diff --git a/custom_hyp.yaml b/custom_hyp.yaml
index cebb06f322e9174a14b5fb85c093bcb7d50af333..3e78301a5764d142b4ed182b39e5e1a26396993e 100644
--- a/custom_hyp.yaml
+++ b/custom_hyp.yaml
@@ -29,3 +29,5 @@ fliplr: 0.0 # image flip left-right (probability)
mosaic: 0.0 # image mosaic (probability)
mixup: 0.3 # image mixup (probability)
copy_paste: 0.1 # segment copy-paste (probability)
+ToGray: 0.0 # image in Grayscale (probability)
+GaussNoise: 0.0 # adding Gaussian/white noise (probability)
diff --git a/get_spectrogram.py b/get_spectrogram.py
index 556ce00cfd53e70dae72f427370d93caea20761c..767ebe72b5a8b2bd1d713b29741526926a117c0d 100755
--- a/get_spectrogram.py
+++ b/get_spectrogram.py
@@ -51,12 +51,14 @@ def main(data, arguments):
name = os.path.join(arguments.directory, folder, f'{file}_{offset}')
utils.create_spectrogram(
- sig, arguments.directory, name, window_size=arguments.window,
- overlap=arguments.hop)
- return folder
+ sig, arguments.directory, name, window_size=arguments.window,
+ overlap=arguments.hop)
+
except Exception as error:
+ folder = 'spectrograms'
print(f'`{filename}` cannot be open : {error}')
+ return folder
if __name__ == "__main__":
parser = argparse.ArgumentParser(
@@ -67,8 +69,8 @@ if __name__ == "__main__":
help='Directory to which spectrograms will be stored')
parser.add_argument('--duration', type=int,
help='Duration for each spectrogram', default=8)
- parser.add_argument(
- '--window', type=int, help='Window size for the Fourier Transform', default=1024)
+ parser.add_argument('--window', type=int, help='Window size for the Fourier Transform',
+ default=1024)
parser.add_argument('--hop', type=float,
help='Ratio of hop in window : 50%% = 0.5', default=0.5)
parser.add_argument('--high', type=int,
diff --git a/get_train_annot.py b/get_train_annot.py
index fa7f0834b15d1f8146dcd90eaad087f884e0430d..9c1749637ceed310001cec5bcfadff538440dd8d 100755
--- a/get_train_annot.py
+++ b/get_train_annot.py
@@ -3,18 +3,15 @@
import argparse
import random
import os
-import pandas as pd
+import sys
import numpy as np
-import matplotlib.pyplot as plt
-import soundfile as sf
-
import cv2
-import sys
-import subprocess
import utils
-
+import matplotlib.pyplot as plt
+import soundfile as sf
from p_tqdm import p_map
from tqdm import tqdm
+import pandas as pd
def process(entry, arguments, species_list):
@@ -63,13 +60,13 @@ def process(entry, arguments, species_list):
sig, fs = sf.read(filename, start=int(
offset*fs), stop=int((offset+arguments.duration)*fs), always_2d=True) # Load the signal
sig = sig[:, 0] # Only take channel 0
- if arguments.rf == None:
+ if not arguments.rf:
arguments.rf = fs
# Apply resample and low/high pass filter
sig = utils.signal_processing(
sig, rf=arguments.rf, fs=fs, high=arguments.high, low=arguments.low)
fig = utils.create_spectrogram(
- sig, arguments.directory, names=None,
+ sig, arguments.directory, names=None,
window_size=arguments.window,
overlap=arguments.hop)
@@ -105,13 +102,14 @@ def process(entry, arguments, species_list):
for folder in ['images', 'labels', 'images/all', 'annotated_images']:
utils.create_directory(os.path.join(
arguments.directory, folder))
- for specie_num in species_list:
+ for specie_num in species_list[species_list.columns[0]]:
utils.create_directory(os.path.join(
arguments.directory, 'images', str(specie_num)))
# Save the images and annotation
plt.savefig(os.path.join(arguments.directory,
- 'images', str(species_list[species_list.species ==
- specie].index[0]), f'{name}.jpg'))
+ 'images',
+ str(species_list[species_list.species ==specie].species.iloc[0]),
+ f'{name}.jpg'))
annotation.to_csv(name_file, sep=' ', header=False, index=False)
plt.savefig(os.path.join(arguments.directory, 'images', 'all',
f'{name}.jpg'))
@@ -167,13 +165,15 @@ if __name__ == '__main__':
parser.add_argument('--overlap', type=int,
help='Overlap in seconds between 2 spectrograms', default=0)
parser.add_argument(
- '--rf', type=int, help='Frequency Resampling ', default=None)
+ '--rf', type=int, help='Frequency Resampling', default=None)
parser.add_argument(
- '--window', type=int, help='Window size for the Fourier Transform', default=1024)
+ '--window', type=int, help='Window size for the Fourier Transform',
+ default=1024)
parser.add_argument(
'--hop', type=float, help='Ratio of hop in window : 50%% = 0.5', default=.5)
parser.add_argument(
- '--cpu', type=int, help='To speed up the process, write 2 or more', default=1)
+ '--cpu', type=int, help='To speed up the process, write 2 or more',
+ default=1)
parser.add_argument('--high', type=int,
help='High Pass Filter value in Hz', default=10)
parser.add_argument('--low', type=int,
@@ -185,13 +185,14 @@ if __name__ == '__main__':
# Load the data and put it into a DataFrame
df = utils.open_file(args.filename_path)
- suffix = input('Which suffix for your recording data? [wav, WAV, Wav, flac, mp3] : ')
+ suffix = input('Which suffix for your recording data? [wav, WAV, Wav, flac, mp3, <other>] : ')
+
if len(df.columns) == 1:
final = []
for file, _ in df.groupby('Path'):
new_df = utils.open_file(file)
if len(new_df) >= 1:
- new_df['Path'] = os.path.join(args.path_to_data, str(file.split('.Table')[0]+f'.{suffix}'))
+ new_df['Path'] = os.path.join(args.path_to_data, str(file.split('/')[-1].split('.Table')[0]+f'.{suffix}'))
final.append(new_df)
else:
continue
@@ -205,7 +206,6 @@ if __name__ == '__main__':
species.to_csv(os.path.join(
args.directory, 'species_list.csv'), index=False)
-
if args.cpu == 1:
for i in tqdm(enumerate(df.groupby('Path')), total=len(df.groupby('Path')),
desc="Processing", ascii='░▒▓█'):
@@ -234,23 +234,10 @@ if __name__ == '__main__':
directory_path = os.path.join(path, args.directory, 'set')
# Create the directory path if not exists
utils.create_directory(directory_path)
- try :
+ try :
+ print(f'Train saved in {directory_path}\n')
# Run the split command
os.system(f'{sys.executable} {script} {data_path} {directory_path} -r 0.7 --test')
-
- print(f'Train saved in {directory_path}\n')
- print('To train your model, use the following command : \n')
-
- yolo_path = os.path.join(path, 'yolov5/train.py')
- data_path = os.path.join(directory_path, 'custom_data.yaml')
- weights_path = os.path.join(path, 'yolov5/weights/yolov5l.pt')
- hyp_path = os.path.join(path, 'custom_hyp.yaml')
-
- command = f'python {yolo_path} --data {data_path} --imgsz 640 --epochs 100 --weights {weights_path} --hyp {hyp_path} --cache'
- print(command,'\n')
- print('\u26A0\uFE0F Be aware that it is recommended to have background images that',
- ' represents 10% of your dataset. To do so, please use the script "get_spectrogram.py"',
- ' with --background arguments. Comptue on recordings that contains multiple type of noise...')
except Exception as error:
print(error)
diff --git a/get_train_val.py b/get_train_val.py
index de6d9a637b3ef979f2e3f47d5b4292388573fd49..0b8f59407190b076b699f59b496f3bcf3e80e1e0 100755
--- a/get_train_val.py
+++ b/get_train_val.py
@@ -123,7 +123,7 @@ if __name__ == '__main__':
command = f'python {yolo_path} --data {data_path} --imgsz 640 --epochs 100 --weights {weights_path} --hyp {hyp_path} --cache'
print(command,'\n')
- print('\u26A0\uFE0F Be aware that it is recommended to have background images that',
+ print('\u26A0\uFE0F Be aware that it is recommended to have background images that',
'represents 10% of your dataset. To do so, please use the script "get_spectrogram.py"',
'with --background arguments. Comptue on recordings that contains multiple type of noise...')
\ No newline at end of file
diff --git a/utils.py b/utils.py
index 947e5bb1dd4b7f6b946cc63620bc6c092ab7c9d9..4918af208ea377a9c00ef0c760fd903e06aae74f 100755
--- a/utils.py
+++ b/utils.py
@@ -66,7 +66,7 @@ def signal_processing(sig, rf, fs, high=None, low=None):
:return array: Processed signal.
"""
# Check if resampling frequency is different than sampling frequency
- if rf == None:
+ if not rf:
rf = fs
if rf != fs:
sig = signal.resample(sig, int(len(sig) * rf / fs)
@@ -97,7 +97,11 @@ def create_spectrogram(sig, directory, names, window_size=1024, overlap=.5):
:param directory (str): Path to save the spectrogram.
:param filename (str): Name of the final spectrogram.
"""
- overlap_size = int(window_size * overlap)
+ if overlap >= 1:
+ print(f'You put a hop value over 1. This has been corrected to have {overlap} as overlap size between window')
+ overlap_size = overlap
+ else:
+ overlap_size = int(window_size * overlap)
stft = librosa.stft(sig, n_fft=window_size,
hop_length=overlap_size, window='hann') # Compute the STFT
diff --git a/yolov5/data/hyps/hyp.Objects365.yaml b/yolov5/data/hyps/hyp.Objects365.yaml
index 74971740f7c73bf661950f339792b790a26b2b1c..f5a1ce15e30601614a23c367e378d089d1b21d7e 100755
--- a/yolov5/data/hyps/hyp.Objects365.yaml
+++ b/yolov5/data/hyps/hyp.Objects365.yaml
@@ -32,3 +32,5 @@ fliplr: 0.5
mosaic: 1.0
mixup: 0.0
copy_paste: 0.0
+ToGray: 0.0 # image in Grayscale (probability)
+GaussNoise: 0.0 # adding Gaussian/white noise (probability)
diff --git a/yolov5/data/hyps/hyp.VOC.yaml b/yolov5/data/hyps/hyp.VOC.yaml
index 0aa4e7d9f8f5162653e3999b04b4636b103c355f..f0a1fef4e986f93ae4d06a1237a4bd5bff07edfa 100755
--- a/yolov5/data/hyps/hyp.VOC.yaml
+++ b/yolov5/data/hyps/hyp.VOC.yaml
@@ -38,3 +38,5 @@ mosaic: 0.85834
mixup: 0.04266
copy_paste: 0.0
anchors: 3.412
+ToGray: 0.0
+GaussNoise: 0.0
diff --git a/yolov5/data/hyps/hyp.scratch-high.yaml b/yolov5/data/hyps/hyp.scratch-high.yaml
index 123cc8407413e9c130e21a3b5dd8ed33a3632db5..dcde050506de265c2f6e147bd6fa485ac63ecc68 100755
--- a/yolov5/data/hyps/hyp.scratch-high.yaml
+++ b/yolov5/data/hyps/hyp.scratch-high.yaml
@@ -32,3 +32,5 @@ fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.1 # image mixup (probability)
copy_paste: 0.1 # segment copy-paste (probability)
+ToGray: 0.0 # image in Grayscale (probability)
+GaussNoise: 0.0 # adding Gaussian/white noise (probability)
diff --git a/yolov5/data/hyps/hyp.scratch-low.yaml b/yolov5/data/hyps/hyp.scratch-low.yaml
index b9ef1d55a3b6ec8873ac87d6f4aa0ca081868bd6..1c65347a5f6d0dc908df5096ff697fe5cfe7e15b 100755
--- a/yolov5/data/hyps/hyp.scratch-low.yaml
+++ b/yolov5/data/hyps/hyp.scratch-low.yaml
@@ -32,3 +32,5 @@ fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.0 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)
+ToGray: 0.0 # image in Grayscale (probability)
+GaussNoise: 0.0 # adding Gaussian/white noise (probability)
diff --git a/yolov5/data/hyps/hyp.scratch-med.yaml b/yolov5/data/hyps/hyp.scratch-med.yaml
index d6867d7557bac73db7f8787db60cff4c4c64b440..75752655527c00629ac0a328c6e3170331b5fe78 100755
--- a/yolov5/data/hyps/hyp.scratch-med.yaml
+++ b/yolov5/data/hyps/hyp.scratch-med.yaml
@@ -32,3 +32,5 @@ fliplr: 0.5 # image flip left-right (probability)
mosaic: 1.0 # image mosaic (probability)
mixup: 0.1 # image mixup (probability)
copy_paste: 0.0 # segment copy-paste (probability)
+ToGray: 0.0 # image in Grayscale (probability)
+GaussNoise: 0.0 # adding Gaussian/white noise (probability)
diff --git a/yolov5/detect.py b/yolov5/detect.py
index fddf942e43c78cf37f1a456124c8475f515be6f0..47d0fb4669e92e2d5dfc6706463a905f5710d4a0 100755
--- a/yolov5/detect.py
+++ b/yolov5/detect.py
@@ -74,6 +74,8 @@ def run(
rf=22050,
window=1024,
hop=0.5,
+ low=None,
+ high=None,
visualize=False, # visualize features
update=False, # update all models
project=ROOT / 'runs/detect', # save results to project/name
@@ -97,10 +99,15 @@ def run(
project_name = input('Please enter the name of your project : ')
date_now = date.today().strftime("%Y%m%d")
+
if sound:
- name = '_'.join([project_name, 'detect', date_now, weights[0].split('/')[-3], 'conf', str(conf_thres).replace('.','_'), str(rf), str(sampleDur), str(window), str(int(window*hop)),''])
+ name = '_'.join([project_name, 'detect', date_now, weights[0].split('/')[-3], 'conf',
+ str(conf_thres).replace('.','_'), str(rf), str(sampleDur), str(window),
+ str(int(window*hop)),'low',str(int(low)), 'high', str(int(high)),''])
else:
- name = '_'.join([project_name, 'detect', date_now, weights[0].split('/')[-3], 'conf', str(conf_thres).replace('.','_'),''])
+ name = '_'.join([project_name, 'detect', date_now, weights[0].split('/')[-3], 'conf',
+ str(conf_thres).replace('.','_'),''])
+
print(f'\nYour train results will be saved as {name}\n')
# Directories
@@ -122,8 +129,12 @@ def run(
elif screenshot:
dataset = LoadScreenshots(source, img_size=imgsz, stride=stride, auto=pt)
elif sound:
- hop = window * hop
- dataset = LoadSpectros(source, sampleDur, rf, window, hop, img_size=imgsz, stride=stride, auto=pt)
+ if hop > 1:
+ hop = hop
+ print(f'You put hop > 1, this has been corrected by putting hop at {hop}')
+ else:
+ hop = window * hop
+ dataset = LoadSpectros(source, sampleDur, rf, window, hop, low, high, img_size=imgsz, stride=stride, auto=pt)
else:
dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt, vid_stride=vid_stride)
vid_path, vid_writer = [None] * bs, [None] * bs
@@ -260,6 +271,8 @@ def parse_opt():
parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels')
parser.add_argument('--sampleDur', default=8, help="Duration for each spectrogram for detection",type=int)
parser.add_argument('--rf', default=22050, help="Resampling Frequency",type=int)
+ parser.add_argument('--low', default=None, help="Low pass filter",type=int)
+ parser.add_argument('--high', default=None, help="High pass filter",type=int)
parser.add_argument('--window', default=1024, help="Window size for each spectrogram for detection",type=int)
parser.add_argument('--hop', default=0.5, help="Hop lenght for each spectrogram for detection",type=float)
parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences')
diff --git a/yolov5/train.py b/yolov5/train.py
index ce470f67a98bed3cc0103916bf9f001d508cf52c..ee3be00d73ba66fe2b64842ff87633f6c66a091f 100755
--- a/yolov5/train.py
+++ b/yolov5/train.py
@@ -82,6 +82,7 @@ def train(hyp, opt, device, callbacks): # hyp is path/to/hyp.yaml or hyp dictio
if isinstance(hyp, str):
with open(hyp, errors='ignore') as f:
hyp = yaml.safe_load(f) # load hyps dict
+
LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items()))
opt.hyp = hyp.copy() # for saving hyps to checkpoints
diff --git a/yolov5/utils/dataloaders.py b/yolov5/utils/dataloaders.py
index e0b047d5ecfe6a19407a5c721a0bd38703b68140..dad59ef8068f583e7481cb744fc4e2f64f8c5598 100755
--- a/yolov5/utils/dataloaders.py
+++ b/yolov5/utils/dataloaders.py
@@ -16,8 +16,9 @@ from itertools import repeat
from multiprocessing.pool import Pool, ThreadPool
from pathlib import Path
from threading import Thread
-import soundfile as sf
from urllib.parse import urlparse
+import librosa
+import soundfile as sf
import matplotlib.pyplot as plt
from scipy import signal
import numpy as np
@@ -238,8 +239,8 @@ class LoadScreenshots:
return str(self.screen), im, im0, None, s # screen, img, original img, im0s, s
class LoadSpectros:
- def __init__(self, folder, sampleDur, rf, window, hop, img_size, stride=32, auto=True):
- self.folder, self.sampleDur, self.rf, self.window, self.hop, self.img_size, self.stride, self.auto = folder, sampleDur, rf, window, hop, img_size, stride, auto
+ def __init__(self, folder, sampleDur, rf, window, hop, low, high, img_size, stride=32, auto=True):
+ self.folder, self.sampleDur, self.rf, self.window, self.hop, self.low, self.high, self.img_size, self.stride, self.auto = folder, sampleDur, rf, window, hop, low, high, img_size, stride, auto
self.files = os.listdir(folder)
self.mode = 'image'
self.samples = []
@@ -271,10 +272,26 @@ class LoadSpectros:
if fs != self.rf:
sig = signal.resample(sig, int(len(sig)*self.rf/fs))
fig = plt.figure()
+
+ if self.low:
+ # Create low pass filter
+ low_pass = signal.butter(1, self.low / (self.rf / 2), 'lp', output='sos')
+ sig = signal.sosfilt(low_pass, sig) # Apply low pass filter
+ if self.high:
+ # Create low pass filter
+ high_pass = signal.butter(2, self.high / (self.rf / 2), 'hp', output='sos')
+ sig = signal.sosfilt(high_pass, sig) # Apply high pass filter
+
+ if self.hop > 1:
+ hop = int(self.hop)
+ else:
+ hop = int(self.window * self.hop)
+
stft = librosa.stft(sig, n_fft=self.window,
- hop_length=self.hop, window='hann') # Compute the STFT
+ hop_length=hop, window='hann') # Compute the STFT
stft = np.log10(np.abs(stft))
- axim = plt.imshow(stft, aspect = "auto", interpolation = None, cmap = 'viridis', vmin=np.mean(stft))
+ axim = plt.imshow(stft, aspect = "auto", interpolation = None,
+ cmap = 'viridis', vmin=np.mean(stft))
plt.subplots_adjust(top=1, bottom=0, left=0, right=1)
im0 = axim.make_image(fig.canvas)[0][:,:,:-1][:,:,::-1]
cv2.imwrite(path, im0)
@@ -510,7 +527,8 @@ class LoadImagesAndLabels(Dataset):
self.mosaic_border = [-img_size // 2, -img_size // 2]
self.stride = stride
self.path = path
- self.albumentations = Albumentations(size=img_size) if augment else None
+ self.albumentations = Albumentations(pToGray=self.hyp['ToGray'],
+ pGaussNoise=self.hyp['GaussNoise'], size=img_size) if augment else None
try:
f = [] # image files