dataset_list = sys.argv[5:]
dataset_num = len(dataset_list)
dis_phone_num = 4
# load data
for dataset_indx in range(dataset_num):
# intial the dist class for different training datasets
train_dist_path = os.path.join(dataset_list[dataset_indx],
"subivector_dist.ark")
dist_trans = dataset.Input_transform(train_dist_path, test_dist_path,
dis_phone_num, prob_trans_path)
ds_path = os.path.join(dataset_list[dataset_indx], "subivector.scp")
ds = dataset.SubivectorDataset(ds_path)
utt_num = ds.__len__()
print("The number of utters in %s is %d." % (ds_path, utt_num))
subivector_dim = np.size(ds.data_list[0], 1)
phone_num = np.size(ds.data_list[0], 0)
# read into training data loader
ds_loader = DataLoader(
ds,
batch_size=batchSize, # batch training
shuffle=False,
num_workers=int(1))
# set hook and store output in csvector
exit()
# (Hyper parameters)
batchSize = 128 # batchsize的大小
niter = 500 # epoch的最大值
# load data
train_path = sys.argv[
1] #'/scratch/njzheng/myprogram/phone_vectors5/subivector.scp'
sre_test_path = sys.argv[
2] #'/scratch/njzheng/myprogram/sre_test/subivector_test.scp'
sre_train_path = sys.argv[
3] #'/scratch/njzheng/myprogram/sre_train/subivector_test.scp'
sre_path = sys.argv[4] #'/scratch/njzheng/myprogram/sre/subivector_test.scp'
train_dataset = dataset.SubivectorDataset(train_path)
sre_test_dataset = dataset.SubivectorDataset(sre_test_path)
sre_train_dataset = dataset.SubivectorDataset(sre_train_path)
sre_dataset = dataset.SubivectorDataset(sre_path)
utt_num = train_dataset.__len__()
train_len = int(0.9 * utt_num)
valid_len = utt_num - train_len
train, valid = torch.utils.data.random_split(train_dataset,
lengths=[train_len, valid_len])
subivector_dim = np.size(train_dataset.data_list[0], 0)
phone_num = np.size(train_dataset.data_list[0], 1)
# load spk2int list
uttid2int_dic_path = sys.argv[
phone_num = 43
subnet_out_dim = 50
out_dim_sqrt = torch.tensor(subnet_out_dim).float().sqrt()
subnet_list = []
# Need loop ==============================================================
for phone_indx in range(34, phone_num + 1):
dnn_name = sys.argv[3] + "." + str(phone_indx) # $sub_train_path/net.pkl
# train_scp_path = os.path.join(train_path,"subivector.scp")
train_scp_path = os.path.join(train_path, "phone_vector",
"triplet_data." + str(phone_indx) + ".scp")
print('scp file is: %s' % (train_scp_path))
train_dataset = dataset.SubivectorDataset(train_scp_path)
subivector_dim = np.size(train_dataset.data_list[0], 1)
in_dim_sqrt = torch.tensor(subivector_dim).float().sqrt()
train_utt_num = train_dataset.__len__()
train_len = int(0.9 * train_utt_num)
valid_len = train_utt_num - train_len
train, valid = torch.utils.data.random_split(
train_dataset, lengths=[train_len, valid_len])
print('The number of training and valid utters are %d %d' %
(train_len, valid_len))
print('The subivector_dim is %d, phone_indx is %d' %
(subivector_dim, phone_indx))
