def printWSD2_scores(
wavF,
true_annF,
template_annF,
WSD2_clf,
WSD2_feExFun,
lt,
scoreClassLabels,
outF,
strini="",
strend="",
m="auto",
readSectionsWSD2="default", # for WSD2
labelsHierarchy="default",
): # for instantiating truth labels
if labelsHierarchy == "default":
labelsHierarchy = ["c"]
# load waveform
waveForm, fs = sT.wav2waveform(wavF)
tf = len(waveForm) / fs
if m == "auto":
m = int(len(waveForm) / 512)
# ground thruth
y_true_names = auf.annotationsFi2instances(true_annF,
m,
tf,
labelsHierarchy=labelsHierarchy)
# predictions
y_pred_names = pT.WSD2predict(wavF,
template_annF,
WSD2_feExFun,
lt,
WSD2_clf,
m,
tf,
readSections=readSectionsWSD2)
# print P, R, f1, support
print_precision_recall_fscore_support(y_true_names,
y_pred_names,
outF,
strini=strini,
strend=strend,
labels=scoreClassLabels)
return outF
def plotAnnotatedSpectro(wavFi, annFi, outDir, callAsTitle=True, figsize=None, dpi=None,
labelsHeight=10, cmapName='seismic', lw = 1, **kwargs):
'''
plots the spectrogram with it's annotations
Parameters
----------
wavAnnCollection : collection of paths to wavs and annotations files
outDir : dir where the plots will be saved
labelHeight : dictionary with the names of the labels and the height
dpi : size of the image
lw : line width for the annotated section
'''
## wav file
try:
waveForm, fs = sT.wav2waveform(wavFi)
except:
return False
M, ff, tf, _ = sT.spectralRep(waveForm, fs)#plt.specgram(waveForm, Fs=fs)
plt.ioff()
fig, ax = plt.subplots(figsize=figsize)
ax.imshow(M.T, aspect='auto', origin='bottom', #interpolation='nearest',
extent=[0, tf, 0, fs/2/1000.], **kwargs)
ax.set_xlabel('time (s)')
ax.set_ylabel('frequency (KHz)')
### read annotations
annD = annT.parseAupFile(annFi)
## custom ann clrs
idx2labs = list(set([item['label'] for item in annD]))
labs2idx = { idx2labs[i] : i for i in range(len(idx2labs)) }
cmap = plt.cm.get_cmap( cmapName, len(idx2labs))
annLabel=''
### annotate
for item in annD: # plot call segments - black line
ax.hlines(labelsHeight, float(item['startTime']), float(item['endTime']),
colors=cmap(labs2idx[item['label']]) , lw=lw)
annLabel += '{} '.format(item['label']) #read labels
if callAsTitle:
ax.set_title('{}'.format(annLabel))
#if callAsTitle: ax.set_title('{}'.format(fp.parseHeikesNameConv(wavFi)['call']))
outPl = os.path.join( outDir, os.path.basename(wavFi).replace('wav', 'png'))
plt.savefig(outPl, dpi=dpi)
del fig, ax
return outPl
def coll_clf_scores(clf,
wavAnnCollection,
featExtFun,
labelTransformer,
labelSet=None):
"""estimates the score of a classifiers, for each wav file in the collection
Parameters:
< clf : classifier
< wavAnnCollection : collection of wavs and annotations
list of tuples (<wav>, <txt>)
< featExtFun : function(wavs, annotLi) --> A0, a_names, _, _
import functools;
featExtFun = functools.partial(sT.waveform2featMatrix,
textWS=textWS, **featConstD)
< labelEncoder : label encoder of the features
< labelSet : list of labels to consider, if None => all labels are kept
"""
n = len(wavAnnCollection)
acc = np.zeros(n)
pre = np.zeros(n)
rec = np.zeros(n)
f1 = np.zeros(n)
sizes = np.zeros(n)
if labelSet is None:
labelSet = labelTransformer.num2nom(clf.classes_)
i = 0
for wavF, annF in wavAnnCollection[:]:
waveForm, fs = sT.wav2waveform(wavF) # read wavs
annotLi_t = sT.aupTxt2annTu(annF) # read annotations
A0, a_names, _, _ = featExtFun(waveForm, fs, annotations=annotLi_t)
mask = np.in1d(a_names, labelSet) # filer unwanted labels
a = labelTransformer.nom2num(
a_names[mask]) # convert labels to numeric
A = A0[mask]
y_pred = clf.predict(A)
## scores
acc[i] = mt.accuracy_score(a, y_pred)
pre[i] = mt.precision_score(a, y_pred)
rec[i] = mt.recall_score(a, y_pred)
f1[i] = mt.f1_score(a, y_pred)
sizes[i] = len(a)
i += 1
return acc, pre, rec, f1, sizes
