def __init__(self, config):
self.config = config
self.filenames = librosa.util.find_files(config['data']['sample_set_dir'])
self.set = {}
for file in self.filenames:
filename = os.path.basename(file)
if filename[-1] == '/':
filename = filename[0:-1]
audio, labels = pp.get_samples_and_labels(filename, config)
if config['data']['type'] == 'mel':
spec = torch.Tensor(audio_utils.wav2spectrogram(audio)).t()
melspec = torch.Tensor(audio_utils.wav2melspectrogram(audio)).t()
self.set[filename] = [melspec, labels, spec]
else:
audio = np.array(audio, dtype = np.float64)
f0, ap, sp, coded_sp = pw.cal_mcep(audio)
coded_sp = torch.Tensor(coded_sp.T)
self.set[filename] = [f0, ap, sp, coded_sp, labels]
def get_session_data(data_dir, exclude_unlabelled=True):
# print(data_dir)
os.chdir(data_dir)
filenames = []
specs = []
mels = []
labels = []
conts = []
conts_dis = []
speakers = []
all_labels = []
for foldername in os.listdir(data_dir):
if not (foldername == "Annotations" or foldername == ".DS_Store"):
for filename in os.listdir(data_dir + "/" + foldername):
if not filename == ".DS_Store":
wav, labels = get_wav_and_labels(filename, data_dir)
if not (exclude_unlabelled
and labels[0] == -1): #ignore some rare emotions
mel = audio_utils.wav2melspectrogram(wav)
spec = audio_utils.wav2spectrogram(wav)
filenames.append(filename)
mels.append(torch.Tensor(mel))
specs.append(torch.Tensor(spec))
# labels.append(label)
# conts.append(cont)
# conts_dis.append(cont_dis)
# speakers.append(speaker)
all_labels.append(labels)
print(foldername + " completed.")
return filenames, mels, specs, all_labels
def all_wavs_and_labels(exclude_unlabelled=True):
os.chdir(dataset_dir)
filenames = []
specs = []
mels = []
labels = []
conts = []
conts_dis = []
speakers = []
for foldername in os.listdir(dataset_dir):
if not (foldername == "Annotations" or foldername == ".DS_Store"
or foldername == "Processed"):
for filename in os.listdir(dataset_dir + "/" + foldername):
if not filename == ".DS_Store":
wav, label, cont, cont_dis, speaker = audio_utils.get_wav_and_labels(
filename)
if not (exclude_unlabelled
and label == -1): #ignore some rare emotions
spec = audio_utils.wav2spectrogram(wav)
mel = audio_utils.spectrogram2melspectrogram(spec)
filenames.append(filename)
mels.append(torch.Tensor(mel))
specs.append(torch.Tensor(spec))
labels.append(label)
conts.append(cont)
conts_dis.append(cont_dis)
speakers.append(speaker)
print(foldername + " done.")
return filenames, mels, specs, labels, conts, conts_dis, speakers
# # print(len(mels))
# save_data(filenames, dataset_dir + '/Processed_data/filenames' + ses_number)
# save_data(mels, dataset_dir + '/Processed_data/melspecs' + ses_number)
# save_data(specs, dataset_dir + '/Processed_data/specs' + ses_number)
# save_data(labels, dataset_dir + '/Processed_data/labels' + ses_number)
# print('Done ' + ses_number + ".")
#
# print(len(mels), len(labels), labels[0].size())
# print(concatenate_labels(1,2,[3,4,5],[6,7,8]))
wav, label = get_wav_and_labels("Ses02F_impro03_F002.wav",
dataset_dir + "/Session2")
# # wav2 = get_wav_and_labels("Ses01F_impro03_F002.wav", dataset_dir + "/Session1")[0]
#
spec = audio_utils.wav2spectrogram(wav)
# # spec2 = audio_utils.wav2spectrogram(wav2)
# print(spec.shape)
# # print(spec2.shape)
audio_utils.plot_spec(spec, type='log')
# #
melspec = audio_utils.spectrogram2melspectrogram(spec)
# print(melspec.shape)
audio_utils.plot_spec(melspec)
#
fn = str(label[0].item()) + 'to' + '2-' + str(label[1].item()) + '.png'
audio_utils.save_spec(melspec, "Test", fn)
# reproduced = audio_utils.spectrogram2wav(spec)
#
# audio_utils.save_wav(reproduced, "/Users/Max/MScProject/datasets/Produced/test1.wav")