te_lst=options.te_lst
pt_file=options.pt_file
class_dict_file=options.lab_dict
data_folder=options.data_folder+'/'
output_folder=options.output_folder
#[windowing]
fs=int(options.fs)
cw_len=int(options.cw_len)
cw_shift=int(options.cw_shift)
#[cnn]
cnn_N_filt=list(map(int, options.cnn_N_filt.split(',')))
cnn_len_filt=list(map(int, options.cnn_len_filt.split(',')))
cnn_max_pool_len=list(map(int, options.cnn_max_pool_len.split(',')))
cnn_use_laynorm_inp=str_to_bool(options.cnn_use_laynorm_inp)
cnn_use_batchnorm_inp=str_to_bool(options.cnn_use_batchnorm_inp)
cnn_use_laynorm=list(map(str_to_bool, options.cnn_use_laynorm.split(',')))
cnn_use_batchnorm=list(map(str_to_bool, options.cnn_use_batchnorm.split(',')))
cnn_act=list(map(str, options.cnn_act.split(',')))
cnn_drop=list(map(float, options.cnn_drop.split(',')))
#[dnn]
fc_lay=list(map(int, options.fc_lay.split(',')))
fc_drop=list(map(float, options.fc_drop.split(',')))
fc_use_laynorm_inp=str_to_bool(options.fc_use_laynorm_inp)
fc_use_batchnorm_inp=str_to_bool(options.fc_use_batchnorm_inp)
fc_use_batchnorm=list(map(str_to_bool, options.fc_use_batchnorm.split(',')))
fc_use_laynorm=list(map(str_to_bool, options.fc_use_laynorm.split(',')))
fc_act=list(map(str, options.fc_act.split(',')))
dev_tgt_dir=options.dev_tgt_dir
test_src_dir=options.test_src_dir
test_tgt_dir=options.test_tgt_dir
output_folder=options.output_folder
save_dir=options.save_dir
restore_file=options.restore_file
#[windowing]
fs=int(options.fs)
#[cnn]
wlen=int(options.wlen)
cnn_N_filt=list(map(int, options.cnn_N_filt.split(',')))
cnn_len_filt=list(map(int, options.cnn_len_filt.split(',')))
cnn_max_pool_len=list(map(int, options.cnn_max_pool_len.split(',')))
cnn_use_laynorm_inp=str_to_bool(options.cnn_use_laynorm_inp)
cnn_use_batchnorm_inp=str_to_bool(options.cnn_use_batchnorm_inp)
cnn_use_laynorm=list(map(str_to_bool, options.cnn_use_laynorm.split(',')))
cnn_use_batchnorm=list(map(str_to_bool, options.cnn_use_batchnorm.split(',')))
cnn_act=list(map(str, options.cnn_act.split(',')))
cnn_drop=list(map(float, options.cnn_drop.split(',')))
#[transformer]
tr_embed_dim=int(options.tr_embed_dim)
tr_max_positions=int(options.tr_max_positions)
tr_pos=options.tr_pos
tr_num_layers=int(options.tr_num_layers)
tr_num_heads=int(options.tr_num_heads)
tr_filter_size=int(options.tr_filter_size)
tr_hidden_size=int(options.tr_hidden_size)
tr_dropout=float(options.tr_dropout)
def get_cfg():
# Reading cfg file
options=read_conf()
#[data]
args = Namespace()
args.tr_lst=options.tr_lst
args.te_lst=options.te_lst
args.pt_file=options.pt_file if options.pt_file_reset=='' else options.pt_file_reset
args.class_dict_file=options.lab_dict
args.data_folder=(options.data_folder if options.data_folder_reset=='' else options.data_folder_reset)+'/'
args.output_folder=options.output_folder
#[windowing]
args.fs=int(options.fs)
args.cw_len=int(options.cw_len)
args.cw_shift=int(options.cw_shift)
#[cnn]
args.cnn_N_filt=list(map(int, options.cnn_N_filt.split(',')))
args.cnn_len_filt=list(map(int, options.cnn_len_filt.split(',')))
args.cnn_max_pool_len=list(map(int, options.cnn_max_pool_len.split(',')))
args.cnn_use_laynorm_inp=str_to_bool(options.cnn_use_laynorm_inp)
args.cnn_use_batchnorm_inp=str_to_bool(options.cnn_use_batchnorm_inp)
args.cnn_use_laynorm=list(map(str_to_bool, options.cnn_use_laynorm.split(',')))
args.cnn_use_batchnorm=list(map(str_to_bool, options.cnn_use_batchnorm.split(',')))
args.cnn_act=list(map(str, options.cnn_act.split(',')))
args.cnn_drop=list(map(float, options.cnn_drop.split(',')))
#[dnn]
args.fc_lay=list(map(int, options.fc_lay.split(',')))
args.fc_drop=list(map(float, options.fc_drop.split(',')))
args.fc_use_laynorm_inp=str_to_bool(options.fc_use_laynorm_inp)
args.fc_use_batchnorm_inp=str_to_bool(options.fc_use_batchnorm_inp)
args.fc_use_batchnorm=list(map(str_to_bool, options.fc_use_batchnorm.split(',')))
args.fc_use_laynorm=list(map(str_to_bool, options.fc_use_laynorm.split(',')))
args.fc_act=list(map(str, options.fc_act.split(',')))
#[class]
args.class_lay=list(map(int, options.class_lay.split(',')))
args.class_drop=list(map(float, options.class_drop.split(',')))
args.class_use_laynorm_inp=str_to_bool(options.class_use_laynorm_inp)
args.class_use_batchnorm_inp=str_to_bool(options.class_use_batchnorm_inp)
args.class_use_batchnorm=list(map(str_to_bool, options.class_use_batchnorm.split(',')))
args.class_use_laynorm=list(map(str_to_bool, options.class_use_laynorm.split(',')))
args.class_act=list(map(str, options.class_act.split(',')))
#[optimization]
args.lr=float(options.lr)
args.batch_size=int(options.batch_size)
args.N_epochs=int(options.N_epochs)
args.N_batches=int(options.N_batches)
args.N_eval_epoch=int(options.N_eval_epoch)
args.seed=int(options.seed)
# training list
args.wav_lst_tr=ReadList(args.tr_lst)
args.snt_tr=len(args.wav_lst_tr)
# test list
args.wav_lst_te=ReadList(args.te_lst)
args.snt_te=len(args.wav_lst_te)
args.eval = options.eval
# Folder creation
try:
os.stat(args.output_folder)
except:
os.mkdir(args.output_folder)
return args
def __init__(self, weights_file=None):
self.device = 'cpu' # torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Reading cfg file
options = read_conf()
#[data]
tr_lst = options.tr_lst
te_lst = options.te_lst
pt_file = options.pt_file
class_dict_file = options.lab_dict
data_folder = options.data_folder + '/'
output_folder = options.output_folder
#[windowing]
fs = int(options.fs)
cw_len = int(options.cw_len)
cw_shift = int(options.cw_shift)
#[cnn]
cnn_N_filt = list(map(int, options.cnn_N_filt.split(',')))
cnn_len_filt = list(map(int, options.cnn_len_filt.split(',')))
cnn_max_pool_len = list(map(int, options.cnn_max_pool_len.split(',')))
cnn_use_laynorm_inp = str_to_bool(options.cnn_use_laynorm_inp)
cnn_use_batchnorm_inp = str_to_bool(options.cnn_use_batchnorm_inp)
cnn_use_laynorm = list(
map(str_to_bool, options.cnn_use_laynorm.split(',')))
cnn_use_batchnorm = list(
map(str_to_bool, options.cnn_use_batchnorm.split(',')))
cnn_act = list(map(str, options.cnn_act.split(',')))
cnn_drop = list(map(float, options.cnn_drop.split(',')))
#[dnn]
fc_lay = list(map(int, options.fc_lay.split(',')))
fc_drop = list(map(float, options.fc_drop.split(',')))
fc_use_laynorm_inp = str_to_bool(options.fc_use_laynorm_inp)
fc_use_batchnorm_inp = str_to_bool(options.fc_use_batchnorm_inp)
fc_use_batchnorm = list(
map(str_to_bool, options.fc_use_batchnorm.split(',')))
fc_use_laynorm = list(
map(str_to_bool, options.fc_use_laynorm.split(',')))
fc_act = list(map(str, options.fc_act.split(',')))
#[class]
self.class_lay = list(map(int, options.class_lay.split(',')))
class_drop = list(map(float, options.class_drop.split(',')))
class_use_laynorm_inp = str_to_bool(options.class_use_laynorm_inp)
class_use_batchnorm_inp = str_to_bool(options.class_use_batchnorm_inp)
class_use_batchnorm = list(
map(str_to_bool, options.class_use_batchnorm.split(',')))
class_use_laynorm = list(
map(str_to_bool, options.class_use_laynorm.split(',')))
class_act = list(map(str, options.class_act.split(',')))
#[optimization]
lr = float(options.lr)
batch_size = int(options.batch_size)
N_epochs = int(options.N_epochs)
N_batches = int(options.N_batches)
N_eval_epoch = int(options.N_eval_epoch)
seed = int(options.seed)
# training list
wav_lst_tr = ReadList(tr_lst)
snt_tr = len(wav_lst_tr)
# test list
wav_lst_te = ReadList(te_lst)
snt_te = len(wav_lst_te)
# setting seed
torch.manual_seed(seed)
np.random.seed(seed)
# Converting context and shift in samples
self.wlen = int(fs * cw_len / 1000.00)
self.wshift = int(fs * cw_shift / 1000.00)
# Batch_dev
self.Batch_dev = 128
# Feature extractor CNN
CNN_arch = {
'input_dim': self.wlen,
'fs': fs,
'cnn_N_filt': cnn_N_filt,
'cnn_len_filt': cnn_len_filt,
'cnn_max_pool_len': cnn_max_pool_len,
'cnn_use_laynorm_inp': cnn_use_laynorm_inp,
'cnn_use_batchnorm_inp': cnn_use_batchnorm_inp,
'cnn_use_laynorm': cnn_use_laynorm,
'cnn_use_batchnorm': cnn_use_batchnorm,
'cnn_act': cnn_act,
'cnn_drop': cnn_drop,
}
self.CNN_net = CNN(CNN_arch)
self.CNN_net.to(self.device)
# Loading label dictionary
lab_dict = np.load(class_dict_file).item()
DNN1_arch = {
'input_dim': self.CNN_net.out_dim,
'fc_lay': fc_lay,
'fc_drop': fc_drop,
'fc_use_batchnorm': fc_use_batchnorm,
'fc_use_laynorm': fc_use_laynorm,
'fc_use_laynorm_inp': fc_use_laynorm_inp,
'fc_use_batchnorm_inp': fc_use_batchnorm_inp,
'fc_act': fc_act,
}
self.DNN1_net = MLP(DNN1_arch)
#self.DNN1_net.cuda()
self.DNN1_net.to(self.device)
DNN2_arch = {
'input_dim': fc_lay[-1],
'fc_lay': self.class_lay,
'fc_drop': class_drop,
'fc_use_batchnorm': class_use_batchnorm,
'fc_use_laynorm': class_use_laynorm,
'fc_use_laynorm_inp': class_use_laynorm_inp,
'fc_use_batchnorm_inp': class_use_batchnorm_inp,
'fc_act': class_act,
}
self.DNN2_net = MLP(DNN2_arch)
#self.DNN2_net.cuda()
self.DNN2_net.to(self.device)
pre_train_file = None
if weights_file != 'none':
#let's load this otherwise load pt_file
pre_train_file = weights_file
elif pt_file != 'none':
pre_train_file = pt_file
if pre_train_file != 'none':
print("loading pre trained file", pre_train_file)
checkpoint_load = torch.load(pre_train_file, map_location=device)
self.CNN_net.load_state_dict(checkpoint_load['CNN_model_par'])
self.DNN1_net.load_state_dict(checkpoint_load['DNN1_model_par'])
self.DNN2_net.load_state_dict(checkpoint_load['DNN2_model_par'])
self.CNN_net.eval()
self.DNN1_net.eval()
self.DNN2_net.eval()
test_flag = 1
loss_sum = 0
err_sum = 0
err_sum_snt = 0
