with multiprocessing.Pool(nproc) as p:
p.map(reverb_mix_partial, range(nproc))
else:
# Easy case without multiprocessing
reverberate_and_mix(reverb_folder,
input_folder,
rir_folder,
mix_info,
part=0,
nparts=1,
chat=True)
# Check the data (fix sox bug)
examples_list = make_example_list(reverb_folder)
for example in examples_list:
check_and_correct_example(example,
reverb_folder,
check_length=True,
fix_length=True,
check_mix=True,
fix_mix=True,
sample_rate=sample_rate,
duration=clip_duration)
rev_subset_folder = osp.join(reverb_folder, subset)
post_process_txt_labels(rev_subset_folder,
output_folder=rev_subset_folder,
output_tsv=out_tsv,
rm_nOn_nOff=True)
print(time.time() - t)
event_time_std,
event_time_min,
event_time_max,
),
events_duration=(event_duration_dist, event_duration_min,
event_duration_max),
snrs=(snr_dist, snr_min, snr_max),
pitch_shifts=(pitch_dist, pitch_min, pitch_max),
time_stretches=(time_stretch_dist, time_stretch_min, time_stretch_max),
txt_file=True,
)
rm_high_polyphony(out_folder_500, 2)
out_tsv = osp.join(out_tsv_folder, "500ms.tsv")
post_process_txt_labels(out_folder_500,
output_folder=out_folder_500,
output_tsv=out_tsv)
# Generate 2 variants of this dataset
jams_to_modify = glob.glob(osp.join(out_folder_500, "*.jams"))
# Be careful, if changing the values of the added onset value,
# you maybe want to rerun the post_processing_annotations to be sure there is no inconsistency
# 5.5s onset files
out_folder_5500 = osp.join(base_out_folder, "5500ms")
add_onset = 5.0
modif_onset_5s = functools.partial(modify_fg_onset,
slice_seconds=add_onset)
modify_jams(jams_to_modify, modif_onset_5s, out_folder_5500)
# we also need to generate a new DataFrame with the right values
df = pd.read_csv(out_tsv, sep="\t")
pitch_max = 3.0
time_stretch_dist = 'uniform'
time_stretch_min = 1
time_stretch_max = 1
out_folder_ls_30 = osp.join(out_folder, "ls_30dB")
create_folder(out_folder_ls_30)
sg.generate(n_soundscapes, out_folder_ls_30, min_events, max_events, labels=('choose', []),
source_files=('choose', []), sources_time=(source_time_dist, source_time),
events_start=(evt_time_dist, evt_time_mean, evt_time_std, evt_time_min, evt_time_max),
events_duration=(event_duration_dist, event_duration_min, event_duration_max),
snrs=('const', 30), pitch_shifts=('uniform', -3.0, 3.0), time_stretches=('uniform', 1, 1),
txt_file=True)
rm_high_polyphony(out_folder_ls_30, 3)
post_process_txt_labels(out_folder_ls_30, output_folder=out_folder_ls_30, output_tsv=args.out_tsv,
background_label=True)
list_jams = glob.glob(osp.join(out_folder_ls_30, "*.jams"))
# We create the same dataset with different background SNR
# Be careful, 15 means the background SNR is 15,
# so the foreground background snr ratio is between -9dB and 15dB
out_folder_ls_15 = osp.join(out_folder, "ls_15dB")
modify_jams(list_jams, change_snr, out_folder_ls_15, db_change=-15)
# Same for 0dB FBSNR, from original fbsnr [6;30]dB, we go to [-24;0]dB FBSNR
out_folder_ls_0 = osp.join(out_folder, "ls_0dB")
modify_jams(list_jams, change_snr, out_folder_ls_15, db_change=-30)
# General output folder, in args
out_folder = args.out_folder
create_folder(out_folder)
create_folder(osp.dirname(args.out_tsv))
# Default parameters
clip_duration = 10.0
sg = SoundscapesGenerator(duration=clip_duration,
fg_folder=args.fg_folder,
bg_folder=args.bg_folder,
ref_db=cfg.ref_db,
samplerate=cfg.samplerate)
# Generate events same way as the training set
out_folder_30 = osp.join(out_folder, "30dB")
create_folder(out_folder_30)
param_json = args.json_params
with open(param_json) as json_file:
co_occur_dict = json.load(json_file)
sg.generate_by_label_occurence(label_occurences=co_occur_dict,
number=args.number,
out_folder=out_folder_30,
save_isolated_events=True)
# Remove files that have polyphony greater than 3
rm_high_polyphony(out_folder_30, 3)
# Combine labels of same events appending at the same time or less than 150ms interval
post_process_txt_labels(out_folder_30, output_folder=out_folder_30, output_tsv=args.out_tsv)
if __name__ == '__main__':
dataset_path = osp.join("..", "..", "dataset")
parser = argparse.ArgumentParser()
parser.add_argument('--in_dir',
type=str,
default=osp.join(dataset_path, "audio", "train",
"one_event_train"))
parser.add_argument('--out_dir',
type=str,
default=osp.join(dataset_path, "audio", "train",
"one_event_train_no_bg"))
parser.add_argument('--outtsv',
type=str,
default=osp.join(dataset_path, "metadata", "train",
"one_event_train_no_bg.tsv"))
parser.add_argument('--soundbank_dir', type=str, default=None)
args = parser.parse_args()
pformat(vars(args))
fg_path = None
bg_path = None
if args.soundbank_dir is not None:
fg_path = osp.join(args.soundbank_dir, "foreground")
bg_path = osp.join(args.soundbank_dir, "background")
generate_new_no_bg_files(args.in_dir, args.out_dir, fg_path, bg_path)
post_process_txt_labels(args.out_dir,
wavdir=args.in_dir,
output_folder=args.out_dir,
output_tsv=args.outtsv)
if random_state is None:
random_states = [None for i in range(nproc)]
else:
random_states = [random_state + i for i in range(nproc)]
print(random_states)
with multiprocessing.Pool(nproc) as p:
p.starmap(generate_multiproc,
zip(list_start, numbers, random_states))
# Single process
else:
sg = SoundscapesGenerator(duration=clip_duration,
fg_folder=fg_folder,
bg_folder=bg_folder,
ref_db=ref_db,
samplerate=sample_rate,
random_state=random_state)
sg.generate_by_label_occurence(label_occurences=co_occur_dict,
number=n_soundscapes,
out_folder=full_out_folder,
save_isolated_events=True,
pitch_shift=pitch_shift)
# ##
# Post processing
rm_high_polyphony(full_out_folder, max_polyphony=2)
# concat same labels overlapping
post_process_txt_labels(full_out_folder,
output_folder=full_out_folder,
output_tsv=out_tsv,
rm_nOn_nOff=True)
def test_postprocessing():
""" Test the preprocessing method
Returns:
Should output Fixed 3 problems
"""
folder = osp.join(absolute_dir_path, "material", "post_processing")
checked_folder = osp.join(absolute_dir_path, "generated",
"post_processing")
out_tsv = osp.join(absolute_dir_path, "generated", "post.tsv")
post_process_txt_labels(folder,
output_folder=checked_folder,
output_tsv=out_tsv)
df = pd.read_csv(out_tsv, sep="\t")
print(df.to_dict())
valid_df = pd.DataFrame({
"filename": {
0: "5.wav",
1: "5.wav",
2: "7.wav",
3: "7.wav",
4: "7.wav",
5: "7.wav",
6: "7.wav",
7: "7.wav",
},
"onset": {
0: 0.008,
1: 4.969,
2: 2.183,
3: 2.406,
4: 3.099,
5: 3.406,
6: 3.684,
7: 6.406,
},
"offset": {
0: 5.546,
1: 9.609,
2: 2.488,
3: 5.2,
4: 3.36,
5: 6.2,
6: 5.624,
7: 10.0,
},
"event_label": {
0: "Cat",
1: "Speech",
2: "Dishes",
3: "Speech",
4: "Dishes",
5: "Cat",
6: "Dishes",
7: "Frying",
},
})
check = df.round(3).sort_values("onset").reset_index(
drop=True) == valid_df.sort_values("onset").reset_index(drop=True)
assert check.all(axis=None), "Problem with post_processing_txt_annotations"