reorganize project

datasets.py → datasets/datasets.py

 from utils import *
 from torch.utils.data import Dataset, DataLoader
+from processors.spectrogram_processors import make_framed_spec
 
 
 class MelSpecDataset(Dataset):
-    def __init__(self, phase='train', ann_root=None, spec_root=None, length=MAX_FRAMES):
+    def __init__(self, phase='train', ann_root=None, spec_root=None, length=MAX_FRAMES, framed=True):
         assert ann_root is not None, logger.error("ann_root (root directory containing annotation files) required")
         assert spec_root is not None, logger.error("spec_root (root directory of spectrograms) required")
         assert phase in ['train', 'validation', 'test'], \
@@ -12,6 +13,7 @@ class MelSpecDataset(Dataset):
         self.ann_root = ann_root
         self.spec_root = spec_root
         self.length = length
+        self.framed = framed
 
         xy = pd.read_csv(os.path.join(self.ann_root, f'{phase}_processed.tsv'), sep='\t')
         self.len = len(xy)
@@ -27,6 +29,9 @@ class MelSpecDataset(Dataset):
         returns framed mel spectrogram and multi-hot encoded labels
         """
         x_melspec = np.load(os.path.join(self.spec_root, self.x_path[index]))
+        if not self.framed:
+            x_melspec = make_framed_spec(x_melspec, frame_length=256, hop=1.0)
+
         reqd_len = self.length
         spec_len = len(x_melspec)
         x_trimmed = x_melspec[:reqd_len] if spec_len>reqd_len else \
@@ -51,12 +56,15 @@ class MelSpecDataset(Dataset):
             return np.array([])
         return tagslist
 
-    def _make_unique_tags_list_(self, labels, saveto='./tagslist'):
+    def _make_unique_tags_list_(self, labels, saveto=None):
         labelslist = []
         for label in labels:
             labelslist.extend(label)
         tagslist = np.sort(np.unique(np.array(labelslist)))
-        if saveto is not None:
+        if saveto is None:
+            saveto = os.path.join(PATH_PROJECT_ROOT, 'tagslist')
+            np.save(saveto, tagslist)
+        else:
             np.save(saveto, tagslist)
         return tagslist
 
@@ -64,7 +72,8 @@ class MelSpecDataset(Dataset):
 if __name__=='__main__':
     # Tests
     torch.manual_seed(6)
-    dataset = MelSpecDataset(phase='train', ann_root=PATH_ANNOTATIONS, spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED)
+    dataset = MelSpecDataset(phase='train', ann_root=PATH_ANNOTATIONS,
+                             spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED, framed=True)
     train_loader = DataLoader(dataset=dataset,
                               batch_size=32,
                               shuffle=True)

experiment_baseline.py → experiments/experiment_baseline.py

@@ -17,7 +17,7 @@ def run():
         trainer = Trainer(experiment=exp, max_nb_epochs=1, train_percent_check=0.1,
                           fast_dev_run=True)
 
-    model = Network()  # TODO num_class
+    model = Network(num_class=56)  # TODO num_class
 
     print(model)
 

experiment_erf.py → experiments/experiment_erf.py

@@ -3,8 +3,8 @@ from pytorch_lightning import Trainer
 from test_tube import Experiment
 from models.cp_resnet import Network
 import torch
-from datasets import MelSpecDataset
-from torch.utils.data import Dataset, DataLoader
+from datasets.datasets import MelSpecDataset
+from torch.utils.data import DataLoader
 
 from matplotlib.transforms import Affine2D
 import mpl_toolkits.axisartist.floating_axes as floating_axes
@@ -96,7 +96,7 @@ def ERF_generate(model, loader):
             accum += me
         counter += 1
     # torch.save({"arr": accum, "counter": counter}, os.path.join(self.config.out_dir, 'ERF_dict.pth'))
-    ERF_plot(accum, savefile=os.path.join('/home/verena/experiments/moodwalk', 'erf.png'))
+    ERF_plot(accum, savefile=os.path.join(CURR_RUN_PATH, 'erf.png'))
     # self.experiment.add_artifact(os.path.join(self.config.out_dir, 'erf.png'), "erf.png", {"dataset": dataset_name})
     return True
 

experiments/main.py

+from utils import *
+from pytorch_lightning import Trainer
+from pytorch_lightning.utilities import arg_parse
+from pytorch_lightning.callbacks.pt_callbacks import ModelCheckpoint
+from test_tube import Experiment
+from models import vgg_basic
+import argparse
+
+
+def run(hparams):
+    logger.info(CURR_RUN_PATH)
+
+    exp = Experiment(
+        name=hparams.tt_name,
+        debug=hparams.debug,
+        save_dir=CURR_RUN_PATH,
+        version='1',
+        autosave=False,
+        description=hparams.tt_description
+    )
+    exp.argparse(hparams)
+    exp.save()
+
+    # model = cp_resnet.Network(model_config)
+    model = vgg_basic.Network()
+
+    model_save_path = '{}/{}/{}'.format(hparams.model_save_path, exp.name, exp.version)
+    checkpoint = ModelCheckpoint(
+        filepath=model_save_path,
+        save_best_only=True,
+        verbose=True,
+        monitor='rocauc',
+        mode='max'
+    )
+
+    if USE_GPU:
+        trainer = Trainer(gpus=[1], distributed_backend='ddp',
+                          experiment=exp, max_nb_epochs=10, train_percent_check=1.0,
+                          fast_dev_run=False, checkpoint_callback=checkpoint)
+    else:
+        trainer = Trainer(experiment=exp, max_nb_epochs=1, train_percent_check=0.1,
+                          fast_dev_run=True)
+    trainer.fit(model)
+
+
+if __name__=='__main__':
+    parent_parser = argparse.ArgumentParser(description='hyperparameters')
+    arg_parse.add_default_args(parent_parser, root_dir=CURR_RUN_PATH)
+    parser = vgg_basic.Network.add_model_specific_args(parent_parser, root_dir=CURR_RUN_PATH)
+    hyperparams = parser.parse_args()
+    run(hyperparams)
\ No newline at end of file

plotting.py → helpers/plotting.py

 from matplotlib import pyplot as plt
 import numpy as np
+from utils import *
 
-def plot_tag_frequencies(df, norm=True, index_sort=True, out=None):
+def plot_tag_frequencies(df, norm=True, index_sort=True, out=None, title=None):
     tag_freqs = df['TAGS'].str.split(',', expand=True).stack().value_counts()
-    title = 'tag counts'
+    title_ = 'tag counts'
     if norm:
         tag_freqs = tag_freqs / max(tag_freqs)
-        title = f'normalized tag frequencies'
+        title_ = f'normalized tag frequencies'
     if index_sort:
         tag_freqs = tag_freqs.sort_index()
-    tag_freqs.plot.bar(title=title)
+
+    from itertools import islice, cycle
+    colors = list(islice(cycle(['b', 'r', 'g', 'y', 'k']), None, len(tag_freqs)))
+
+    if title is not None:
+        title_=title
+    tag_freqs.plot.bar(title=title_, color=colors)
 
     if out is None:
         plt.show()
@@ -38,4 +45,10 @@ def plot_melspectrogram(melspec, scale=False, title=None, out=None):
     if out is None:
         plt.show()
     else:
-        plt.savefig(out)
\ No newline at end of file
+        plt.savefig(out)
+
+if __name__=='__main__':
+    dataset_path = os.path.join(PATH_ANNOTATIONS, 'test_processed.tsv')
+    df = pd.read_csv(dataset_path, sep='\t')
+    plot_tag_frequencies(df, title='normalized tag frequencies, test')
+

helpers/stats.py

+from utils import *
+
+def compute_duration_stats(filepath):
+    df = pd.read_csv(filepath, sep='\t')
+    total_secs = df.DURATION.sum()
+    hours = int(total_secs//3600)
+    mins = int((total_secs%3600)//60)
+    secs = np.round(total_secs%60, 1)
+    return {'average':time.strftime('%H:%M:%S', time.gmtime(df.DURATION.mean())),
+            'total':f'{hours}:{mins}:{secs}'} #todo: improve format for better readability
+
+
+def compute_melspec_length_stats(filepath):
+    df = pd.read_csv(filepath, sep='\t')
+    lengths = []
+    paths = df.PATH.str.split(pat='.').str[0]+'.npy'
+    for i in tqdm(df.index):
+        x_melspec = np.load(os.path.join(PATH_MELSPEC_DOWNLOADED, paths[i]))
+        lengths.append(x_melspec.shape[1])
+    lengths = np.array(lengths)
+    return {'avg':lengths.mean(), 'max':lengths.max(), 'min':lengths.min()}
+
+
+def compute_tag_stats():
+    trainset_path = os.path.join(PATH_ANNOTATIONS, 'train_processed.tsv')
+    validset_path = os.path.join(PATH_ANNOTATIONS, 'validation_processed.tsv')
+    testset_path = os.path.join(PATH_ANNOTATIONS, 'test_processed.tsv')
+
+    print(f"\nComputing stats for:\n{trainset_path}\n{validset_path}\n{testset_path}\n")
+
+    trainset = pd.read_csv(trainset_path, sep='\t')
+    validset = pd.read_csv(validset_path, sep='\t')
+    testset = pd.read_csv(testset_path, sep='\t')
+
+    trainset_tags = trainset.TAGS.str.split(pat=',')
+    validset_tags = validset.TAGS.str.split(pat=',')
+    testset_tags = testset.TAGS.str.split(pat=',')
+
+    trainset_tags_unique = np.sort(np.unique(np.hstack(trainset_tags)))
+    validset_tags_unique = np.sort(np.unique(np.hstack(validset_tags)))
+    testset_tags_unique = np.sort(np.unique(np.hstack(testset_tags)))
+
+    print("TRAINING SET")
+    print(f"Number of songs \t\t\t\t{len(trainset)}")
+    print(f"Number of unique tags\t\t\t{len(trainset_tags_unique)}")
+    print(f"Avg number of tags per song\t\t{np.round(trainset_tags.apply(len).mean(),2)}")
+    print(f"Max number of tags per song\t\t{trainset_tags.apply(len).max()}")
+    print(f"Min number of tags per song\t\t{trainset_tags.apply(len).min()}")
+
+    print("\nVALIDATION SET")
+    print(f"Number of songs \t\t\t\t{len(validset)}")
+    print(f"Number of unique tags\t\t\t{len(validset_tags_unique)}")
+    print(f"Avg number of tags per song\t\t{np.round(validset_tags.apply(len).mean(),2)}")
+    print(f"Max number of tags per song\t\t{validset_tags.apply(len).max()}")
+    print(f"Min number of tags per song\t\t{validset_tags.apply(len).min()}")
+
+    print("\nTEST SET")
+    print(f"Number of songs \t\t\t\t{len(testset)}")
+    print(f"Number of unique tags\t\t\t{len(testset_tags_unique)}")
+    print(f"Avg number of tags per song\t\t{np.round(testset_tags.apply(len).mean(),2)}")
+    print(f"Max number of tags per song\t\t{testset_tags.apply(len).max()}")
+    print(f"Min number of tags per song\t\t{testset_tags.apply(len).min()}")
+
+
+def correlations():
+    data_path = os.path.join(PATH_ANNOTATIONS, 'train_processed.tsv')
+    df = pd.read_csv(data_path, sep='\t')
+    tags = df.TAGS.str.split(pat=',')
+    artists = df.ARTIST_ID
+    u_tags = np.sort(np.unique(np.hstack(tags)))
+    u_artists, counts = np.unique(artists, return_counts=True)
+
+    counts_idx = counts.argsort()
+    relevant_artists = u_artists[counts_idx[::-1]][:sum((counts>10).astype(int))]
+
+    contingency = pd.DataFrame(0, index=relevant_artists, columns=u_tags)
+
+    for tag in tqdm(u_tags):
+        for i in range(len(tags)):
+            if tag in df.TAGS[i]:
+                if df.ARTIST_ID[i] in relevant_artists:
+                    contingency.loc[df.ARTIST_ID[i]][tag] += 1
+                else:
+                    continue
+
+    pass
+
+
+
+if __name__=='__main__':
+    # stats = compute_duration_stats(os.path.join(PATH_ANNOTATIONS, 'train_processed.tsv'))
+    # stats = compute_melspec_length_stats(os.path.join(PATH_ANNOTATIONS, 'train_processed.tsv'))
+    # print(stats)
+    # compute_tag_stats()
+    correlations()
\ No newline at end of file

main.py

-from utils import *
-from pytorch_lightning import Trainer
-from test_tube import Experiment
-from models.vgg_basic import MultiTagger
-from models import cp_resnet
-
-
-def run():
-    logger.info(CURR_RUN_PATH)
-    exp = Experiment(save_dir=CURR_RUN_PATH)
-
-    if USE_GPU:
-        trainer = Trainer(gpus=[0], distributed_backend='ddp',
-                          experiment=exp, max_nb_epochs=10, train_percent_check=1.0,
-                          fast_dev_run=False)
-    else:
-        trainer = Trainer(experiment=exp, max_nb_epochs=1, train_percent_check=0.1,
-                          fast_dev_run=True)
-
-    from strategies import model_config
-    model = cp_resnet.Network(model_config)
-
-    trainer.fit(model)
-
-
-if __name__=='__main__':
-    run()
\ No newline at end of file

models/baseline.py

 import torch.nn as nn
 
 from utils import *
-from datasets import MelSpecDataset
+from datasets.datasets import MelSpecDataset
 import torch
 import torch.nn as nn
 import torch.nn.functional as F

models/cp_resnet.py

 # coding: utf-8
 import math
 
-import torch
-import torch.nn as nn
 import torch.nn.functional as F
 
 from librosa.filters import mel as librosa_mel_fn
 
 from utils import *
-from datasets import MelSpecDataset
+from datasets.datasets import MelSpecDataset
 from torch.utils.data import DataLoader
 import pytorch_lightning as pl
 
@@ -390,18 +388,16 @@ class Network(pl.LightningModule):
         y_hat_probs = F.softmax(y_hat, dim=1)
         y_hat_binary = (y_hat_probs > 0.5).type(torch.int)
         rocauc = roc_auc_score(y.t().cpu(), y_hat_probs.t().cpu())
-        fscore = f1_score(y.t().cpu(), y_hat_probs.t().cpu(), average='micro')
+        # fscore = f1_score(y.t().cpu(), y_hat_probs.t().cpu(), average='micro')
         return {'val_loss': self.my_loss(y_hat, y),
-                'rocauc':rocauc,
-                'fscore': fscore}
+                'rocauc':rocauc}
 
     def validation_end(self, outputs):
         avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
         avg_auc = torch.stack([torch.tensor([x['rocauc']]) for x in outputs]).mean()
-        avg_f = torch.stack([torch.tensor([x['fscore']]) for x in outputs]).mean()
+        # avg_f = torch.stack([torch.tensor([x['fscore']]) for x in outputs]).mean()
         return {'val_loss':avg_loss,
-                'rocauc':avg_auc,
-                'fscore':avg_f}
+                'rocauc':avg_auc}
 
     def configure_optimizers(self):
         return [torch.optim.Adam(self.parameters(), lr=0.02)]
@@ -409,21 +405,26 @@ class Network(pl.LightningModule):
     @pl.data_loader
     def tng_dataloader(self):
         trainset = MelSpecDataset(phase='train', ann_root=PATH_ANNOTATIONS,
-                                  spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED)
+                                  spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED, framed=True)
         return DataLoader(dataset=trainset, batch_size=32, shuffle=True)
 
     @pl.data_loader
     def val_dataloader(self):
         validationset = MelSpecDataset(phase='validation', ann_root=PATH_ANNOTATIONS,
-                                       spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED)
+                                       spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED, framed=True)
         return DataLoader(dataset=validationset, batch_size=128, shuffle=True)
 
     @pl.data_loader
     def test_dataloader(self):
         testset = MelSpecDataset(phase='test', ann_root=PATH_ANNOTATIONS,
-                                 spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED)
+                                 spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED, framed=True)
         return DataLoader(dataset=testset, batch_size=32, shuffle=True)
 
+    @staticmethod
+    def add_model_specific_args(parent_parser, root_dir):
+        return parent_parser
+        pass
+
 
 if __name__=='__main__':
     model_config = {

models/legacy_models.py

-from utils import *
-from datasets import MelSpecDataset
-import torch
 import torch.nn as nn
-import torch.nn.functional as F
-from torch.utils.data import DataLoader
-import pytorch_lightning as pl
 
-from sklearn.metrics import roc_auc_score
 
 def initialize_weights(module):
     if isinstance(module, nn.Conv2d):

models/vgg_basic.py

 from utils import *
-from datasets import MelSpecDataset
+from datasets.datasets import MelSpecDataset
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -9,9 +9,9 @@ import pytorch_lightning as pl
 from sklearn.metrics import roc_auc_score
 
 
-class MultiTagger(pl.LightningModule):
+class Network(pl.LightningModule):
     def __init__(self, num_tags=8):
-        super(MultiTagger, self).__init__()
+        super(Network, self).__init__()
         self.num_tags = num_tags
 
         self.conv1 = nn.Sequential(
@@ -147,17 +147,22 @@ class MultiTagger(pl.LightningModule):
     @pl.data_loader
     def tng_dataloader(self):
         trainset = MelSpecDataset(phase='train', ann_root=PATH_ANNOTATIONS,
-                                  spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED)
+                                  spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED, framed=True)
         return DataLoader(dataset=trainset, batch_size=32, shuffle=True)
 
     @pl.data_loader
     def val_dataloader(self):
         validationset = MelSpecDataset(phase='validation', ann_root=PATH_ANNOTATIONS,
-                                       spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED)
+                                       spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED, framed=True)
         return DataLoader(dataset=validationset, batch_size=128, shuffle=True)
 
     @pl.data_loader
     def test_dataloader(self):
         testset = MelSpecDataset(phase='test', ann_root=PATH_ANNOTATIONS,
-                                 spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED)
+                                 spec_root=PATH_MELSPEC_DOWNLOADED_FRAMED, framed=True)
         return DataLoader(dataset=testset, batch_size=32, shuffle=True)
+
+    @staticmethod
+    def add_model_specific_args(parent_parser, root_dir):
+        return parent_parser
+        pass
\ No newline at end of file

processors/annotation_processors.py

+from utils import *
+
+def preprocess_and_save_annotation_files():
+    """
+    Removes 'mood/theme---' from tag names, and replaces tabs between multiple tag names with commas.
+    Writes processed filename.ext as filename_processed.ext
+    """
+    import re
+    filelist = os.listdir(PATH_ANNOTATIONS)
+    for file in filelist:
+        # Check if the current file is processed or has a processed copy.
+        # Skip if either of this is true. Else process.
+        if 'processed' in os.path.splitext(file)[0].split('_') or\
+            f'{os.path.splitext(file)[0]}_processed{os.path.splitext(file)[1]}' in filelist:
+            continue
+        else:
+            with open(os.path.join(PATH_ANNOTATIONS, file), 'r') as f:
+                text = f.read()
+            text = re.sub(r'mood/theme---(\w*)\n', r'\1\n', text) # matches last or singular tags
+            text = re.sub(r'mood/theme---(\w*)(\s*)', r'\1,', text) # matches all other tags
+
+            with open(os.path.join(PATH_ANNOTATIONS,
+                                   f'{os.path.splitext(file)[0]}_processed{os.path.splitext(file)[1]}'), 'w') as fw:
+                fw.write(text)

processors/griffin_lim.py

+
+import math
+import sys
+import time
+import numpy as np
+import wave
+import scipy
+import scipy.signal
+from pylab import *
+import array
+import os
+from os.path import expanduser
+import scipy.io.wavfile
+
+# Author: Brian K. Vogel
+# brian.vogel@gmail.com
+
+
+def hz_to_mel(f_hz):
+    """Convert Hz to mel scale.
+    This uses the formula from O'Shaugnessy's book.
+    Args:
+        f_hz (float): The value in Hz.
+    Returns:
+        The value in mels.
+    """
+    return 2595*np.log10(1.0 + f_hz/700.0)
+
+
+def mel_to_hz(m_mel):
+    """Convert mel scale to Hz.
+    This uses the formula from O'Shaugnessy's book.
+    Args:
+        m_mel (float): The value in mels
+    Returns:
+        The value in Hz
+    """
+    return 700*(10**(m_mel/2595) - 1.0)
+
+
+def fft_bin_to_hz(n_bin, sample_rate_hz, fft_size):
+    """Convert FFT bin index to frequency in Hz.
+    Args:
+        n_bin (int or float): The FFT bin index.
+        sample_rate_hz (int or float): The sample rate in Hz.
+        fft_size (int or float): The FFT size.
+    Returns:
+        The value in Hz.
+    """
+    n_bin = float(n_bin)
+    sample_rate_hz = float(sample_rate_hz)
+    fft_size = float(fft_size)
+    return n_bin*sample_rate_hz/(2.0*fft_size)
+
+
+def hz_to_fft_bin(f_hz, sample_rate_hz, fft_size):
+    """Convert frequency in Hz to FFT bin index.
+    Args:
+        f_hz (int or float): The frequency in Hz.
+        sample_rate_hz (int or float): The sample rate in Hz.
+        fft_size (int or float): The FFT size.
+    Returns:
+        The FFT bin index as an int.
+    """
+    f_hz = float(f_hz)
+    sample_rate_hz = float(sample_rate_hz)
+    fft_size = float(fft_size)
+    fft_bin = int(np.round((f_hz*2.0*fft_size/sample_rate_hz)))
+    if fft_bin >= fft_size:
+        fft_bin = fft_size-1
+    return fft_bin
+
+
+def make_mel_filterbank(min_freq_hz, max_freq_hz, mel_bin_count,
+                        linear_bin_count, sample_rate_hz):
+    """Create a mel filterbank matrix.
+    Create and return a mel filterbank matrix `filterbank` of shape (`mel_bin_count`,
+    `linear_bin_couont`). The `filterbank` matrix can be used to transform a
+    (linear scale) spectrum or spectrogram into a mel scale spectrum or
+    spectrogram as follows:
+    `mel_scale_spectrum` = `filterbank`*'linear_scale_spectrum'
+    where linear_scale_spectrum' is a shape (`linear_bin_count`, `m`) and
+    `mel_scale_spectrum` is shape ('mel_bin_count', `m`) where `m` is the number
+    of spectral time slices.
+    Likewise, the reverse-direction transform can be performed as:
+    'linear_scale_spectrum' = filterbank.T`*`mel_scale_spectrum`
+    Note that the process of converting to mel scale and then back to linear
+    scale is lossy.
+    This function computes the mel-spaced filters such that each filter is triangular
+    (in linear frequency) with response 1 at the center frequency and decreases linearly
+    to 0 upon reaching an adjacent filter's center frequency. Note that any two adjacent
+    filters will overlap having a response of 0.5 at the mean frequency of their
+    respective center frequencies.
+    Args:
+        min_freq_hz (float): The frequency in Hz corresponding to the lowest
+            mel scale bin.
+        max_freq_hz (flloat): The frequency in Hz corresponding to the highest
+            mel scale bin.
+        mel_bin_count (int): The number of mel scale bins.
+        linear_bin_count (int): The number of linear scale (fft) bins.
+        sample_rate_hz (float): The sample rate in Hz.
+    Returns:
+        The mel filterbank matrix as an 2-dim Numpy array.
+    """
+    min_mels = hz_to_mel(min_freq_hz)
+    max_mels = hz_to_mel(max_freq_hz)
+    # Create mel_bin_count linearly spaced values between these extreme mel values.
+    mel_lin_spaced = np.linspace(min_mels, max_mels, num=mel_bin_count)
+    # Map each of these mel values back into linear frequency (Hz).
+    center_frequencies_hz = np.array([mel_to_hz(n) for n in mel_lin_spaced])
+    mels_per_bin = float(max_mels - min_mels)/float(mel_bin_count - 1)
+    mels_start = min_mels - mels_per_bin
+    hz_start = mel_to_hz(mels_start)
+    fft_bin_start = hz_to_fft_bin(hz_start, sample_rate_hz, linear_bin_count)
+    #print('fft_bin_start: ', fft_bin_start)
+    mels_end = max_mels + mels_per_bin
+    hz_stop = mel_to_hz(mels_end)
+    fft_bin_stop = hz_to_fft_bin(hz_stop, sample_rate_hz, linear_bin_count)
+    #print('fft_bin_stop: ', fft_bin_stop)
+    # Map each center frequency to the closest fft bin index.
+    linear_bin_indices = np.array([hz_to_fft_bin(f_hz, sample_rate_hz, linear_bin_count) for f_hz in center_frequencies_hz])
+    # Create filterbank matrix.
+    filterbank = np.zeros((mel_bin_count, linear_bin_count))
+    for mel_bin in range(mel_bin_count):
+        center_freq_linear_bin = linear_bin_indices[mel_bin]
+        # Create a triangular filter having the current center freq.
+        # The filter will start with 0 response at left_bin (if it exists)
+        # and ramp up to 1.0 at center_freq_linear_bin, and then ramp
+        # back down to 0 response at right_bin (if it exists).
+
+        # Create the left side of the triangular filter that ramps up
+        # from 0 to a response of 1 at the center frequency.
+        if center_freq_linear_bin > 1:
+            # It is possible to create the left triangular filter.
+            if mel_bin == 0:
+                # Since this is the first center frequency, the left side