def Train(params, sess, saver, qn, corpus, envs, envsTest):
print("Start training")
for epoch in range(params.max_epochs):
#print("epoch", epoch)
for env in envs:
TIMER.Start("Trajectory")
arrRL, totReward, totDiscountedReward = Trajectory(env, params, sess, qn, corpus, False)
TIMER.Pause("Trajectory")
lastLangVisited = corpus.transitions[-1].langsVisited
print("epoch train", epoch, env.rootURL, arrRL[-1], totReward, totDiscountedReward, lastLangVisited)
if epoch % params.updateFrequency == 0:
SavePlot(params, env, params.saveDir, epoch, "train", arrRL, totReward, totDiscountedReward)
TIMER.Start("CalcGrads")
qn.CalcGrads(sess, corpus)
TIMER.Pause("CalcGrads")
if epoch % params.updateFrequency == 0:
RunRLSavePlots(sess, qn, corpus, params, envsTest, params.saveDir, epoch, "test")
if epoch != 0:
print("UpdateGrads & Validating")
TIMER.Start("UpdateGrads")
qn.UpdateGrads(sess, corpus)
TIMER.Pause("UpdateGrads")
sys.stdout.flush()
def Train(params, sess, saver, qns, envs, envsTest):
print("Start training")
for epoch in range(params.max_epochs):
#print("epoch", epoch)
for env in envs:
TIMER.Start("Trajectory")
arrRL, totReward, totDiscountedReward = Trajectory(
env, params, sess, qns, False, 1)
TIMER.Pause("Trajectory")
print("epoch train", epoch, env.rootURL, totReward,
totDiscountedReward)
SavePlot(params, env, params.saveDirPlots, epoch, "train", arrRL,
totReward, totDiscountedReward)
TIMER.Start("Train")
qns.q[0].corpus.Train(sess, params)
qns.q[1].corpus.Train(sess, params)
TIMER.Pause("Train")
if epoch > 0 and epoch % params.walk == 0:
print("Validating")
#SavePlots(sess, qns, params, envs, params.saveDirPlots, epoch, "train")
RunRLSavePlots(sess, qns, params, envsTest, params.saveDirPlots,
epoch, "test")
def plot_results(sr, audio, onsets, savefile):
if audio.ndim > 1:
audio = np.mean(audio, axis=1)
t = np.linspace(0, len(audio) / sr, len(audio))
plt.plot(t, audio, color='dodgerblue')
if len(onsets) > 0:
for o in onsets:
plt.axvline(o / sr, color='red', linestyle='--')
plt.xlabel('Time (s)')
plt.grid()
sp = SavePlot(True, savefile=savefile, auto_overwrite=True)
sp.plot(plt)
def plot_file_data(files, savefile=None, plotlabels=None):
""" Plot radar graph of Fmes, P and R from onset analysis file(s)
"""
data = []
measures = ['F-measure', 'Precision', 'Recall']
for file in files:
results = read_onset_analysis(file)
data.append([results[measure] for measure in measures])
if savefile is not None:
sp = SavePlot(True, savefile, auto_overwrite=True)
else:
sp = SavePlot(False, auto_overwrite=True)
make_radar_graph(data, sp, labels=plotlabels)
def RunRLSavePlot(sess, qn, corpus, params, env, saveDir, epoch, sset):
arrRL, totReward, totDiscountedReward = Trajectory(env, params, sess, qn, corpus, True)
SavePlot(params, env, saveDir, epoch, sset, arrRL, totReward, totDiscountedReward)
# 2.11864583, 2.21 , 2.29 , 2.37 ,
# 2.47])
#
# falsepos = np.array([0.59, 1.55, 2.55, 2.61, 2.69, 2.77])
#
# falseneg = None
savepath = os.path.join(user, 'onsets', 'hsj', 'Visualisation', 'Iowa')
saveroot, _ = os.path.splitext(file)
savefiledmp = '{}_{:.0e}'.format(saveroot, damping)
ext = '.pdf'
savefile = os.path.join(savepath, savefiledmp + ext)
savelegfile = os.path.join(savepath, savefiledmp + '_legend' + ext)
saveaudio = os.path.join(savepath, saveroot + ext)
spa = SavePlot(save=False, savefile=saveaudio)
sp = SavePlot(save=False, savefile=savefile)
sl = SaveLegend(savelegfile)
get_responses(
filepath,
damping=damping,
sp=sp,
tp=truepos,
fp=falsepos,
fn=falseneg,
# sl=sl,
# spa=spa,
# t1=3
)
def mix(f0, amp):
audio = np.zeros(sr)
for n in range(len(f0)):
t = np.linspace(0, 2 * np.pi * f0[n], sr)
audio += np.sin(t) * amp[n]
audio = np.append(audio, np.zeros(2 * sr))
return audio
sp = SavePlot(False)
method = DetectorBank.central_difference
f_norm = DetectorBank.freq_unnormalized
a_norm = DetectorBank.amp_unnormalized
if method == DetectorBank.runge_kutta:
mthd = 'rk'
elif method == DetectorBank.central_difference:
mthd = 'cd'
if f_norm == DetectorBank.freq_unnormalized:
nrml = 'un'
elif f_norm == DetectorBank.search_normalized:
nrml = 'sn'
def plot_segments(file, threshold=None, save=False, audio_file=None):
sns.set_style('whitegrid')
plot_onsets = True # False #
plot_ms = False #True #
savepath = '/home/keziah/onsets/hsj/Visualisation/Iowa/'
# if plot_onsets:
# figdir = '{}_octaves_with_onsets'.format('xylophone') #'piano'
# else:
# figdir = '{}_octaves_without_onsets'.format('xylophone') #'piano'
figdir = ''
figdir = os.path.join(savepath, figdir)
if not os.path.exists(figdir):
os.makedirs(figdir)
arr = np.loadtxt(file)
# arr = arr[:-1]
# plot first 500 segments
i0 = 0
i1 = len(arr) # 500 # 20 #
seg_size = 20e-3
title = file.split(os.path.sep)[-2]
sr, manual_onsets = get_manual_onsets(file)
sr, found_onsets = get_found_onsets(file)
found_onsets /= sr
if audio_file is not None:
sp = SavePlot(True,
os.path.join(savepath, title + '_waveform.pdf'),
auto_overwrite=True)
plot_audio(audio_file, sp) #, manual_onsets, found_onsets)
t = np.arange(0, len(arr), dtype=float)
t *= seg_size
if plot_ms:
t *= 1000
plt.plot(t[i0:i1], arr[i0:i1])
correct, diff = CheckOnsets.compare(manual_onsets, found_onsets, 50)
if plot_onsets:
if correct.size > 0:
for n, idx in enumerate(correct):
# correctly found onset
onset = found_onsets[idx]
# remove corresponding onset from manual_onsets
man = manual_onsets + (diff[n] / 1000)
w = np.where(np.isclose(man, onset))[0]
manual_onsets = np.delete(manual_onsets, w)
# plot onset
if onset <= i1 * seg_size:
if plot_ms:
onset *= 1000
plt.axvline(onset, color='lime', label='True positive')
# remove correct onsets from found
found_onsets = np.delete(found_onsets, correct)
# else:
# for onset in manual_onsets:
# plt.axvline(onset, color='red', linestyle='--')
# plot false negatives
for onset in manual_onsets:
if onset <= i1 * seg_size:
if plot_ms:
onset *= 1000
plt.axvline(onset,
color='indigo',
linestyle='--',
label='False negative')
# plot false positives
for onset in found_onsets:
if onset <= i1 * seg_size: # t[i1-1]:
if plot_ms:
onset *= 1000
plt.axvline(onset,
color='mediumorchid',
linestyle='--',
label='False positive')
if threshold is not None:
threshold = np.log(threshold)
plt.axhline(threshold, color='green', linestyle='--')
# plt.title(title)
plt.grid(True)
if plot_ms:
xlabel = 'Time (ms)'
else:
xlabel = 'Time (s)'
plt.xlabel(xlabel)
plt.ylabel('Mean log')
# plt.legend()
sp = SavePlot(
save,
os.path.join(figdir, title + '.pdf'), #+'_w_legend.pdf'),
auto_overwrite=True)
sp.plot(plt)
"""
Example file to show the use of SavePlot and SaveLegend.
"""
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from save_plot import SavePlot, SaveLegend
# make save_plot objects
sp = SavePlot(True, 'waves.pdf')
sl = SaveLegend('legend.pdf', auto_overwrite=True)
# parameters for sine waves
x = np.linspace(0, 2*np.pi, 1000)
freq = np.arange(1,5)
amp = [0.5, 1, 0.75, 0.25]
# colours, linestyles and labels
c = ['darkorange','dodgerblue', 'firebrick', 'green']
ls = ['-', '--', ':', '-.']
mk = ['o', None, None, None]
labels = list('{} rad/s'.format(f) for f in freq)
# using seaborn style
# feel free to omit this
sns.set_style('darkgrid')
# plot some waves
for idx, f in enumerate(freq):
plt.plot(x, amp[idx]*np.sin(f*x), color=c[idx], linestyle=ls[idx])
outpath = '/home/keziah/onsets/hsj/Sandpit/results/Iowa/'
# sys.stdout = open(os.path.join(outpath, 'compare_mirex_summary.txt'), 'w')
modes = ['Precision', 'Recall', 'F-measure']
which = modes[2]
tab_d = {}
df = getAllResults(which)
for idx in df.index:
name = getName(idx)
savefile = '{}_{}.pdf'.format(name, which)
savefile = os.path.join(savepath, savefile)
sp = SavePlot(False, savefile, auto_overwrite=True, mode='quiet')
row = df.loc[idx]
plotResults(row, ylabel=which, sp=sp)
if name == 'percussion_single_note':
name = 'Percussion'
elif name == 'pizz':
name = 'Pizzicato strings'
elif name == 'arco':
name = 'Arco strings'
print(name.capitalize())
d = getStats(row)
print()
tab_d[name.capitalize()] = d
f = files[0]
user = os.path.expanduser('~')
root = os.path.join(user, 'Iowa', 'all')
file = os.path.join(root, f)
audio, sr = sf.read(file)
onset = 0.076
orig_found = [0.000, 0.216, 0.334, 0.487, 0.590, 0.788, 1.067, 2.690]
k = -9
freq = np.array([440 * 2**(i / 12) for i in [k]]) # [k-1, k, k+1]])
onsets = getOnsets(audio, sr, freq)
found = onsets[0] # onsets[1] #set(onsets[0] + onsets[1] + onsets[2])
found = sorted(list(found))
found = np.array(found)
found = found / sr
print(found)
savefile = 'trumpet_C4_right_cap.pdf'
savepath = '/home/keziah/onsets/hsj/Visualisation/onset_detection/'
sp = SavePlot(False, os.path.join(savepath, savefile), auto_overwrite=True)
# plotMean(audio, sr, freq[0], onset, [orig_found[1]], sp)
plotMean(audio, sr, freq[0], onset, found, sp)
if __name__ == '__main__':
user = os.path.expanduser('~')
audio_root_dir = os.path.join(user, 'Iowa', 'all')
percussion_dir = os.path.join(audio_root_dir, 'Percussion')
savepath = '.'
params = {
'bow': {
'file': 'Vibraphone.bow/Vibraphone.bow.A4.stereo.wav',
'savename': 'vibraphone_bow_400ms.pdf',
'onset': [0.174],
'found': [0.022]
},
# 'strike':
# {'file': 'Vibraphone.dampen/Vibraphone.dampen.ff.A4.stereo.wav',
# 'savename':'vibraphone_strike.pdf',
# 'onset':0.015,
# 'found':0.019}
}
for k, dct in params.items():
file = os.path.join(percussion_dir, dct['file'])
savefile = os.path.join(savepath, dct['savename'])
sp = SavePlot(False, savefile, auto_overwrite=True)
plot_audio(file,
sp,
manual_onsets=dct['onset'],
found_onsets=dct['found'])
if __name__ == '__main__':
method = DetectorBank.runge_kutta
f_norm = DetectorBank.freq_unnormalized
a_norm = DetectorBank.amp_unnormalized
sr = 48000
f0 = 440 # 100 #
gain = 1.35
d = [1e-4, 2e-4, 3e-4, 4e-4, 5e-4]
# make SavePlot objects before calling get_rise_relax_times()
sp0 = SavePlot(False)
sp1 = SavePlot(False)
sp = [sp0, sp1]
rstms, rxtms = get_rise_relax_times(d, f0, sr, method, f_norm, a_norm,
gain, True, sp)
if a_norm == 2**16:
s = 'AMP UNNORMALIZED'
elif a_norm == 2**17:
s = 'AMP NORMALIZED'
# print('Generated by ' + __file__ + '\n')
print(s + '\n' + '-' * len(s) + '\n')
