def calc_erbs(low_freq, fs, n_filters):
ear_q = 9.26449 # Glasberg and Moore Parameters
min_bw = 24.7
order = 1
erbs = ((centre_freqs(fs, n_filters, low_freq)/ear_q)**order + min_bw**order)**(1/order)
return erbs
def calc_erbs(low_freq, fs, n_filters):
ear_q = 9.26449 # Glasberg and Moore Parameters
min_bw = 24.7
order = 1
erbs = ((centre_freqs(fs, n_filters, low_freq)/ear_q)**order + min_bw**order)**(1/order)
return erbs
def test_make_erb_filters():
hi = 100
lo = 11025
oldcf = oldfilt.centre_freqs(44100, 100, 20)
newcf = newfilt.centre_freqs(44100, 100, 20)
t0 = time.time()
old = oldfilt.make_erb_filters(44100, oldcf, width=1.0)
t1 = time.time()
new = newfilt.make_erb_filters(44100, newcf, width=1.0)
t2 = time.time()
print(
f'Old method took {t1 - t0} seconds, New method took {t2 - t1} seconds.'
)
assert np.allclose(old, new)
def srmr(x, fs, n_cochlear_filters=23, low_freq=125, min_cf=4, max_cf=128, fast=True, norm=False):
wLengthS = .256
wIncS = .064
# Computing gammatone envelopes
if fast:
mfs = 400.0
gt_env = fft_gtgram(x, fs, 0.010, 0.0025, n_cochlear_filters, low_freq)
else:
cfs = centre_freqs(fs, n_cochlear_filters, low_freq)
fcoefs = make_erb_filters(fs, cfs)
gt_env = np.abs(hilbert(erb_filterbank(x, fcoefs)))
mfs = fs
wLength = int(np.ceil(wLengthS*mfs))
wInc = int(np.ceil(wIncS*mfs))
# Computing modulation filterbank with Q = 2 and 8 channels
mod_filter_cfs = compute_modulation_cfs(min_cf, max_cf, 8)
MF = modulation_filterbank(mod_filter_cfs, mfs, 2)
n_frames = int(1 + (gt_env.shape[1] - wLength)//wInc)
w = hamming(wLength+1)[:-1] # window is periodic, not symmetric
energy = np.zeros((n_cochlear_filters, 8, n_frames))
for i, ac_ch in enumerate(gt_env):
mod_out = modfilt(MF, ac_ch)
for j, mod_ch in enumerate(mod_out):
mod_out_frame = segment_axis(mod_ch, wLength, overlap=wLength-wInc, end='pad')
energy[i,j,:] = np.sum((w*mod_out_frame[:n_frames])**2, axis=1)
if norm:
energy = normalize_energy(energy)
erbs = np.flipud(calc_erbs(low_freq, fs, n_cochlear_filters))
avg_energy = np.mean(energy, axis=2)
total_energy = np.sum(avg_energy)
AC_energy = np.sum(avg_energy, axis=1)
AC_perc = AC_energy*100/total_energy
AC_perc_cumsum=np.cumsum(np.flipud(AC_perc))
K90perc_idx = np.where(AC_perc_cumsum>90)[0][0]
BW = erbs[K90perc_idx]
cutoffs = calc_cutoffs(mod_filter_cfs, fs, 2)[0]
if (BW > cutoffs[4]) and (BW < cutoffs[5]):
Kstar=5
elif (BW > cutoffs[5]) and (BW < cutoffs[6]):
Kstar=6
elif (BW > cutoffs[6]) and (BW < cutoffs[7]):
Kstar=7
elif (BW > cutoffs[7]):
Kstar=8
return np.sum(avg_energy[:, :4])/np.sum(avg_energy[:, 4:Kstar]), energy
def srmr(x, fs, n_cochlear_filters=23, low_freq=125, min_cf=4, max_cf=128, fast=True, norm=False):
wLengthS = .256
wIncS = .064
# Computing gammatone envelopes
if fast:
mfs = 400.0
gt_env = fft_gtgram(x, fs, 0.010, 0.0025, n_cochlear_filters, low_freq)
else:
cfs = centre_freqs(fs, n_cochlear_filters, low_freq)
fcoefs = make_erb_filters(fs, cfs)
gt_env = np.abs(hilbert(erb_filterbank(x, fcoefs)))
mfs = fs
wLength = np.ceil(wLengthS*mfs)
wInc = np.ceil(wIncS*mfs)
# Computing modulation filterbank with Q = 2 and 8 channels
mod_filter_cfs = compute_modulation_cfs(min_cf, max_cf, 8)
MF = modulation_filterbank(mod_filter_cfs, mfs, 2)
n_frames = 1 + (gt_env.shape[1] - wLength)//wInc
w = hamming(wLength+1)[:-1] # window is periodic, not symmetric
energy = np.zeros((n_cochlear_filters, 8, n_frames))
for i, ac_ch in enumerate(gt_env):
mod_out = modfilt(MF, ac_ch)
for j, mod_ch in enumerate(mod_out):
mod_out_frame = segment_axis(mod_ch, wLength, overlap=wLength-wInc, end='pad')
energy[i,j,:] = np.sum((w*mod_out_frame[:n_frames])**2, axis=1)
if norm:
energy = normalize_energy(energy)
erbs = np.flipud(calc_erbs(low_freq, fs, n_cochlear_filters))
avg_energy = np.mean(energy, axis=2)
total_energy = np.sum(avg_energy)
AC_energy = np.sum(avg_energy, axis=1)
AC_perc = AC_energy*100/total_energy
AC_perc_cumsum=np.cumsum(np.flipud(AC_perc))
K90perc_idx = np.where(AC_perc_cumsum>90)[0][0]
BW = erbs[K90perc_idx]
cutoffs = calc_cutoffs(mod_filter_cfs, fs, 2)[0]
if (BW > cutoffs[4]) and (BW < cutoffs[5]):
Kstar=5
elif (BW > cutoffs[5]) and (BW < cutoffs[6]):
Kstar=6
elif (BW > cutoffs[6]) and (BW < cutoffs[7]):
Kstar=7
elif (BW > cutoffs[7]):
Kstar=8
return np.sum(avg_energy[:, :4])/np.sum(avg_energy[:, 4:Kstar]), energy
def _computeSingleFrameFeature(self,sig):
'''Feature computation for a single time-series frame/segment
Args:
sig (numpy array): The signal segment for which feature will be computed
Returns:
feature (numpy array): Computed feature vector
單個時間序列幀/段的特徵計算
(只限 ”SubEnv” 子帶包絡(Sub-band envelopes)特徵計算)
- 輸入變數 : sig (numpy array)
- 輸出變數 : feature (numpy array)
'''
if self.name=='SubEnv':
'''Sub-band envelopes feature computation 子帶包絡特徵計算'''
#Computing sub-band signals /計算子帶信號
timeRes=self.dimensions[0]
numBands=self.dimensions[1]
low_cut_off=2#lower cut off frequency = 2Hz /較低的截止頻率= 2Hz
centre_freqVals = centre_freqs(self.samplerate,numBands,low_cut_off)
fcoefs = make_erb_filters(self.samplerate, centre_freqVals, width=1.0)
y = erb_filterbank(sig, fcoefs)
subenv = np.array([]).reshape(timeRes,0)
for i in range(numBands):
subBandSig=y[i,:]
analytic_signal = hilbert(subBandSig)
amp_env = np.abs(analytic_signal)
np.nan_to_num(amp_env)
#amp_env=resampy.resample(amp_env, len(amp_env), timeRes, axis=-1)#resampy library used resampling /resampy庫使用重新取樣
#resampling may lead to unexpected computation errors, /重新採樣可能會導致意外的計算錯誤，
#I prefered average amplitudes for short-time windows /我更喜歡短時間窗口的平均幅度
downSampEnv=np.zeros((timeRes,1))
winSize=int(len(amp_env)/timeRes)
for ind in range(timeRes):
downSampEnv[ind]=np.log2(np.mean(amp_env[ind*winSize:(ind+1)*winSize]))
subenv=np.hstack([subenv,downSampEnv])
#removing mean and normalizing /刪除均值和正常化
subenv=subenv-np.mean(subenv)
subenv=subenv/(np.max(np.abs(subenv)))
feature=subenv
else:
print('Error: feature '+self.name+' is not recognized')
feature=[]
return feature
def EvaluateRandom(count=None, LPF=False, CUTOFF=50):
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' # Silence tensorflow logs
TotalTime = time.time()
if not os.path.isdir("graphs"):
os.mkdir('graphs')
os.mkdir(os.path.join('graphs', 'FallingOrRising'))
# Get all the WAV files under resources/fcnn
wavFiles = glob.glob(os.path.join('resources', 'f2cnn', '*', '*.WAV'))
print(
"\n###############################\nEvaluating network on {} WAV files in '{}'."
.format(len(wavFiles),
os.path.split(wavFiles[0])[0]))
if not wavFiles:
print("NO WAV FILES FOUND")
exit(-1)
# Reading the config file
config = ConfigParser()
config.read('configF2CNN.conf')
framerate = config.getint('FILTERBANK', 'FRAMERATE')
nchannels = config.getint('FILTERBANK', 'NCHANNELS')
lowcutoff = config.getint('FILTERBANK', 'LOW_FREQ')
# CENTER FREQUENCIES ON ERB SCALE
CENTER_FREQUENCIES = filters.centre_freqs(framerate, nchannels, lowcutoff)
FILTERBANK_COEFFICIENTS = filters.make_erb_filters(framerate,
CENTER_FREQUENCIES)
# Selecting some random files, or all of them
if count is None:
numpy.random.shuffle(wavFiles)
elif count > 1:
wavFiles = numpy.random.choice(wavFiles, count)
for file in wavFiles:
EvaluateOneWavFile(file,
LPF=LPF,
CUTOFF=CUTOFF,
CENTER_FREQUENCIES=CENTER_FREQUENCIES,
FILTERBANK_COEFFICIENTS=FILTERBANK_COEFFICIENTS)
print("Evaluating network on all files.")
print(' Total time:', time.time() - TotalTime)
print('')
def FilterAllOrganisedFiles():
TotalTime = time.time()
# Get all the WAV files under resources
# wavFiles = glob.glob(join("resources", "f2cnn", "*", "*.WAV"))
wavFiles = glob.glob(os.path.join("resources", "f2cnn", "**", "*.WAV"))
print(
"\n###############################\nApplying FilterBank to files in '{}'."
.format(os.path.split(wavFiles[0])[0]))
if not wavFiles:
print("NO WAV FILES FOUND, PLEASE ORGANIZE FILES")
exit(-1)
print(len(wavFiles), "files found")
# #### READING CONFIG FILE
config = ConfigParser()
config.read('configF2CNN.conf')
framerate = config.getint('FILTERBANK', 'FRAMERATE')
nchannels = config.getint('FILTERBANK', 'NCHANNELS')
lowcutoff = config.getint('FILTERBANK', 'LOW_FREQ')
# ##### PREPARATION OF FILTERBANK
# CENTER FREQUENCIES ON ERB SCALE
CENTER_FREQUENCIES = filters.centre_freqs(framerate, nchannels, lowcutoff)
# Filter coefficient for a Gammatone filterbank
FILTERBANK_COEFFICIENTS = filters.make_erb_filters(framerate,
CENTER_FREQUENCIES)
# Usage of multiprocessing, to reduce computing time
proc = cpu_count()
counter = Value('i', 0)
multiproc_pool = Pool(processes=proc,
initializer=InitProcesses,
initargs=(
FILTERBANK_COEFFICIENTS,
counter,
))
multiproc_pool.starmap(GammatoneFiltering,
zip(wavFiles, repeat(len(wavFiles))))
print("Filtered and Saved all files.")
print(' Total time:', time.time() - TotalTime)
print('')
def gammatone_bank(
wav: NDVar,
f_min: float,
f_max: float,
n: int,
integration_window: float = 0.010,
tstep: float = None,
location: str = 'right',
pad: bool = True,
name: str = None,
) -> NDVar:
"""Gammatone filterbank response
Parameters
----------
wav : NDVar
Sound input.
f_min : scalar
Lower frequency cutoff.
f_max : scalar
Upper frequency cutoff.
n : int
Number of filter channels.
integration_window : scalar
Integration time window in seconds (default 10 ms).
tstep : scalar
Time step size in the output (default is same as ``wav``).
location : str
Location of the output relative to the input time axis:
- ``right``: gammatone sample at end of integration window (default)
- ``left``: gammatone sample at beginning of integration window
- ``center``: gammatone sample at center of integration window
Since gammatone filter response depends on ``integration_window``, the
filter response will be delayed relative to the analytic envlope. To
ignore this delay, use `location='left'`
pad : bool
Pad output to match time axis of input.
name : str
NDVar name (default is ``wav.name``).
Notes
-----
Requires the ``fmax`` branch of the gammatone library to be installed:
$ pip install https://github.com/christianbrodbeck/gammatone/archive/fmax.zip
"""
from gammatone.filters import centre_freqs, erb_filterbank
from gammatone.gtgram import make_erb_filters
wav_ = wav
if location == 'left':
if pad:
wav_ = _pad_func(wav, wav.time.tmin - integration_window)
elif location == 'right':
# tmin += window_time
if pad:
wav_ = _pad_func(wav, tstop=wav.time.tstop + integration_window)
elif location == 'center':
dt = integration_window / 2
# tmin += dt
if pad:
wav_ = _pad_func(wav, wav.time.tmin - dt, wav.time.tstop + dt)
else:
raise ValueError(f"mode={location!r}")
fs = 1 / wav.time.tstep
if tstep is None:
tstep = wav.time.tstep
wave = wav_.get_data('time')
# based on gammatone library, rewritten to reduce memory footprint
cfs = centre_freqs(fs, n, f_min, f_max)
integration_window_len = int(round(integration_window * fs))
output_n_samples = floor((len(wave) - integration_window_len) * wav.time.tstep / tstep)
output_step = tstep / wav.time.tstep
results = []
for i, cf in tqdm(enumerate(reversed(cfs)), "Gammatone spectrogram", total=len(cfs), unit='band'):
fcoefs = np.flipud(make_erb_filters(fs, cf))
xf = erb_filterbank(wave, fcoefs)
results.append(aggregate(xf[0], output_n_samples, output_step, integration_window_len))
result = np.sqrt(results)
# package output
freq_dim = Scalar('frequency', cfs[::-1], 'Hz')
time_dim = UTS(wav.time.tmin, tstep, output_n_samples)
if name is None:
name = wav.name
return NDVar(result, (freq_dim, time_dim), name)
def gammatone_bank(wav: NDVar,
f_min: float,
f_max: float,
n: int,
integration_window: float = 0.010,
tstep: float = None,
location: str = 'right',
pad: bool = True,
name: str = None) -> NDVar:
"""Gammatone filterbank response
Parameters
----------
wav : NDVar
Sound input.
f_min : scalar
Lower frequency cutoff.
f_max : scalar
Upper frequency cutoff.
n : int
Number of filter channels.
integration_window : scalar
Integration time window in seconds (default 10 ms).
tstep : scalar
Time step size in the output (default is same as ``wav``).
location : str
Location of the output relative to the input time axis:
- ``right``: gammatone sample at end of integration window (default)
- ``left``: gammatone sample at beginning of integration window
- ``center``: gammatone sample at center of integration window
Since gammatone filter response depends on ``integration_window``, the
filter response will be delayed relative to the analytic envlope. To
ignore this delay, use `location='left'`
pad : bool
Pad output to match time axis of input.
name : str
NDVar name (default is ``wav.name``).
Notes
-----
Requires the ``fmax`` branch of the gammatone library to be installed:
$ pip install https://github.com/christianbrodbeck/gammatone/archive/fmax.zip
"""
from gammatone.filters import centre_freqs
from gammatone.gtgram import gtgram
tmin = wav.time.tmin
wav_ = wav
if location == 'left':
if pad:
wav_ = _pad_func(wav, wav.time.tmin - integration_window)
elif location == 'right':
# tmin += window_time
if pad:
wav_ = _pad_func(wav, tstop=wav.time.tstop + integration_window)
elif location == 'center':
dt = integration_window / 2
# tmin += dt
if pad:
wav_ = _pad_func(wav, wav.time.tmin - dt, wav.time.tstop + dt)
else:
raise ValueError(f"mode={location!r}")
sfreq = 1 / wav.time.tstep
if tstep is None:
tstep = wav.time.tstep
x = gtgram(wav_.get_data('time'), sfreq, integration_window, tstep, n,
f_min, f_max)
freqs = centre_freqs(sfreq, n, f_min, f_max)
# freqs = np.round(freqs, out=freqs).astype(int)
freq_dim = Scalar('frequency', freqs[::-1], 'Hz')
time_dim = UTS(tmin, tstep, x.shape[1])
return NDVar(x, (freq_dim, time_dim), name or wav.name)
print("gtgram_function.shape", gtgram_function.shape)
print("gtgram_function.T", gtgram_function.T.shape)
matplotlib.pyplot.plot(gtgram_function.T)
axes.set_title("gtgram_function.T 1ch. " + os.path.basename(new_file_name_path))
matplotlib.pyplot.show()
ipdb.set_trace()
"""
#ssc = ssc(sig,samplerate=rate)
#print(logfbank_feat[1:3,:])
#filters.centre_freqs(fs, num_freqs, cutoff)
centre_freqs = filters.centre_freqs(rate, sig.shape[0], 100)
axes.set_title("centre_freqs"+ str(centre_freqs.shape)+" " + os.path.basename(new_file_name_path))
axes.set_xlabel("Time (s)")
axes.set_ylabel("Frequency")
print("centre_freqs.shape", centre_freqs.shape)
matplotlib.pyplot.plot(centre_freqs)
matplotlib.pyplot.show()
ipdb.set_trace()
erb_filters = filters.make_erb_filters(rate, centre_freqs, width=1.0)
axes.set_title("erb_filters"+ str(erb_filters.shape)+" " + os.path.basename(new_file_name_path))
axes.set_xlabel("Time (s)")
axes.set_ylabel("Frequency")
print("erb_filters.shape", erb_filters.shape)
def predict(self, clean, mixture, noise):
# Computing gammatone envelopes
if self.fast:
mfs = 400.0
gt_env = fft_gtgram(mixture, self.fs, 0.010, 0.0025,
self.n_cochlear_filters, self.low_freq)
else:
cfs = centre_freqs(self.fs, self.n_cochlear_filters, self.low_freq)
fcoefs = make_erb_filters(self.fs, cfs)
gt_env = np.abs(hilbert(erb_filterbank(mixture, fcoefs)))
mfs = self.fs
wLength = np.ceil(self.wLengthS*mfs)
wInc = np.ceil(self.wIncS*mfs)
# Computing modulation filterbank with Q = 2 and 8 channels
mod_filter_cfs = compute_modulation_cfs(self.min_cf, self.max_cf, 8)
MF = modulation_filterbank(mod_filter_cfs, mfs, 2)
n_frames = np.ceil((gt_env.shape[1])/wInc)
w = hamming(wLength)
energy = np.zeros((self.n_cochlear_filters, 8, n_frames))
for i, ac_ch in enumerate(gt_env):
mod_out = modfilt(MF, ac_ch)
for j, mod_ch in enumerate(mod_out):
mod_out_frame = segment_axis(mod_ch, wLength, overlap=wLength-wInc, end='delay')
energy[i,j,:] = np.sum((w*mod_out_frame)**2, axis=1)
if self.norm:
peak_energy = np.max(np.mean(energy, axis=0))
min_energy = peak_energy*0.001
energy[energy < min_energy] = min_energy
energy[energy > peak_energy] = peak_energy
erbs = np.flipud(self.calc_erbs(self.low_freq, self.fs,
self.n_cochlear_filters))
avg_energy = np.mean(energy, axis=2)
total_energy = np.sum(avg_energy)
AC_energy = np.sum(avg_energy, axis=1)
AC_perc = AC_energy*100/total_energy
AC_perc_cumsum=np.cumsum(np.flipud(AC_perc))
K90perc_idx = np.where(AC_perc_cumsum>90)[0][0]
BW = erbs[K90perc_idx]
cutoffs = self.calc_cutoffs(mod_filter_cfs, self.fs, 2)[0]
if (BW > cutoffs[4]) and (BW < cutoffs[5]):
Kstar=5
elif (BW > cutoffs[5]) and (BW < cutoffs[6]):
Kstar=6
elif (BW > cutoffs[6]) and (BW < cutoffs[7]):
Kstar=7
elif (BW > cutoffs[7]):
Kstar=8
out = {'p': {
'srmr': np.sum(avg_energy[:, :4]) / np.sum(avg_energy[:, 4:Kstar])},
'avg_energy': avg_energy
}
return out
def compute(filepath, file):
modelpath = 'C:/Users/user/Desktop/cnn/data/model/M_uocSeq1SubEnv32by16_nASyn2000len_1000hopt.h5'
#dir='C:/Users/Lab606B/Desktop/result/'#txt 儲存路徑
#wildcard="txt"
# fileLabels=['1']
timeDim = 32
freqDim = 16
frameSizeMs = 2000
hopSizeMs = 1000
signal, samplerate = sf.read(filepath + file)
lenSigSamp = len(signal)
lenSigMs = 1000 * lenSigSamp / samplerate
lenSigMs = lenSigMs
startsMs = list(np.arange(0, lenSigMs - frameSizeMs, hopSizeMs))
stopsMs = [x + frameSizeMs for x in startsMs]
#windowing using segmentation info and performing feature extraction /使用分段信息進行窗口化並執行特徵提取
starts = [int(round(x * samplerate / 1000)) for x in startsMs]
stops = [int(round(x * samplerate / 1000)) for x in stopsMs]
globalInd = 0
allFeatures = np.zeros((1, timeDim, freqDim))
# allLabels=[]
#(無用) fileSegmentMap={}#map containing filename versus indexes of segments/features within all samples in this set /包含文件名的映射與此集合中所有樣本中的段/要素的索引
for ind in range(len(starts)):
segment = signal[starts[ind]:stops[ind]]
#applying windowing function to the segment /將窗口函數應用於段
segment = segment * create_window(
stops[ind] - starts[ind], 'tukey', r=0.08)
if (np.max(segment) > 0): #normalization /正規化
segment = segment / np.max(segment)
#feature=Feature._computeSingleFrameFeature(segment)
'''Sub-band envelopes feature computation 子帶包絡特徵計算'''
#Computing sub-band signals /計算子帶信號
low_cut_off = 2 #lower cut off frequency = 2Hz /較低的截止頻率= 2Hz
centre_freqVals = centre_freqs(samplerate, freqDim, low_cut_off)
fcoefs = make_erb_filters(samplerate, centre_freqVals, width=1.0)
y = erb_filterbank(segment, fcoefs)
subenv = np.array([]).reshape(timeDim, 0)
for i in range(freqDim):
subBandSig = y[i, :]
analytic_signal = hilbert(subBandSig)
amp_env = np.abs(analytic_signal)
np.nan_to_num(amp_env)
#amp_env=resampy.resample(amp_env, len(amp_env), timeRes(timeDim), axis=-1)#resampy library used resampling /resampy庫使用重新取樣
#resampling may lead to unexpected computation errors, /重新採樣可能會導致意外的計算錯誤，
#I prefered average amplitudes for short-time windows /我更喜歡短時間窗口的平均幅度
downSampEnv = np.zeros((timeDim, 1))
winSize = int(len(amp_env) / timeDim)
for ind in range(timeDim):
downSampEnv[ind] = np.log2(
np.mean(amp_env[ind * winSize:(ind + 1) * winSize]))
subenv = np.hstack([subenv, downSampEnv])
#removing mean and normalizing /刪除均值和正常化
subenv = subenv - np.mean(subenv)
subenv = subenv / (np.max(np.abs(subenv)))
feature = subenv
#adding computed feature /添加計算特徵
if globalInd == 0: #if this is the first feature assign it directly /如果這是第一個功能直接分配它
allFeatures[0] = feature
else: #add one more element in the feature vector and then assign /在特徵向量中添加一個元素，然後分配
allFeatures = np.vstack(
[allFeatures, np.zeros((1, timeDim, freqDim))])
allFeatures[globalInd] = feature
#(無用) #adding segment to file-segment map /將段添加到文件段映射
#(無用) if file in fileSegmentMap:#if file already exists, append segment /如果文件已存在，則追加段
#(無用) val=fileSegmentMap[file]
#(無用) val.append(globalInd)
#(無用) fileSegmentMap[file]=val
#(無用) else:#file does not exist in map, add the first file-segment map /文件在地圖中不存在，添加第一個文件段映射
#(無用) fileSegmentMap[file]=[globalInd]
#(無用) allLabels.append(fileLabels)
globalInd += 1
#(無用) allFeatures=allFeatures.reshape(allFeatures.shape[0],timeRes,numBands,1)
#(無用) allLabels=np.array(allLabels,dtype = np.int)
#(無用) allLabels = to_categorical(allLabels)
allFeatures = np.reshape(allFeatures,
[len(allFeatures), timeDim, freqDim, 1])
#(無用) with open(filepath+'Test_Features.pkl', 'wb') as f:
#(無用) pickle.dump(allFeatures, f, 1)
#(無用) with open(filepath+'Test_Labels.pkl' , 'wb') as f:
#(無用) pickle.dump(allLabels, f, 1)
#(無用) with open(filepath+'Test_Map.pkl', 'wb') as f:
#(無用) pickle.dump(fileSegmentMap, f, 1)
model = keras.models.load_model(modelpath)
y_probs = model.predict(allFeatures,
batch_size=allFeatures.shape[0],
verbose=0)
#normal = -1 = 0 ; abnormal = 1
normal = 0
abnormal = 0
for i in range(len(y_probs)):
if (y_probs[i, 0] > y_probs[i, 1]):
normal = normal + 1
else:
abnormal = abnormal + 1
if (normal > abnormal):
result = 'normal'
resultRate = normal / len(y_probs) * 100
elif (normal < abnormal):
result = 'abnormal'
resultRate = abnormal / len(y_probs) * 100
else:
result = 'not sure'
resultRate = 50
#建立txt檔
text_file_predict = open(
'C:/Users/user/Desktop/cnn/DataSpaceFoeFTP/Predict_Result/nxp/' +
file.replace('.wav', '') + ".txt",
"w",
encoding='utf-8')
#text_file_predict.write('test result(predict)\n')
text_file_predict.write('檔案:' + str(file))
text_file_predict.write('\n')
text_file_predict.write('\n診斷結果 =\t' + str(result))
text_file_predict.write('\n概率為 =\t' + str(resultRate) + '%')
text_file_predict.write('\n------------------------------------------\n')
# ListFilesToTxt(dir,file,wildcard, 1)
text_file_predict.close()
print('診斷結果為 : ', result)
print('機率為 : ', resultRate, '%')
def EvaluateOneWavArray(wavArray,
framerate,
wavFileName,
model='last_trained_model',
LPF=False,
CUTOFF=100,
CENTER_FREQUENCIES=None,
FILTERBANK_COEFFICIENTS=None):
# #### READING CONFIG FILE
config = ConfigParser()
config.read('configF2CNN.conf')
RADIUS = config.getint('CNN', 'RADIUS')
SAMPPERIOD = config.getint('CNN', 'SAMPLING_PERIOD')
NCHANNELS = config.getint('FILTERBANK', 'NCHANNELS')
DOTSPERINPUT = RADIUS * 2 + 1
USTOS = 1 / 1000000.
# Extracting labels, for accuracy computation
labels = ExtractLabel(wavFileName, config)
labels = [(entry[-4], entry[-1])
for entry in labels] if labels is not None else None
if CENTER_FREQUENCIES is None:
NCHANNELS = config.getint('FILTERBANK', 'NCHANNELS')
lowcutoff = config.getint('FILTERBANK', 'LOW_FREQ')
# ##### PREPARATION OF FILTERBANK
# CENTER FREQUENCIES ON ERB SCALE
CENTER_FREQUENCIES = filters.centre_freqs(framerate, NCHANNELS,
lowcutoff)
# Filter coefficients for a Gammatone filterbank
FILTERBANK_COEFFICIENTS = filters.make_erb_filters(
framerate, CENTER_FREQUENCIES)
print("Applying filterbank...")
filtered = GetFilteredOutputFromArray(wavArray, FILTERBANK_COEFFICIENTS)
del wavArray
if not LPF:
print("Extracting Envelope...")
else:
print(
"Extraction Envelope with {}Hz Low Pass Filter...".format(CUTOFF))
print(LPF, CUTOFF)
envelopes = ExtractEnvelopeFromMatrix(filtered, LPF, CUTOFF)
del filtered
print("Extracting Formants...")
fbPath = os.path.splitext(wavFileName)[0] + '.FB'
formants, sampPeriod = ExtractFBFile(fbPath)
print("Extracting Phonemes...")
phnPath = os.path.splitext(wavFileName)[0] + '.PHN'
phonemes = ExtractPhonemes(phnPath)
print("Generating input data for CNN...")
STEP = int(framerate * SAMPPERIOD * USTOS)
START = int(STEP * RADIUS)
nb = int(len(envelopes[0]) - DOTSPERINPUT * STEP)
input_data = numpy.zeros([nb, DOTSPERINPUT, NCHANNELS])
print("INPUT SHAPE:", input_data.shape)
for i in range(0, nb):
input_data[i] = [[
channel[START + i + (k - RADIUS) * STEP] for channel in envelopes
] for k in range(DOTSPERINPUT)]
for i, matrix in enumerate(input_data):
input_data[i] = normalizeInput(matrix)
input_data.astype('float32')
print("Evaluating the data with the pretrained model...")
import keras
model = keras.models.load_model(model)
scores = model.predict(input_data.reshape(nb, DOTSPERINPUT, NCHANNELS, 1),
verbose=1)
simplified_scores = [1 if score[1] > score[0] else 0 for score in scores]
# Attempt to compute an accuracy for the file. TODO: Doesn't take into account phonemes we use, step values
keras.backend.clear_session()
del model
del input_data
accuracy = None
if labels is not None:
accuracy = 0
total_valid = 0
for timepoint, score in enumerate(simplified_scores):
for index in range(len(labels) - 1):
before = labels[index][0]
after = labels[index + 1][0]
if before < timepoint < after and (
abs(timepoint - before) < STEP
or abs(timepoint - after) < STEP):
if abs(before - timepoint) <= abs(after - timepoint):
if score == labels[index][1]:
accuracy += 1
else:
if score == labels[index + 1][1]:
accuracy += 1
total_valid += 1
accuracy /= total_valid
print("Plotting...")
PlotEnvelopesAndCNNResultsWithPhonemes(envelopes, scores, accuracy,
CENTER_FREQUENCIES, phonemes,
formants, wavFileName)
del envelopes
del phonemes
atdata = np.empty([0, 5, 128, 128])
iter = 0
fold = 0
for x in range(1, len(meta)):
if int(meta[x][5]) == fold + 1:
filename, foldno, classID = meta[x][0], int(meta[x][5]), int(
meta[x][6])
s, sr = librosa.load('UrbanSound8K/audio/fold' + str(foldno) + '/' +
filename,
sr=44100)
fcoefs = filters.make_erb_filters(sr,
filters.centre_freqs(sr, 128, 40),
odr=4)
g = filters.erb_filterbank(s, fcoefs)
c = cp.asarray(g)
c = cp.power(c, 2)
if len(c[0]) // 66536 > 0 and len(c[0]) % 65536 < 65536 / 2:
nspecs = len(c[0]) // 65536
else:
nspecs = (len(c[0]) // 65536) + 1
if len(c[0]) < 65536 + 512:
c = cp.pad(c, ((0, 0), (0, 66048 - len(c[0]))),
'constant',
constant_values=0)
def erb_filter(self):
"""
For the input sampling frequency, get the ERB filters.
"""
return filters.make_erb_filters(self.fs,
filters.centre_freqs(self.fs, 64, 50))
def make_gammatone_filters(num_bins = 1024, cutoff_low = 30, sample_rate = 44100):
center_freqs = gt_filters.centre_freqs(sample_rate, num_bins, cutoff_low)
gammatone_filters = gt_filters.make_erb_filters(sample_rate, center_freqs)
return gammatone_filters