def __init__(self, weight_file, params_file):
"""Performs detection on an audio file.
The structure of the network is hard coded to a network with 2
convolution layers with pooling, 1 or 2 fully connected layers, and a
final softmax layer.
weight_file is the path to the numpy weights to the network
params_file is the path to the network parameters
"""
self.weights = np.load(weight_file, encoding='bytes', allow_pickle=True)
if not all([weight.dtype == np.float32 for weight in self.weights]):
for i in range(self.weights.shape[0]):
self.weights[i] = self.weights[i].astype(np.float32)
with open(params_file, 'rb') as fp:
params = pickle.load(fp)
self.chunk_size = 4.0 # seconds
self.win_size = params['win_size']
self.max_freq = params['max_freq']
self.min_freq = params['min_freq']
self.slice_scale = params['slice_scale']
self.overlap = params['overlap']
self.crop_spec = params['crop_spec']
self.denoise = params['denoise']
self.smooth_spec = params['smooth_spec']
self.nms_win_size = int(params['nms_win_size'])
self.smooth_op_prediction_sigma = params['smooth_op_prediction_sigma']
self.sp = Spectrogram()
def loadFile(path, sr=16000, time=3.0):
spectrogram = Spectrogram(input_shape=(int(time * sr), 1), sr=sr)
X = []
curr = 0 # indicates 1st and last file of speaker
speaker_dict = {}
print('Total speakers:', len(os.listdir(path)))
c = 0
for speaker in os.listdir(path):
if c % 10 == 0:
print('Speakers done:', c)
c += 1
print('Loading speaker: ' + speaker)
speaker_path = os.path.join(path, speaker)
speaker_dict[speaker] = [curr, None]
X_speaker = []
for file_path in os.listdir(speaker_path):
files = os.path.join(speaker_path, file_path)
audio, rate = librosa.load(files, sr=sr)
audio = np.expand_dims(audio, axis=1)
spec = spectrogram.spectrogram(audio)
X_speaker.append(spec)
curr += 1
speaker_dict[speaker][1] = curr - 1
X.append(X_speaker)
return X, speaker_dict
def saveBaselineIntensityImages(files,
saveLoc: str = None,
imageExt: str = "png"):
"""
Save a image file of the baseline as a function of time to the folder specified by saveLoc.
"""
if saveLoc == None:
saveLoc = r"../baselineIntensityMaps"
for i in tqdm.trange(len(files)):
filename = f"../dig/{files[i]}"
MySpect = Spectrogram(filename)
peaks, _, heights = baselines_by_squash(MySpect)
plt.plot(
np.array(MySpect.time) * 1e6,
MySpect.intensity[MySpect._velocity_to_index(peaks[0])])
plt.xlabel("Time ($\mu$s)")
plt.ylabel("Intensity (db)")
plt.title(f"{MySpect.data.filename}")
path = os.path.join(
saveLoc,
files[i].replace(".dig", "") + f"_baselineIntensity.{imageExt}")
if not os.path.exists(os.path.split(path)[0]):
os.makedirs(os.path.split(path)[0])
plt.savefig(path)
# Reset the state for the next data file.
plt.clf()
del MySpect
def make_spectrogram(pipe: PNSPipe, **kwargs):
"""
Compute a spectrogram using the passed kwargs or
built-in defaults and set the pipe.spectrogram field.
Default values are:
wavelength = 1.55e-6
points_per_spectrum = 4096
overlap = 7/8
window_function = None
form = 'power'
"""
from spectrogram import Spectrogram
defaults = dict(
t_start=kwargs.get('t_start'),
ending=kwargs.get('ending'),
wavelength=kwargs.get('wavelength', 1.55e-6),
points_per_spectrum=kwargs.get('points_per_spectrum', 4096),
overlap=kwargs.get('overlap', 7 / 8),
window_function=kwargs.get('window_function'),
form=kwargs.get('form', 'power'),
)
allargs = dict({**defaults, **kwargs})
# list the arguments we use in the log
for k, v in allargs.items():
pipe.log(f"+ {k} = {v}")
pipe.spectrogram = Spectrogram(pipe.df, **allargs)
def createGallery(digsToLookAt: list = None,
colormap="blue-orange-div",
fileext="jpeg",
transformData=False):
if type(digsToLookAt) == type(None):
digsToLookAt = DigFile.inventory(justSegments=True)['file']
digDir = DigFile.dig_dir()
imageDir, _ = os.path.split(digDir)
imageFold = "ImageGallery"
imageDir = os.path.join(imageDir, imageFold)
print(digsToLookAt)
for i in range(len(digsToLookAt)):
filename = digsToLookAt[i]
print(filename)
spec = Spectrogram(os.path.join(digDir, filename))
pcms, lastAxes = spec.plot(transformData, cmap=colormap)
fileloc, filename = os.path.split(
os.path.splitext(os.path.join(imageDir, filename))[0])
fileloc = os.path.join(fileloc, filename)
for key in pcms.keys():
if "complex" in key:
continue # There is not a graph with a name that contains complex.
if not os.path.exists(fileloc):
os.makedirs(fileloc)
plt.figure(num=key) # Get to the appropriate figure.
plt.savefig(
os.path.join(fileloc, filename) +
f' {key} spectrogram.{fileext}')
plt.clf()
del spec
def __init__(self, file_pth: str, train=True, random_state=None):
super(Dataset, self).__init__()
self.sp = Spectrogram()
self.file_pth = file_pth
self.train = train
self.random_state = None
self.train = self.read_data()
def get_data(self):
print('Reading Data...')
file_list = os.listdir(self.params['data_path'])
file_list = [
os.path.join(self.params['data_path'], i) for i in file_list[:10]
]
x_data = Spectrogram(filenames=file_list)
self.x_train, self.x_test = train_test_split(x_data.spectrogram,
test_size=0.3)
def test(self, filenames):
filenames = filenames[5:15]
for filename in filenames:
print(filename)
test_data = Spectrogram(filenames=[filename])
decoded_spectrogram = self.network.model.predict(
test_data.spectrogram)
#print_summary(self.network.model, line_length=80)
test_data.spectrogram_to_wav(
spectrogram=copy.deepcopy(decoded_spectrogram),
filename=filename.split("/")[-1])
test_data.visualize(filename=filename,
spectrogram=decoded_spectrogram)
def __init__(self, file_pth: str, train=True, random_state=None):
super(Dataset, self).__init__()
self.sp = Spectrogram()
self.file_pth = file_pth
self.train = train
self.random_state = None
self.data = self.read_data()
datasets = train_test_split(self.data)
if train:
self.train = datasets[0]
else:
# actually this is test dataset
self.train = datasets[1]
def generate_graphs(directory):
for file in os.listdir(directory):
print(directory + file)
if file.endswith(".dig"):
# Then we have the right file.
Spectrogram_Object = Spectrogram(directory + file)
sampleSize = 256
full_length_spectra = Spectrogram_Object.spectrogram(
0, Spectrogram_Object.samples, sampleSize)
colormaps = ["gist_stern", "tab20c", "tab20b", "terrain"]
for colormap in colormaps:
Spectrogram_Object.plot(full_length_spectra,
sample_window=sampleSize,
cmap=colormap)
def setupForDocumentation():
# Set up
filename = "../dig/CH_4_009.dig"
t1 = 14.2389/1e6
t2 = 31.796/1e6
MySpect = Spectrogram(filename, mode = "complex")
sTime = MySpect._time_to_index(t1)
eTime = MySpect._time_to_index(t2)
bottomVel = 1906.38
botInd = MySpect._velocity_to_index(bottomVel)
topVel = 5000
topVelInd = MySpect._velocity_to_index(topVel)
visualSignalStart = 2652.21
visSigIndex = MySpect._velocity_to_index(visualSignalStart)
return MySpect, sTime, eTime, botInd, topVelInd, visSigIndex
def setupForDocumentation():
# Set up
filename = "../See_if_this_is_enough_to_get_started/CH_4_009.dig"
t1 = 14.2389/1e6
t2 = 31.796/1e6
MySpect = Spectrogram(filename)
sTime = MySpect._time_to_index(t1)
eTime = MySpect._time_to_index(t2)
bottomVel = 1906.38
botInd = MySpect._velocity_to_index(bottomVel)
topVel = 5000
topVelInd = MySpect._velocity_to_index(topVel)
signalJump = 25
visualSignalStart = 2652.21
visSigIndex = MySpect._velocity_to_index(visualSignalStart)
return MySpect, sTime, eTime, signalJump, botInd, topVelInd, visSigIndex
def runAgainstTestSet(testSetFilename,
guessAtOptimalThres,
guessAtOptimalSkipUntil,
errorPower: int = 1):
data = pd.read_excel(testSetFilename)
# Ignore the samples that we do not have the full data for.
data = data.dropna()
data.reset_index() # Reset so that it can be easily indexed.
files = data["Filename"].to_list()
bestGuessTimes = (data[data.columns[1]] * 1e6).to_list()
# This will be the same times for the probe destruction dataset. It is generally ignored currently.
bestGuessVels = data[data.columns[-1]].to_list()
d = {"Filename": files}
d["ProbeDestructionEstimateTime"] = np.zeros(len(files))
d["Error Versus Label"] = np.zeros(len(files))
for i in tqdm.trange(len(files)):
filename = os.path.join(os.path.join("..", "dig"), f"{files[i]}")
MySpect = Spectrogram(filename, overlap=0)
peaks, _, heights = baselines_by_squash(MySpect, min_percent=0.75)
timeEstimate = baselineTracking(MySpect, peaks[0], guessAtOptimalThres,
guessAtOptimalSkipUntil)
d["ProbeDestructionEstimateTime"][i] = timeEstimate
d["Error Versus Label"][i] = np.power(bestGuessTimes[i] - timeEstimate,
errorPower)
errors = d["Error Versus Label"]
print(f"The statistics for the test set specified by {testSetFilename}\n")
print(
f"Avg: {np.mean(errors)} Std: {np.std(errors)} Median: {np.median(errors)} Max: {np.max(errors)} Min: {np.min(errors)}"
)
return d
def __init__(self, user_dir, contains_vocals, n_samples):
os.chdir(user_dir)
# The line below makes a list of all the filenames in the specified directory,
# while excluding names of subdirectories/folders
self.filename_list = [name for name in os.listdir('.') if os.path.isfile(name)]
print('Processing songs from:' + user_dir)
# For test set, n_samples = 24505
# For training set, n_samples = 46200
freq_size = 513
time_size = 23
self.n_samples = n_samples
# These are dimensions of the matrix [X, Y]
self.X = np.zeros((self.n_samples, freq_size, time_size,1), dtype=int)
self.Y = np.zeros((self.n_samples,1), dtype=int)
# This will ensure we are indexing correctly to put the x for our sample into our X matrix of all songs' data
n_samples_so_far = 0
for i in range(len(self.filename_list)):
filename = self.filename_list[i] # Iterates through each file name in the list
print(filename)
self.contains_vocals = contains_vocals
spect = Spectrogram(filename, contains_vocals)
x_is = spect.get_X()
y_is = spect.get_Y()
#print("x_is"+str(x_is.shape))
#print("y_is"+str(y_is.shape))
#print("samples so far "+str(n_samples_so_far))
n_song_samples = np.shape(y_is)[0] # I.e., how many samples did we make from the song
# (song length (sec)*2, since 500ms segments used)
upper_index = n_samples_so_far + n_song_samples # Each sample is a row in our array/matrix
#print("upper index "+str(upper_index))
self.X[n_samples_so_far:upper_index, :, :,:] = x_is
self.Y[n_samples_so_far:upper_index,:] = y_is
n_samples_so_far = n_samples_so_far + n_song_samples # Updates how many samples we've done so far
# sort the peaks and heights into descending order
peaks, heights = peaks[ordering], heights[ordering]
neighborhoods = []
width = 50
for peak_center in peaks:
low, high = max(0, peak_center -
width), min(len(powers) - 1, peak_center + width)
neighborhoods.append([velocities[low:high], powers[low:high]])
return neighborhoods
if __name__ == '__main__':
import os
os.chdir('../dig')
sgram = Spectrogram('./CH_4_009/seg00.dig', 0.0, 50.0e-6)
peaks, uncertainties, peak_heights = baselines_by_squash(sgram)
velos = []
times = [x for x in range(0,30)]
pcms, axes = sgram.plot(min_time=0, min_vel=100, max_vel=10000, cmap='3w_gby')
# add the peaks to the plot and change the color map range
pcm = pcms['intensity raw']
pcm.set_clim(-40, -65)
for peak in peaks:
velo_index = sgram._velocity_to_index(peak)
WHITE_CH4_SHOT/seg00.dig -- ??? opencv_long_start_pattern4 span=200
BLUE_CH1_SHOT/seg00.dig -- opencv_long_start_pattern4 span=150
BLUE_CH2_SHOT/seg00.dig -- ??? opencv_long_start_pattern3 span=200
BLUE_CH3_SHOT/seg00.dig -- opencv_long_start_pattern4 span=200
CH_1_009/seg00.dig -- opencv_long_start_pattern2 span=200
CH_3_009/seg00.dig -- opencv_long_start_pattern2 span=200
CH_4_009/seg01.dig -- opencv_long_start_pattern2 span=200
CH_4_009/seg02.dig -- opencv_long_start_pattern4 span=200
"""
from ProcessingAlgorithms.preprocess.digfile import DigFile
path = "../dig/BLUE_CH1_SHOT/seg00.dig"
df = DigFile(path)
spec = Spectrogram(df, 0.0, 60.0e-6, overlap_shift_factor= 1/8, form='db')
spec.availableData = ['intensity']
# print(spec.time[200])
# gives user the option to click, by default it searches from (0,0)
template_matcher = TemplateMatcher(spec,template=Templates.opencv_long_start_pattern5.value,
span=200,
k=20,
methods=['cv2.TM_CCOEFF_NORMED', 'cv2.TM_SQDIFF', 'cv2.TM_SQDIFF_NORMED'])
# masks the baselines to try and avoid matching with baselines or echoed signals
template_matcher.mask_baselines()
# get the times and velocities from matching
if x < 0:
return 0
if x > nmax:
return nmax
return x
for t in range(len(time) - 1):
if power[amax[t], t] > threshold:
acenter = amax[t]
vfrom, vto = peg(acenter - span), peg(acenter + span)
gus = Gaussian(velocity[vfrom:vto], power[vfrom:vto, t])
if gus.valid:
times.append(time[t])
centers.append(gus.center)
widths.append(gus.width)
amps.append(gus.amplitude)
else:
print(f"[{t}] {gus.error}")
#power[power>=threshold] = 1
plt.pcolormesh(time * 1e6, velocity, power)
plt.plot(np.array(times) * 1e6, centers, 'r.', alpha=0.5)
plt.show()
if __name__ == "__main__":
sp = Spectrogram('../dig/sample')
roi = dict(time=(13.6636e-6, 35.1402e-6), velocity=(2393.33, 4500.377))
analyze_region(sp, roi)
print("Hi!")
add_row(next(loc), 0.9, 0.8, 0.5) # to faded yellow
add_row(next(loc), 0.5, 0, 0.5)
add_row(next(loc), 0, 0.5, 0)
add_row(next(loc), 1, 0.75, 0.2)
add_row(next(loc), 0.2, 0.2, 0.2)
add_row(next(loc), 0, 0, 0)
except:
pass
add_row(1, 0, 0, 0)
return cdict
roids = calculate_centroids()
bounds = normalize_bounds(roids)
cdict = build_cdict(bounds)
from matplotlib.colors import LinearSegmentedColormap
COLORMAPS[name] = LinearSegmentedColormap(name, cdict)
return dict(name=name, centroids=roids, cmap=COLORMAPS[name])
if __name__ == '__main__':
sg = Spectrogram('../dig/CH_4_009.dig')
roids = make_spectrogram_color_map(sg, 4, "Kitty")
sg.plot(cmap = roids["cmap"])
import matplotlib.pyplot as plt
plt.xlim((0, 50))
plt.ylim((1500, 5000))
plt.show()
def runExperiment(trainingFilePath,
thresholds: list,
skipUntilTimes: list = [],
cmap: str = DEFMAP,
errorPower: int = 1):
data = pd.read_excel(trainingFilePath)
# Ignore the samples that we do not have the full data for.
data = data.dropna()
data.reset_index() # Reset so that it can be easily indexed.
files = data["Filename"].to_list()
bestGuessTimes = (data[data.columns[1]] * 1e6).to_list()
# This will be the same times for the probe destruction dataset. It is generally ignored currently.
bestGuessVels = data[data.columns[-1]].to_list()
d = {"Filename": files}
if skipUntilTimes == []:
skipUntilTimes = [12e-6]
cmapAttempt = cmap
if cmapAttempt in COLORMAPS.keys():
cmapAttempt = COLORMAPS[cmapAttempt]
for i in range(len(thresholds)):
d[f"threshold {thresholds[i]}"] = np.zeros(
(len(skipUntilTimes), len(files)))
d[f"threshold {thresholds[i]} error"] = np.zeros(
(len(skipUntilTimes), len(files)))
print("Running the experiment")
for i in tqdm.trange(len(files)):
filename = os.path.join(os.path.join("..", "dig"), f"{files[i]}")
MySpect = Spectrogram(filename, overlap=0)
peaks, _, heights = baselines_by_squash(MySpect, min_percent=0.75)
baselineInd = MySpect._velocity_to_index(peaks[0])
intensity = MySpect.intensity[baselineInd]
for skipTimeInd in range(len(skipUntilTimes)):
skipXInds = MySpect._time_to_index(skipUntilTimes[skipTimeInd])
outputTimeInds = np.array(baselineExperiment(
intensity, thresholds, skipXInds),
dtype=int)
for thresInd, thres in enumerate(thresholds):
timeEstimate = MySpect.time[outputTimeInds[thresInd]] * 1e6
d[f"threshold {thres}"][skipTimeInd][i] = timeEstimate
d[f"threshold {thres} error"][skipTimeInd][
i] = bestGuessTimes[i] - timeEstimate
del MySpect
# Compute summary statistics
summaryStatistics = np.zeros((len(thresholds), len(skipUntilTimes), 5))
stats = ["Avg", "Std", "Median", "Max", "Min"]
print(f"Computing the summary statistics: [{stats}]")
for i in tqdm.trange(len(thresholds)):
for j in range(len(skipUntilTimes)):
q = np.power(d[f"threshold {thresholds[i]} error"][j], errorPower)
summaryStatistics[i][j][0] = np.mean(q)
summaryStatistics[i][j][1] = np.std(q)
summaryStatistics[i][j][2] = np.median(q)
summaryStatistics[i][j][3] = np.max(q)
summaryStatistics[i][j][4] = np.min(q)
for i in tqdm.trange(len(stats)):
fig = plt.figure()
ax = plt.gca()
pcm = ax.pcolormesh(np.array(skipUntilTimes) * 1e6,
thresholds,
summaryStatistics[:, :, i],
cmap=cmapAttempt)
plt.title(f"{stats[i]} Error over the Files Tested")
plt.ylabel("Thresholds")
plt.xlabel("Time that you skip at the beginning ($\mu$s)")
plt.gcf().colorbar(pcm, ax=ax)
plt.show() # Need this for Macs.
summaryFolder = r"../JumpOffTimeEstimates"
if not os.path.exists(summaryFolder):
os.makedirs(summaryFolder)
for i in range(len(thresholds)):
q = pd.DataFrame(summaryStatistics[i])
internal_a = str(thresholds[i]).replace(".", "_")
saveSummaryFilename = f"summaryThresh{internal_a}.csv"
q.to_csv(os.path.join(summaryFolder, saveSummaryFilename))
return summaryStatistics, d
def __init__(self, *args, **kwargs):
"""
If one passes in a single unnamed arg, it can either be a digfile,
a string pointing to a digfile, or a two-dimensional ndarray.
If we are founded on a dig file, it is possible to recompute
things. Only a subset of operations are possible when we're
based on a two-dimensional array, but perhaps that is sometimes
desirable.
"""
self.digfile = None
if len(args) == 1:
arg = args[0]
if isinstance(arg, str):
self.digfile = DigFile(arg)
elif isinstance(arg, DigFile):
self.digfile = arg
if self.digfile == None and len(args):
# Let's see if we have enough information to display a spectrogram
# That means we have a two-dimensional ndarray, and possibly corresponding
# time and velocity arrays. The signature would be (intensity, [times, velocities]).
arg = args[0]
if isinstance(arg, np.ndarray) and len(arg.shape) == 2:
self._static = {
'intensity': arg,
}
if len(args) == 3:
self._static['time'] = np.asarray(args[1])
self._static['velocity'] = np.asarray(args[2])
else:
self._static['time'] = np.arange(0, -1 + arg.shape[1])
self._static['velocity'] = np.arange(0, -1 + arg.shape[0])
assert hasattr(
self, '_static'
), "Inappropriate arguments passed to the SpectrogramWidget constructor"
# If LaTeX is enabled in matplotlib, underscores in the title
# cause problems in displaying the histogram. However, we can
# solve this by not using latex in displaying the title, so there
# is no need to alter characters here.
self.title = "" if self.static else self.digfile.filename.split(
'/')[-1]
self.baselines = []
# Compute the base spectrogram (do we really need this?)
self.spectrogram = None
if self.dig:
self.spectrogram = Spectrogram(self.digfile, None, None, **kwargs)
self.spectrogram_fresh = True # flag for the first pass
self.spectrogram.overlap = 0.875
self.fig, axes = plt.subplots(nrows=1,
ncols=2,
sharey=True,
squeeze=True,
gridspec_kw=self._gspec)
self.axSpectrogram, self.axSpectrum = axes
self.subfig = None
self.axTrack = None
self.axSpare = None
# At the moment, clicking on the image updates the spectrum
# shown on the left axes. It would be nice to be more sophisticated and
# allow for more sophisticated interactions, including the ability
# to display more than one spectrum.
self.fig.canvas.mpl_connect('button_press_event',
lambda x: self.handle_click(x))
self.fig.canvas.mpl_connect('key_press_event',
lambda x: self.handle_key(x))
self.spectrum(None, "")
self.image = None # we will set in update_spectrogram
self.colorbar = None # we will set this on updating, based on the
self.peak_followers = [] # will hold any PeakFollowers
self.spectra = [] # will hold spectra displayed at right
self.spectra_in_db = True # should spectra be displayed in db?
self.controls = dict() # widgets stored by name
self.layout = None # how the controls get laid out
self.selecting = False # we are not currently selecting a ROI
self.roi = [] # and we have no regions of interest
self.threshold = None
self.make_controls(**kwargs)
display(self.layout)
self.update_spectrogram()
interesting, fixed = horizontal_lines(squashed)
fig = plt.figure()
ax1 = fig.add_subplot(1, 3, 1)
im1 = ax1.pcolormesh(self.time * 1e6, self.velocity,
20 * np.log10(self.intensity))
im1.set_cmap(COLORMAPS['3w_gby'])
fig.colorbar(im1, ax=ax1)
int2, fixed2 = horizontal_lines(20 * np.log10(self.intensity))
ax2 = fig.add_subplot(1, 3, 2)
im2 = ax2.pcolormesh(self.time * 1e6, self.velocity,
fixed2)
im2.set_cmap(COLORMAPS['3w_gby'])
fig.colorbar(im2, ax=ax2)
ax3 = fig.add_subplot(1, 3, 3)
im3 = ax3.pcolormesh(self.time * 1e6, self.velocity, fixed)
im3.set_cmap(COLORMAPS['3w_gby'])
fig.colorbar(im3, ax=ax3)
plt.show()
if __name__ == '__main__':
sp = Spectrogram('../dig/PDV_CHAN1BAK001')
gf = GrossFeatures(sp, time_chunk=8)
gf.k_means()
os.makedirs(fileloc)
plt.figure(num=key) # Get to the appropriate figure.
plt.savefig(
os.path.join(fileloc, filename) +
f' {key} spectrogram.{fileext}')
plt.clf()
del spec
if __name__ == "__main__":
# process command line args
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--file_name', type=str, default=False)
args = parser.parse_args()
#Just set directory here, I can't be bothered
directory = '/home/lanl/Documents/dig/new/CH_1_009/'
plt.rcParams["figure.figsize"] = [
20, 10
] # This is setting the default parameters to "20 by 10" inches by inches.
# I would prefer if this was set to full screen.
if args.file_name:
spec = Spectrogram(directory + args.file_name)
spec.plot()
plt.savefig(args.file_name + '.png')
plt.clf()
del spec
for ind in range(len(files)):
d = {"Template": Templates}
for k in range(len(keyedMethods)):
d[f"{keyedMethods[k]} Time microseconds"] = zeros
d[f"{keyedMethods[k]} Velocity (m/s)"] = zeros
d[f"{keyedMethods[k]} L2 error"] = zeros
d[f"{keyedMethods[k]} L1 error"] = zeros
outputFiles[ind] = d
print("The experiment is now set up")
# Run the experiment.
for i in tqdm.trange(len(files)):
filename = files[i]
fname = os.path.join(baseFolder, filename)
MySpect = Spectrogram(fname, overlap=7/8)
bestGuessTime = bestGuessTimes[i]
bestGuessVel = bestGuessVels[i]
template_matcher = TemplateMatcher(MySpect, None, Templates.start_pattern.value, span = 200, k=1)
valuesList = template_matcher.matchMultipleTemplates(methods, Templates)
for ind in range(len(valuesList)):
times, velos, scores, methodUsed = valuesList[ind]
for k in range(len(times)):
output[ind][keyedMethods[methodUsed[k]] + " Time microseconds"][i] = times[k]
output[ind][keyedMethods[methodUsed[k]] + " Velocity (m/s)"][i] = velos[k]
output[ind][keyedMethods[methodUsed[k]] + " L2 error"][i] = np.sqrt((times[k] - bestGuessTime)**2 + (velos[k] - bestGuessVel)**2)
output[ind][keyedMethods[methodUsed[k]] + " L1 error"][i] = np.abs(times[k] - bestGuessTime) + np.abs(velos[k] - bestGuessVel)
# importing os module
import os, os.path
DIR = "C:/Users/.../musdb18/test/Instrumental Versions" #The directory where your songs are stored
os.chdir(
DIR
) #Sets working directory to where song files are; this is so we can load them and write the dataset file in there
#The line below makes a list of all the filenames in the specified directory, while excluding names of subdirectories/folders
filename_list = [name for name in os.listdir('.') if os.path.isfile(name)]
# In[4]:
print('Processing songs from:' + DIR)
os.chdir(DIR)
for i in range(1, len(filename_list)):
filename = filename_list[i]
#Iterates through each file name in the list
print(filename)
contains_vocals = 0
#If the songs are instrumental/karaoke, it should be 0; if it has vocals, value should be 1
Spectrogram(
filename, 0
) #Calling the Spectrogram class. This is what creates/updates our dataset textfile
# In[5]:
filename_list
def buildXCSpectrogram(filename, fileid, speciesid, outDirSpectrogram,
outDirTrainingVectors):
"""
Build spectrogram data for given file
:param filename:
:param fileid
:param speciesid:
:param outDirSpectrogram:
:param outDirTrainingVectors:
:return:
"""
genSpec = Spectrogram()
#pick up audio file, make the spectrogram from it and save spec, plus serialise the data for training
genSpec.load(filename)
#spectrogram computation on entire waveform file
spectrogramDBFS = genSpec.spectrogramDBFS #this is the db relative full scale power spectra
spectrogramMag = genSpec.spectrogramMag #and this is the raw linear power spectra
freq = genSpec.freq #these are the frequency bands that go with the above
print("spec check 1: ", np.min(spectrogramDBFS[0]),
np.max(spectrogramDBFS[0]))
#now plot the data and save the training frames
print("spectrogram feature frames: ", len(spectrogramDBFS))
print("spectrogram=", np.shape(spectrogramDBFS))
#At this point we have a number of possibilities. There is the spectrumDBFS, which is the DB relative
#full scale log magnitude power spectrum, or the raw linear magnitude power spectrum which is just from
#the raw fft data directly (magnitude of the im, re vector). Either of these can be median filtered
#and then converted to a mel scale.
genSpec.plotSpectrogram(
spectrogramDBFS, freq,
os.path.join(outDirSpectrogram, 'xc' + fileid + '_spec_dbfs.png'))
genSpec.plotSpectrogram(
spectrogramMag, freq,
os.path.join(outDirSpectrogram, 'xc' + fileid + '_spec_mag.png'))
spectrogramDBFS_Filt = genSpec.spectrogramMedianFilter(spectrogramDBFS)
spectrogramMag_Filt = genSpec.spectrogramMedianFilter(spectrogramMag)
genSpec.plotSpectrogram(
spectrogramDBFS_Filt, freq,
os.path.join(outDirSpectrogram, 'xc' + fileid + '_spec_dbfs_med.png'))
genSpec.plotSpectrogram(
spectrogramMag_Filt, freq,
os.path.join(outDirSpectrogram, 'xc' + fileid + '_spec_mag_med.png'))
#now you can do a mel frequency one off of either spectrogram[Mag|DBFS] or spectrogram[Mag|DBFS]_Filt (median filtered)
#melfreq, spectrogramMels = genSpec.melSpectrum(spectrogramMag_Filt,freq)
#print "mels=",np.shape(spectrogramMels)
#genSpec.plotSpectrogram(spectrogramMels,melfreq,os.path.join(outDirSpectrogram,'xc'+fileid+'_spec_mel_mag_med.png'))
#spectrogramMelsLog = genSpec.logSpectrogram(spectrogramMels) #this applies a log function to the magnitudes
#genSpec.plotSpectrogram(spectrogramMelsLog,melfreq,os.path.join(outDirSpectrogram,'xc'+fileid+'_spec_mel_log_mag_med.png'))
#and rms normalisation
#and silence removal
#serialise the data from the spectrogram here so we can go back to it quickly later if needed
#Save vector data needed for learning
output = open(
os.path.join(outDirTrainingVectors,
speciesid + '_xc' + fileid + '_dbfs.pkl'), 'wb')
pickle.dump(spectrogramDBFS, output)
output.close()
output = open(
os.path.join(outDirTrainingVectors,
speciesid + '_xc' + fileid + '_mag.pkl'), 'wb')
pickle.dump(spectrogramMag, output)
output.close()
output = open(
os.path.join(outDirTrainingVectors,
speciesid + '_xc' + fileid + '_freq.pkl'), 'wb')
pickle.dump(freq, output)
output.close()
print("spec check 2: ", np.min(spectrogramDBFS[0]),
np.max(spectrogramDBFS[0]))
return spectrogramDBFS, spectrogramMag, freq
allDigs = DigFile.inventory()[
"file"] # Just the files that are not segments.
saveLoc = os.path.join(
os.path.split(digFolder)[0], "TotalIntensityMaps\\Median\\")
if not os.path.exists(saveLoc):
os.makedirs(saveLoc)
fractions = np.arange(10)
fractionsL = len(fractions)
data = np.zeros((fractionsL, len(allDigs)))
data[:, 0] = fractions
for i in range(len(allDigs)):
filename = allDigs[i]
path = os.path.join(digFolder, filename)
spec = Spectrogram(path)
inTen = np.sum(spec.intensity, axis=0)
# Offset the values so that everything is non negative.
inTen = inTen - np.min(inTen)
for j in range(fractionsL):
frac = np.power(10.0, -1 * fractions[j])
data[j, i] = len(threshold(inTen, frac))
del spec
name = os.path.join(saveLoc, "NumberOfLowValuesVsFracOfMax.txt")
np.savetxt(name, data)
from spectrogram import Spectrogram
def get_file_name(instrument, pitch, index = "000", velocity = "050"):
return "../nsynth-train/" + instrument + "/" + instrument + "_" \
+ index + "-" + pitch + "-" + velocity + ".wav"
guitar = [get_file_name("guitar_acoustic", i) for i in \
["055", "060", "080", "100"]]
keyboard = [get_file_name("keyboard_acoustic", i) for i in \
["055", "060", "080", "100"]]
for f in guitar + keyboard:
d = Spectrogram(f)
d.Visualize("random")
