Download this notebook by clicking on the download icon or find it in the repository

Blue A Classify

Copyright (c) 2022, MBARI

• Distributed under the terms of the GPL License
• Maintainer: dcline@mbari.org
• Authors: Danelle Cline dcline@mbari.org, John Ryan ryjo@mbari.org

Kernel Selection¶

If running in SageMaker, the Python 3 (Data Science) kernel is sufficient. The Python 3 (TensorFlow 2.3 Python 3.7 GPU Optimized) will run the inference code faster for a higher cost. For more advanced users, SageMaker Batch Transform is recommended to process data in bulk.

Applying Machine Learning to classify blue whale A calls¶

---

Essential to detection and classification of marine mammal vocalizations are the distinct acoustic attributes of those vocalizations. Machine learning (ML) is an effective way to recognize acoustic attributes and reliably classify such vocalizations.

In this brief tutorial, we will:

• tap into an extensive (6+ years and growing) archive of sound recordings from a deep-sea location along the eastern margin of the North Pacific Ocean,
• illustrate the beautiful songs produced by baleen whales, and
• demonstrate the application of ML to classify one of the three types of calls produced by blue whales in their songs.

If you use this data set or tutorial, please cite our project[1].

Install required dependencies¶

First, let's install the required software dependencies.

If you are using this notebook in a cloud environment, select a Python3 compatible kernel and run this next section. This only needs to be done once for the duration of this notebook.

If you are working on local computer, you can skip this next cell. Change your kernel to pacific-sound-notebooks, which you installed according to the instructions in the README - this has all the dependencies that are needed.

In [15]:

!apt-get update -y && apt-get install -y libsndfile1
!python -m pip install --upgrade pip
!pip install tensorflow==2.4.1 --quiet
!pip install boto3 --quiet
!pip install oceansoundscape  --quiet

!apt-get update -y && apt-get install -y libsndfile1 !python -m pip install --upgrade pip !pip install tensorflow==2.4.1 --quiet !pip install boto3 --quiet !pip install oceansoundscape --quiet

0% [Working]
            
Hit:1 https://cloud.r-project.org/bin/linux/ubuntu bionic-cran40/ InRelease

0% [Connecting to archive.ubuntu.com] [Connecting to security.ubuntu.com (91.18
0% [1 InRelease gpgv 3,626 B] [Connecting to archive.ubuntu.com] [Waiting for h
                                                                               
Ign:2 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  InRelease

0% [1 InRelease gpgv 3,626 B] [Connecting to archive.ubuntu.com] [Waiting for h
                                                                               
Hit:3 https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64  InRelease

0% [1 InRelease gpgv 3,626 B] [Connecting to archive.ubuntu.com] [Waiting for h
                                                                               
Hit:4 https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64  Release
Hit:5 http://security.ubuntu.com/ubuntu bionic-security InRelease
Hit:6 http://ppa.launchpad.net/c2d4u.team/c2d4u4.0+/ubuntu bionic InRelease
Hit:7 http://archive.ubuntu.com/ubuntu bionic InRelease
Hit:9 http://ppa.launchpad.net/cran/libgit2/ubuntu bionic InRelease
Hit:10 http://archive.ubuntu.com/ubuntu bionic-updates InRelease
Hit:11 http://archive.ubuntu.com/ubuntu bionic-backports InRelease
Hit:12 http://ppa.launchpad.net/deadsnakes/ppa/ubuntu bionic InRelease
Hit:13 http://ppa.launchpad.net/graphics-drivers/ppa/ubuntu bionic InRelease
Reading package lists... Done
Reading package lists... Done
Building dependency tree       
Reading state information... Done
libsndfile1 is already the newest version (1.0.28-4ubuntu0.18.04.2).
The following package was automatically installed and is no longer required:
  libnvidia-common-460
Use 'apt autoremove' to remove it.
0 upgraded, 0 newly installed, 0 to remove and 41 not upgraded.
Looking in indexes: https://pypi.org/simple, https://us-python.pkg.dev/colab-wheels/public/simple/
Requirement already satisfied: pip in /usr/local/lib/python3.7/dist-packages (22.2.2)
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv

Import all packages¶

In [16]:

import boto3
from botocore import UNSIGNED
from botocore.client import Config
import cv2
from oceansoundscape.spectrogram.signal import psd_1sec
from oceansoundscape.raven import BLEDParser
from oceansoundscape.spectrogram import conf, denoise, utils
import os
from pathlib import Path
import soundfile as sf
import json
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from matplotlib.patches import Rectangle

import boto3 from botocore import UNSIGNED from botocore.client import Config import cv2 from oceansoundscape.spectrogram.signal import psd_1sec from oceansoundscape.raven import BLEDParser from oceansoundscape.spectrogram import conf, denoise, utils import os from pathlib import Path import soundfile as sf import json import numpy as np import matplotlib.pyplot as plt import tensorflow as tf from matplotlib.patches import Rectangle

Why use 2 kHz data?¶

Because we are studying the low-frequency calls of blue whales, we don't need the original recordings sampled at 256 kHz. Instead, we will use the 2 kHz decimated data in WAV format; these are stored in an s3 bucket named pacific-sound-2khz. For more information on the storage and organization of the 2kHz data, please see the 2kHz example.

List the contents of a monthly directory¶

Between August 2015 and July 2021 (a 6-year period), the highest levels of blue whale song activity off central California were detected during November 2017. Let's start by listing the daily 2 kHz files for that month.

In [17]:

s3 = boto3.client('s3',
    aws_access_key_id='',
    aws_secret_access_key='',
    config=Config(signature_version=UNSIGNED))

year = 2017
month = 11
bucket = 'pacific-sound-2khz'

for obj in s3.list_objects_v2(Bucket=bucket, Prefix=f'{year:04d}/{month:02d}')['Contents']:
    print(obj['Key'])

s3 = boto3.client('s3', aws_access_key_id='', aws_secret_access_key='', config=Config(signature_version=UNSIGNED)) year = 2017 month = 11 bucket = 'pacific-sound-2khz' for obj in s3.list_objects_v2(Bucket=bucket, Prefix=f'{year:04d}/{month:02d}')['Contents']: print(obj['Key'])

2017/11/MARS-20171101T000000Z-2kHz.wav
2017/11/MARS-20171102T000000Z-2kHz.wav
2017/11/MARS-20171103T000000Z-2kHz.wav
2017/11/MARS-20171104T000000Z-2kHz.wav
2017/11/MARS-20171105T000000Z-2kHz.wav
2017/11/MARS-20171106T000000Z-2kHz.wav
2017/11/MARS-20171107T000000Z-2kHz.wav
2017/11/MARS-20171108T000000Z-2kHz.wav
2017/11/MARS-20171109T000000Z-2kHz.wav
2017/11/MARS-20171110T000000Z-2kHz.wav
2017/11/MARS-20171111T000000Z-2kHz.wav
2017/11/MARS-20171112T000000Z-2kHz.wav
2017/11/MARS-20171113T000000Z-2kHz.wav
2017/11/MARS-20171114T000000Z-2kHz.wav
2017/11/MARS-20171115T000000Z-2kHz.wav
2017/11/MARS-20171116T000000Z-2kHz.wav
2017/11/MARS-20171117T000000Z-2kHz.wav
2017/11/MARS-20171118T000000Z-2kHz.wav
2017/11/MARS-20171119T000000Z-2kHz.wav
2017/11/MARS-20171120T000000Z-2kHz.wav
2017/11/MARS-20171121T000000Z-2kHz.wav
2017/11/MARS-20171122T000000Z-2kHz.wav
2017/11/MARS-20171123T000000Z-2kHz.wav
2017/11/MARS-20171124T000000Z-2kHz.wav
2017/11/MARS-20171125T000000Z-2kHz.wav
2017/11/MARS-20171126T000000Z-2kHz.wav
2017/11/MARS-20171127T000000Z-2kHz.wav
2017/11/MARS-20171128T000000Z-2kHz.wav
2017/11/MARS-20171129T000000Z-2kHz.wav
2017/11/MARS-20171130T000000Z-2kHz.wav
2017/11/copy.sh

A view of baleen whale song¶

Let's produce a spectrogram with sufficient resolution in time and frequency to see the blue whale song with enough resolution to visually identify. We'll limit the exercise to a single hour from a day with calls of variable received intensity (signal strength).

Download a single 2 kHz file¶

In [18]:

year = 2017
month = 11
wav_filename = 'MARS-20171101T000000Z-2kHz.wav'
bucket = 'pacific-sound-2khz'
key = f'{year:04d}/{month:02d}/{wav_filename}'

s3 = boto3.resource('s3',
    aws_access_key_id='',
    aws_secret_access_key='',
    config=Config(signature_version=UNSIGNED))
 
# only download if needed
if not Path(wav_filename).exists():

    # Alternatively, it can be downloaded directly in SageMaker with
    # !aws s3 cp s3://{bucket}/{key} . 

    print('Downloading') 
    s3.Bucket(bucket).download_file(key, wav_filename)
    print('Done') 

year = 2017 month = 11 wav_filename = 'MARS-20171101T000000Z-2kHz.wav' bucket = 'pacific-sound-2khz' key = f'{year:04d}/{month:02d}/{wav_filename}' s3 = boto3.resource('s3', aws_access_key_id='', aws_secret_access_key='', config=Config(signature_version=UNSIGNED)) # only download if needed if not Path(wav_filename).exists(): # Alternatively, it can be downloaded directly in SageMaker with # !aws s3 cp s3://{bucket}/{key} . print('Downloading') s3.Bucket(bucket).download_file(key, wav_filename) print('Done')

Subset to the 5th hour of the day¶

In [19]:

sample_rate = int(2e3)
start_hour = 5
start_frame = int(sample_rate * start_hour * 3600)
duration_frames =  int(sample_rate* 3600)

pacsound_file = sf.SoundFile(wav_filename)
pacsound_file.seek(start_frame)
x = pacsound_file.read(duration_frames, dtype='float32')

sample_rate = int(2e3) start_hour = 5 start_frame = int(sample_rate * start_hour * 3600) duration_frames = int(sample_rate* 3600) pacsound_file = sf.SoundFile(wav_filename) pacsound_file.seek(start_frame) x = pacsound_file.read(duration_frames, dtype='float32')

Plot the full 2 kHz spectrogram¶

Lots of biophony (sounds of ocean life) are represented in this spectrogram. Humpback whale songs are dominant above ~ 100 Hz, while blue and fin whale songs are dominant below ~ 100 Hz. The energy of blue whale A calls is largely between ~70 and 90 Hz (between the white dashed lines).

For more details about creating a calibrated spectrogram, see the 2 kHz tutorial.

In [20]:

sg, f = psd_1sec(x, sample_rate, 177.9) # create calibrated psd
plt.figure(dpi=300)
plt.axhline(73,linestyle='--', color='white')
plt.axhline(91,linestyle='--', color='white')
plt.imshow(sg,extent=[0, 3600, min(f), max(f)],aspect='auto',origin='lower',vmin=30,vmax=100)
plt.yscale('log')
plt.ylim(10,1000)
plt.colorbar()
plt.xlabel('November 01, 2017 Hour 5')
plt.ylabel('Frequency (Hz)')
plt.title('Calibrated spectrum levels')

sg, f = psd_1sec(x, sample_rate, 177.9) # create calibrated psd plt.figure(dpi=300) plt.axhline(73,linestyle='--', color='white') plt.axhline(91,linestyle='--', color='white') plt.imshow(sg,extent=[0, 3600, min(f), max(f)],aspect='auto',origin='lower',vmin=30,vmax=100) plt.yscale('log') plt.ylim(10,1000) plt.colorbar() plt.xlabel('November 01, 2017 Hour 5') plt.ylabel('Frequency (Hz)') plt.title('Calibrated spectrum levels')

Out[20]:

Text(0.5, 1.0, 'Calibrated spectrum levels')

Band Limited Energy Detections¶

In this example, we will focus on classification of calls. Here, detections for potential Blue whale A calls have been precomputed using Band-Limited-Energy-Detectors BLEDs in the Raven software. There are many way to do detection, including spectrogram correlation and neural network approaches such as single-shot detectors to name a few. The best detector depends on the acoustic application and its cost/performance tradeoff.

The approach taken here is to use the BLED detectors to identify not only strong and clear calls, but also many uncertain signals, so we can avoid missing calls.

An output from the BLED detectors might look something like this following simple table.

Note that this only represents a subset of the detections in the 5th hour for the purposes of this simple example.

In [21]:

bled_file = 'MARS-20171101T000000Z-2kHz.wav.Table01.txt'

bled_file = 'MARS-20171101T000000Z-2kHz.wav.Table01.txt'

In [22]:

%%writefile {bled_file}
Selection	View	Channel	Low Freq (Hz)	High Freq (Hz)	Begin Time (s)	End Time (s)
453	Spectrogram 1	1	70.0	90.0	19076.535750	19088.027750
454	Spectrogram 1	1	70.0	90.0	19115.990750	19132.331750
455	Spectrogram 1	1	70.0	90.0	19145.565750	19161.217750
456	Spectrogram 1	1	70.0	90.0	19182.693750	19200.880750
457	Spectrogram 1	1	70.0	90.0	19221.550750	19235.590750
458	Spectrogram 1	1	70.0	90.0	19244.183750	19270.027750
459	Spectrogram 1	1	70.0	90.0	19310.171750	19328.345750
460	Spectrogram 1	1	70.0	90.0	19342.502750	19360.689750
461	Spectrogram 1	1	70.0	90.0	19373.676750	19389.107750
462	Spectrogram 1	1	70.0	90.0	19398.727750	19410.193750
463	Spectrogram 1	1	70.0	90.0	19421.425750	19436.804750
464	Spectrogram 1	1	70.0	90.0	19445.189750	19461.803750
465	Spectrogram 1	1	70.0	90.0	19494.706750	19507.173750
466	Spectrogram 1	1	70.0	90.0	19527.622750	19538.555750
467	Spectrogram 1	1	70.0	90.0	19548.071750	19561.370750
468	Spectrogram 1	1	70.0	90.0	19568.377750	19585.186750
469	Spectrogram 1	1	70.0	90.0	19625.057750	19645.753750
470	Spectrogram 1	1	70.0	90.0	19653.293750	19676.381750
471	Spectrogram 1	1	70.0	90.0	19680.944750	19704.734750
472	Spectrogram 1	1	70.0	90.0	19708.543750	19723.051750
473	Spectrogram 1	1	70.0	90.0	19760.907750	19778.990750

%%writefile {bled_file} Selection View Channel Low Freq (Hz) High Freq (Hz) Begin Time (s) End Time (s) 453 Spectrogram 1 1 70.0 90.0 19076.535750 19088.027750 454 Spectrogram 1 1 70.0 90.0 19115.990750 19132.331750 455 Spectrogram 1 1 70.0 90.0 19145.565750 19161.217750 456 Spectrogram 1 1 70.0 90.0 19182.693750 19200.880750 457 Spectrogram 1 1 70.0 90.0 19221.550750 19235.590750 458 Spectrogram 1 1 70.0 90.0 19244.183750 19270.027750 459 Spectrogram 1 1 70.0 90.0 19310.171750 19328.345750 460 Spectrogram 1 1 70.0 90.0 19342.502750 19360.689750 461 Spectrogram 1 1 70.0 90.0 19373.676750 19389.107750 462 Spectrogram 1 1 70.0 90.0 19398.727750 19410.193750 463 Spectrogram 1 1 70.0 90.0 19421.425750 19436.804750 464 Spectrogram 1 1 70.0 90.0 19445.189750 19461.803750 465 Spectrogram 1 1 70.0 90.0 19494.706750 19507.173750 466 Spectrogram 1 1 70.0 90.0 19527.622750 19538.555750 467 Spectrogram 1 1 70.0 90.0 19548.071750 19561.370750 468 Spectrogram 1 1 70.0 90.0 19568.377750 19585.186750 469 Spectrogram 1 1 70.0 90.0 19625.057750 19645.753750 470 Spectrogram 1 1 70.0 90.0 19653.293750 19676.381750 471 Spectrogram 1 1 70.0 90.0 19680.944750 19704.734750 472 Spectrogram 1 1 70.0 90.0 19708.543750 19723.051750 473 Spectrogram 1 1 70.0 90.0 19760.907750 19778.990750

Writing MARS-20171101T000000Z-2kHz.wav.Table01.txt

Focus on blue whale A calls¶

Zooming in on a more narrow frequency and time, we can see the pulse-train nature of the A calls, an important distinction for classification. In this brief segment, we can see A calls of variable received intensity.

Here, we parse the detections and display the identifying Raven Selection number for reference.

Notes:

• The first A call in this time period, which was quite strong, does not have a start marker because its start precedes the beginning of this audio segment.
• The full detail of the A calls is still not visible at the resolution presented above. The full detail will become clear when we apply the model and examine individual classified calls.

In [23]:

# an optimum configuration for the call spectrogram generation is defined in the oceanscoundscape package based on extensive
# hyper parameter sweeps. Let's grab those here
blue_a_conf = conf.CONF_DICT['blueA']

# BLEDParser needs a path object
bled_path = Path(bled_file)

# parse the detections
parser = BLEDParser(bled_path, blue_a_conf, max_samples=len(x), sampling_rate=sample_rate)

# plot the spectrogram
plt.figure(dpi=300)
plt.imshow(sg,aspect='auto',origin='lower',vmin=30,vmax=100)

# add boxes for each normalized detection - detections are normalized to a fixed size (frequency/time) for classification
start_secs = start_hour * 3600
high_freq = blue_a_conf['high_freq']
low_freq = blue_a_conf['low_freq']
duration_secs = blue_a_conf['duration_secs']
freq_window = int(high_freq - low_freq)
for row, item in sorted(parser.data.iterrows()):
    detection_start = item.call_start/sample_rate - start_secs # detections are stored in sample numbers so here we convert back to time
    plt.gca().add_patch(Rectangle((detection_start,low_freq),duration_secs,freq_window,
                        edgecolor='black',
                        facecolor='none',
                        ls='--',
                        lw=1))
    plt.text(detection_start,high_freq - 8,item.Selection, size='x-small', rotation=45)

plt.yscale('linear')
plt.ylim(10,high_freq + 10)
plt.xlim(1000, 1800)
plt.xlabel('Second of 01 November 2017')
plt.ylabel('Frequency (Hz)')
plt.title('Detected blue whale A calls')

# an optimum configuration for the call spectrogram generation is defined in the oceanscoundscape package based on extensive # hyper parameter sweeps. Let's grab those here blue_a_conf = conf.CONF_DICT['blueA'] # BLEDParser needs a path object bled_path = Path(bled_file) # parse the detections parser = BLEDParser(bled_path, blue_a_conf, max_samples=len(x), sampling_rate=sample_rate) # plot the spectrogram plt.figure(dpi=300) plt.imshow(sg,aspect='auto',origin='lower',vmin=30,vmax=100) # add boxes for each normalized detection - detections are normalized to a fixed size (frequency/time) for classification start_secs = start_hour * 3600 high_freq = blue_a_conf['high_freq'] low_freq = blue_a_conf['low_freq'] duration_secs = blue_a_conf['duration_secs'] freq_window = int(high_freq - low_freq) for row, item in sorted(parser.data.iterrows()): detection_start = item.call_start/sample_rate - start_secs # detections are stored in sample numbers so here we convert back to time plt.gca().add_patch(Rectangle((detection_start,low_freq),duration_secs,freq_window, edgecolor='black', facecolor='none', ls='--', lw=1)) plt.text(detection_start,high_freq - 8,item.Selection, size='x-small', rotation=45) plt.yscale('linear') plt.ylim(10,high_freq + 10) plt.xlim(1000, 1800) plt.xlabel('Second of 01 November 2017') plt.ylabel('Frequency (Hz)') plt.title('Detected blue whale A calls')

Reading MARS-20171101T000000Z-2kHz.wav.Table01.txt {'low_freq': 70, 'high_freq': 100, 'duration_secs': 25, 'blur_axis': 'frequency', 'num_fft': 1024, 'center': True, 'padding_secs': 3, 'num_mels': 30}
Found 21 detections in MARS-20171101T000000Z-2kHz.wav.Table01.txt

Out[23]:

Text(0.5, 1.0, 'Detected blue whale A calls')

Classification of blue whale A calls¶

We can turn these detected audio signals into images, one for each detected signal, and classify them using a model trained for only A calls.

Create optimized spectrograms¶

Because the model we are demonstrating is image based, the first step to classify the calls is to create optimized spectrograms for each candidate detection.

In [24]:

# read the wav file
xc, sample_rate = sf.read(wav_filename)
 
# parse the detections, but this time let's use the entire wav file to allow for padding during denoising
parser_full = BLEDParser(bled_path, blue_a_conf, len(xc), sample_rate)

# grab the spectrogram parameters
call_width = int(blue_a_conf['duration_secs'] * sample_rate)
num_fft = blue_a_conf['num_fft']
hop_length = round(num_fft * (1 - conf.OVERLAP))

# generate optimized spectrograms
print(f'Denoising {bled_file} num detections {parser_full.num_detections}')

# denoise them and store the results in a temporary hdf file cache
cache = denoise.PCENSmooth(Path(f'{bled_path}.hdf'), parser_full, blue_a_conf, xc, sample_rate)

for row, item in sorted(parser_full.data.iterrows()):
    try:
        start = int(call_width / hop_length)
        end = int(2 * call_width / hop_length)

        data = cache.get_data(item.Selection)

        # subset the call, leaving off the padding for PCEN
        subset_data = data[:, start:end]

        # save to a denoised color image
        utils.ImageUtils.colorizeDenoise(subset_data, Path(item.image_filename))
    except Exception as ex:
        print(ex)
        continue

print('Done')

# read the wav file xc, sample_rate = sf.read(wav_filename) # parse the detections, but this time let's use the entire wav file to allow for padding during denoising parser_full = BLEDParser(bled_path, blue_a_conf, len(xc), sample_rate) # grab the spectrogram parameters call_width = int(blue_a_conf['duration_secs'] * sample_rate) num_fft = blue_a_conf['num_fft'] hop_length = round(num_fft * (1 - conf.OVERLAP)) # generate optimized spectrograms print(f'Denoising {bled_file} num detections {parser_full.num_detections}') # denoise them and store the results in a temporary hdf file cache cache = denoise.PCENSmooth(Path(f'{bled_path}.hdf'), parser_full, blue_a_conf, xc, sample_rate) for row, item in sorted(parser_full.data.iterrows()): try: start = int(call_width / hop_length) end = int(2 * call_width / hop_length) data = cache.get_data(item.Selection) # subset the call, leaving off the padding for PCEN subset_data = data[:, start:end] # save to a denoised color image utils.ImageUtils.colorizeDenoise(subset_data, Path(item.image_filename)) except Exception as ex: print(ex) continue print('Done')

Reading MARS-20171101T000000Z-2kHz.wav.Table01.txt {'low_freq': 70, 'high_freq': 100, 'duration_secs': 25, 'blur_axis': 'frequency', 'num_fft': 1024, 'center': True, 'padding_secs': 3, 'num_mels': 30}
Found 21 detections in MARS-20171101T000000Z-2kHz.wav.Table01.txt
Denoising MARS-20171101T000000Z-2kHz.wav.Table01.txt num detections 21
Done

Display the spectrogram images¶

In [25]:

images = list( sorted((Path.cwd()).glob('*.jpg')))

plt.figure(figsize=(14,14))
plt.rcParams['axes.titlelocation'] = 'center'
for i, image_path in enumerate(images):
    # the selection number of part of the filename name,
    # e.g. selection 3 20150925T072604.53528828.53552892.sel.03.ch01.spectrogram.jpg
    selection = image_path.name.split('.')[4]
    image_path = str(image_path)
    plt.subplot(5, 5, i + 1)
    img = plt.imread(image_path)
    plt.title(selection)
    plt.imshow(img)
    plt.axis('off')

images = list( sorted((Path.cwd()).glob('*.jpg'))) plt.figure(figsize=(14,14)) plt.rcParams['axes.titlelocation'] = 'center' for i, image_path in enumerate(images): # the selection number of part of the filename name, # e.g. selection 3 20150925T072604.53528828.53552892.sel.03.ch01.spectrogram.jpg selection = image_path.name.split('.')[4] image_path = str(image_path) plt.subplot(5, 5, i + 1) img = plt.imread(image_path) plt.title(selection) plt.imshow(img) plt.axis('off')