def getTrainingFilenames(annotation_path, cv_filename):
"""
annotation filenames - cv test filenames
:param annotation_path:
:param cv_filename:
:return:
"""
annotation_fns = getRecordings(annotation_path)
test_fns = annotationCvParser(cv_filename)
train_fns = [x for x in annotation_fns if x not in test_fns]
return train_fns
def dump_feature_onset_helper(audio_path, annotation_path, fn, channel):
audio_fn = join(audio_path, fn + '.flac')
annotation_fn = join(annotation_path, fn + '.onsets')
mfcc = getMFCCBands2DMadmom(audio_fn, fs, hopsize_t, channel)
print('Feature collecting ...', fn)
times_onset = annotationCvParser(annotation_fn)
times_onset = [float(to) for to in times_onset]
# syllable onset frames
frames_onset = np.array(np.around(np.array(times_onset) / hopsize_t), dtype=int)
# line start and end frames
frame_start = 0
frame_end = mfcc.shape[0] - 1
return mfcc, frames_onset, frame_start, frame_end
def schluter_eval_subroutine(nfolds, pp_threshold, obs_cal, len_seq,
architecture, bock_cv_path, bock_cnn_model_path,
bock_audio_path, bock_annotations_path,
bock_results_path, detection_results_path,
jingju_cnn_model_path, full_path_jingju_scaler):
for ii in range(nfolds):
# load scaler
if 'bidi_lstms' not in architecture: # not CRNN
# only for jingju + schulter datasets trained model
# scaler_name_0 = 'scaler_jan_madmom_simpleSampleWeighting_early_stopping_schluter_jingju_dataset_'
# + str(ii)+'.pickle.gz'
if 'pretrained' in architecture:
scaler_0 = pickle.load(open(full_path_jingju_scaler, 'rb'))
else:
if 'temporal' in architecture:
scaler_name_0 = 'scaler_bock_' + str(ii) + '.pickle.gz'
else:
scaler_name_0 = 'scaler_bock_temporal_' + str(
ii) + '.pickle.gz'
with gzip.open(join(bock_cnn_model_path, scaler_name_0),
'rb') as f:
scaler_0 = cPickle.load(f)
else: # CRNN
scaler_name_0 = 'scaler_bock_phrase.pkl'
scaler_0 = pickle.load(
open(join(bock_cnn_model_path, scaler_name_0), 'rb'))
# load model
if 'pretrained' in architecture:
model_name_0 = '5_layers_cnn0'
else:
model_name_0 = architecture + str(ii)
model_name_1 = ''
if obs_cal != 'tocal':
model_keras_cnn_0 = None
stateful = None
else:
if 'bidi_lstms' not in architecture:
if 'pretrained' in architecture:
model_keras_cnn_0 = load_model(
join(jingju_cnn_model_path, model_name_0 + '.h5'))
else:
model_keras_cnn_0 = load_model(
join(bock_cnn_model_path, model_name_0 + '.h5'))
print(model_keras_cnn_0.summary())
else:
from training_scripts.models_CRNN import jan_original
# initialize the model
stateful = False
bidi = True
input_shape = (1, len_seq, 1, 80, 15)
model_keras_cnn_0 = jan_original(filter_density=1,
dropout=0.5,
input_shape=input_shape,
batchNorm=False,
dense_activation='sigmoid',
channel=1,
stateful=stateful,
training=False,
bidi=bidi)
# load weights
model_keras_cnn_0.load_weights(
join(bock_cnn_model_path, model_name_0 + '.h5'))
# load cross validation filenames
test_cv_filename = join(bock_cv_path,
'8-fold_cv_random_' + str(ii) + '.fold')
test_filenames = annotationCvParser(test_cv_filename)
if 'pretrained' in architecture:
model_name_0 = architecture + str(ii)
# delete detection results path if it exists
detection_results_path_model = join(detection_results_path,
model_name_0)
if os.path.exists(detection_results_path_model) and os.path.isdir(
model_name_0):
shutil.rmtree(model_name_0)
for fn in test_filenames:
if 'bidi_lstms' not in architecture:
batch_process_onset_detection(
audio_path=bock_audio_path,
annotation_path=bock_annotations_path,
filename=fn,
scaler_0=scaler_0,
model_keras_cnn_0=model_keras_cnn_0,
model_name_0=model_name_0,
model_name_1=model_name_1,
pp_threshold=pp_threshold,
channel=1,
obs_cal=obs_cal,
architecture=architecture,
detection_results_path=detection_results_path)
else:
batch_process_onset_detection_phrase(
audio_path=bock_audio_path,
annotation_path=bock_annotations_path,
filename=fn,
scaler_0=scaler_0,
model_keras_cnn_0=model_keras_cnn_0,
model_name_0=model_name_0,
model_name_1=model_name_1,
pp_threshold=pp_threshold,
stateful=stateful,
obs_cal=obs_cal,
len_seq=len_seq,
detection_results_path=detection_results_path)
print('threshold', pp_threshold)
recall_precision_f1_fold, recall_precision_f1_overall = eval_bock(
architecture=architecture,
detection_results_path=detection_results_path,
bock_annotations_path=bock_annotations_path)
log_path = join(bock_results_path, varin['sample_weighting'],
architecture + '_' + 'threshold.txt')
# log_path = join(schluter_results_path, weighting, 'schluter_jingju_model_threshold.txt')
append_write = append_or_write(log_path)
write_results_2_txt_schluter(log_path, append_write, pp_threshold,
recall_precision_f1_overall)
return recall_precision_f1_fold, recall_precision_f1_overall