["minnan", "nanchang", "kejia", "changsha", "shanghai", "hebei"])
sentences = []
with open("./label_dev_list_fb.txt", "r") as s:
for line in s.readlines():
sentences.append(line.strip().split("/")[-1].split()[0].replace(
"fb", "pcm"))
sentences = np.array(sentences)
#print len(sentences)
## ======================================
dev_dataset = TorchDataSet(dev_list, batch_size, chunk_num, dimension)
logging.info('finish reading all train data')
train_module = LanNet(input_dim=dimension,
hidden_dim=128,
bn_dim=30,
output_dim=language_nums)
logging.info(train_module)
train_module.load_state_dict(
torch.load('/inference/models/model9.model',
map_location=lambda storage, loc: storage))
train_module.eval()
epoch_tic = time.time()
dev_loss = 0.
dev_acc = 0.
dev_batch_num = 0
ACC = 0
dev_size = 0
start = 0
size = 0
language_nums = 10
learning_rate = 0.1
batch_size = 20
chunk_num = 10
train_iteration = 10
display_fre = 50
half = 4
## ======================================
train_dataset = TorchDataSet(train_list, batch_size, chunk_num, dimension)
dev_dataset = TorchDataSet(dev_list, batch_size, chunk_num, dimension)
logging.info('finish reading all train data')
# 优化器，SGD更新梯度
train_module = LanNet(input_dim=dimension,
hidden_dim=512,
bn_dim=30,
output_dim=language_nums)
logging.info(train_module)
optimizer = torch.optim.SGD(train_module.parameters(),
lr=learning_rate,
momentum=0.9)
# 将模型放入GPU中
if use_cuda:
train_module = train_module.to(device)
for epoch in range(train_iteration):
if epoch >= half:
learning_rate /= 2.
optimizer = torch.optim.SGD(train_module.parameters(),
lr=learning_rate,
learning_rate = 0.1
batch_size = 64
chunk_num = 10
train_iteration = 1
display_fre = 50
half = 4
half_1 = 7
epoch = 0
## ======================================
train_dataset = TorchDataSet(train_list, batch_size, chunk_num, dimension)
dev_dataset = TorchDataSet(dev_list, batch_size, chunk_num, dimension)
logging.info('finish reading all train data')
# 优化器，SGD更新梯度
train_module = LanNet(input_dim=dimension,
hidden_dim=128,
bn_dim=30,
output_dim=language_nums)
logging.info(train_module)
#optimizer = torch.optim.SGD(train_module.parameters(), lr=learning_rate, momentum=0.9)
train_module.load_state_dict(
torch.load('./inference/model9.model',
map_location=lambda storage, loc: storage))
#optimizer = torch.optim.RMSprop(train_module.parameters(), lr=learning_rate, alpha=0.9)
optimizer = torch.optim.SGD(train_module.parameters(),
lr=learning_rate,
momentum=0.9)
# 将模型放入GPU中
if use_cuda:
train_module = train_module.to(device)
train_module.eval()