pred_dict = taggrams_to_songs(taggrams_pred)\n",
    "\n",
    "# Convert frame-by-frame tagged labels to a dictionary that maps songs to tags\n",
    "# for both the testing set and the predicted values\n",
    "s_test_dict = frame_predictions_to_song_predictions(y_test_labels, fnames_test)\n",
    "s_test, s_pred, s_fnames = song_prediction_dict_to_lists(s_test_dict, s_pred_dict, fnames_test )\n",
    "\n",
    "# One-hot representation is used for computing average precision and auc scores \n",
    "s_test_onehot = np.zeros((len(s_test), len(TAGS)))\n",
    "for n in range(len(s_test)):\n",
    "    s_test_onehot[n,tag2idx[s_test[n]]] = 1.0\n",
    "    \n",
    "# Convert the predicted taggram to probabilities\n",
    "s_pred_proba_dict = taggrams_to_proba(taggrams_pred)\n",
    "\n",
    "s_pred_proba = []\n",
    "for fname in fnames_test:\n",
    "    s_pred_proba.append(s_pred_proba_dict[fname])\n",
    "s_pred_proba = np.vstack(s_pred_proba)\n",
    "\n",
    "plot_confusion_matrix(s_test, s_pred)\n",
    "print(classification_report(s_test, s_pred))\n",
    "\n",
    "print(\"Average accuracy: {:.2f}\".format(100*accuracy_score(s_test, s_pred)))\n",
    "print(\"Average precision from PR curve: {:.4f}\".format(average_precision_score(s_test_onehot, s_pred_proba)))\n",
    "print(\"ROC-AUC: {:.4f}\".format(roc_auc_score(s_test_onehot, s_pred_proba)))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Results on the *testing* split\n",
    "\n",
    "|Classifier                    |Avg. Acc| Avg. P| ROC-AUC| Time (training) | Time (testing) |\n",
    "|------------------------------|--------|-------|--------|----   | ----|\n",
    "|SVM(RBF,C=1)<sup>1</sup>|0.7724   | --    |  --    |      ---     |--- |\n",
    "|SVM(RBF,C=1)  |         0.7860   | 0.8822      | 0.9740       |4m40s| 45.9s|\n",
    "|XGBoost       |         0.7826   |  0.8861     | 0.9707       |26.2s| 569ms |\n",
    "\n",
    "We achieve an average best accuracy of 78.60% which is more than the 77.24% reported in https://github.com/jordipons/sklearn-audio-transfer-learning. \n",
    "\n",
    "Also while avg. accuracy is better in the case of an SVM, XGBoost is 2 orders of magnitude faster regarding training and testing times, and should be the classifier chosen for production.\n",
    "\n",
    "We observe that blues are easily confused for jazz. Pop and reggae can be sometimes also confused for hiphop. While these results are not entirely unsurprising (humans can also confuse jazz and blues) some study into this is warranted.\n",
    "\n",
    "Taggrams are probabilities of each frame that the song is a specific tag. In the diagrams below lighter boxes mean higher probabilities. \n",
    "\n",
    "<sup>1</sup>https://github.com/jordipons/sklearn-audio-transfer-learning."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1080x432 with 6 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# Plot some taggrams\n",
    "plt.figure(figsize=(15,6))\n",
    "\n",
    "plt.subplot(1,3,1)\n",
    "plot_taggram('blues/blues.00012.wav', taggrams_pred)\n",
    "plt.subplot(1,3,2)\n",
    "plot_taggram('jazz/jazz.00073.wav', taggrams_pred)\n",
    "plt.subplot(1,3,3)\n",
    "plot_taggram('classical/classical.00079.wav', taggrams_pred)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Final step: Save the trained model \n",
    "\n",
    "We save the final classifier in order to use it for deployment. We used a OneVsRest classification scheme to be able to derive probabilities and scores and evaluate it. We save this and also the multi-class (stock) classifier directly (without the OVR wrapper) which might behave slightly differently but is a lot faster:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 7min, sys: 363 ms, total: 7min 1s\n",
      "Wall time: 52.8 s\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "['models/features_classifier.pkl']"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "clf_ovr = Pipeline(steps=[\n",
    "    ('pca', PCA(128, whiten=True)), \n",
    "    ('xgb', OneVsRestClassifier(\n",
    "            XGBClassifier()\n",
    "        )\n",
    "    )\n",
    "])\n",
    "\n",
    "clf = Pipeline(steps=[\n",
    "    ('pca', PCA(128, whiten=True)), \n",
    "    ('xgb', XGBClassifier()),\n",
    "])\n",
    "\n",
    "clf_ovr.fit(X_train, y_train)\n",
    "clf.fit(X_train, y_train)\n",
    "\n",
    "joblib.dump(clf_ovr, 'models/features_classifier_ovr.pkl')\n",
    "joblib.dump(clf, 'models/features_classifier.pkl')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Below we compare times for the OVR and the stock classifiers:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 838 ms, sys: 3.36 ms, total: 841 ms\n",
      "Wall time: 105 ms\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "y_pred = clf.predict(X_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 1.49 s, sys: 26.6 ms, total: 1.52 s\n",
      "Wall time: 190 ms\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "y_pred = clf_ovr.predict(X_test)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Demo: Steps to evaluate an audio file"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Computing spectrogram (w/ librosa) and tags (w/ tensorflow).. done!\n",
      "Assigned genre: classical\n",
      "CPU times: user 5.39 s, sys: 1.88 s, total: 7.27 s\n",
      "Wall time: 18.5 s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "# Specify filename \n",
    "fname = 'data/genres/classical/classical.00006.wav'\n",
    "\n",
    "# Extract the `max_pool' features using musicnn\n",
    "_, _, features           = extractor(fname, \n",
    "                           model='MSD_musicnn', \n",
    "                           input_overlap=1,\n",
    "                           extract_features=True)\n",
    "\n",
    "maxpool_features = features['max_pool']\n",
    "\n",
    "# Classify each feature with the trained classifier\n",
    "y_pred = clf.predict(maxpool_features)\n",
    "\n",
    "# Assign a string tag (e.g. 'classical' instead of 1) to each predicted class\n",
    "y_pred_labels = [idx2tag[n] for n in y_pred]\n",
    "\n",
    "# The assigned genre is the most common tag\n",
    "tag = Counter(y_pred_labels).most_common()[0][0]\n",
    "\n",
    "print('Assigned genre:', tag)"
   ]
  }
 ],
 "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": "3.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}