def load_embeddings():
return {
'img_embeds_train': read_pickle(DATA_DIR + 'image_captioning/train_img_embeds.pickle'),
'img_filenames_train': read_pickle(DATA_DIR + 'image_captioning/train_img_fns.pickle'),
'img_embeds_val': read_pickle(DATA_DIR + 'image_captioning/val_img_embeds.pickle'),
'img_filenames_val': read_pickle(DATA_DIR + 'image_captioning/val_img_fns.pickle')
}
def load_data(_x, _y, load_mode):
x = None
if load_mode == 'txt':
x = np.loadtxt(_x)
elif load_mode == 'pickle':
x = util.read_pickle(_x)
y = pd.read_csv(_y, header=None)
y.columns = ['patient_id', 'diagnosis']
y.diagnosis = pd.Categorical(y.diagnosis)
y['diag_code'] = y.diagnosis.cat.codes
return x, y
def concatenate_list_of_deltas2(original, deltas1, deltas2):
list_conc = []
for item1, delta1, delta2 in zip(original, deltas1, deltas2):
conc = np.concatenate((item1, delta1, delta2))
list_conc.append(np.float32(conc))
return list_conc
if __name__ == '__main__':
# Loading fbanks
file_fbanks = 'C:/Users/Win10/PycharmProjects/the_speech/data/fbanks/fbanks_dem_40'
file_fbanks_bea = 'C:/Users/Win10/PycharmProjects/the_speech/data/fbanks/fbanks_ubm_dem_40'
list_fbanks = util.read_pickle(file_fbanks)
fbanks_bea = np.vstack(util.read_pickle(file_fbanks_bea))
# Output files
file_pca_fbanks = 'C:/Users/Win10/PycharmProjects/the_speech/data/fbanks/fbanks_dem_40_PCA'
file_pca_bea = 'C:/Users/Win10/PycharmProjects/the_speech/data/fbanks/fbanks_ubm_dem_40_PCA_SP'
# Scaling and selecting best number of components
bea_scaled, list_dem_scaled = scale_min_max(fbanks_bea, list_fbanks)
c = best_components(bea_scaled)
# Reducing dimensions PCA
fbanks_reduced, bea_reduced = run_pca(bea_scaled, list_dem_scaled, c)
# Computing deltas
# list_fbanks_deltas = compute_deltas(list_fbanks, 1)
#bea_deltas = python_speech_features.base.delta(feat=fbanks_bea, N=1)
def main():
work_dir = '/opt/project'
set_ = ''
set_models = ''
obs1 = 'aug_2del'
obs = '2del_aug-ubm'
obs_ivec = '2del_aug-ubm-augv1-mf'
num_mfccs = '20'
ivecs_dim = 256
# ---Input Files---
# MFCCs
file_mfccs_ivec = '/opt/project/data/hc/alzheimer/mfccs_dem_{}_{}'.format(
num_mfccs, obs1)
file_mfccs_ubm = '/opt/project/data/hc/alzheimer/mfccs_ubm_dem_{}_{}'.format(
num_mfccs, obs)
# Load MFCCs for UBM
mfccs_wav_ubm = np.vstack(util.read_pickle(file_mfccs_ubm))
# Load MFCCs for i-vectors extraction
list_mfccs_ivecs = util.read_pickle(file_mfccs_ivec)
# group per type (original, noised, stretched, pitched) corresponding to each spk.
# and join (concatenate) 3 wavs per speaker
# list_mfccs_joint = util.join_speakers_wavs(util.group_per_audio_type(list_mfccs_ivecs))
num_gauss = [2, 4, 8, 16, 32, 64, 128]
for g in num_gauss:
# ---OUTPUT FILES---
# i-vecs
ivector_2D_file = work_dir + '/data/ivecs/alzheimer/ivecs-' + str(
g) + '-{}mf-{}--{}i'.format(num_mfccs, obs_ivec, ivecs_dim)
# models for i-vecs
file_diag_ubm_model = work_dir + '/data/models/ivecs_alz/dubm_mdl_{}g_dem_{}'.format(
g, obs_ivec)
file_full_ubm_model = work_dir + '/data/models/ivecs_alz/fubm_mdl_{}g_dem_{}'.format(
g, obs_ivec)
file_ivec_extractor_model = work_dir + '/data/models/ivecs_alz/ivec_mdl_{}g_dem_{}'.format(
g, obs_ivec)
# Train models
model_dubm, model_fubm, model_ivector = train_models(
mfccs_wav_ubm, list_mfccs_ivecs, file_diag_ubm_model,
file_full_ubm_model, file_ivec_extractor_model, g, ivecs_dim)
# Extract ivectors
print("Extracting i-vecs...")
ivectors_list = []
n_gselect = int(np.log2(g))
print(n_gselect)
for i2 in list_mfccs_ivecs:
ivector_array = bob.kaldi.ivector_extract(i2,
model_fubm,
model_ivector,
num_gselect=n_gselect)
ivectors_list.append(ivector_array)
a_ivectors = np.vstack(ivectors_list)
# a_ivectors_3d = np.expand_dims(a_ivectors, axis=1)
print("i-vectors shape:", a_ivectors.shape)
# Save i-vectors to a txt file
np.savetxt(ivector_2D_file, a_ivectors)
print("i-vectors saved to:", ivector_2D_file)