def load_nlg_model(self, model_path):
""" load the trained NLG model """
model_params = pickle.load(open(model_path))
hidden_size = model_params['model']['Wd'].shape[0]
output_size = model_params['model']['Wd'].shape[1]
if model_params['params']['model'] == 'lstm_tanh': # lstm_tanh
diaact_input_size = model_params['model']['Wah'].shape[0]
input_size = model_params['model']['WLSTM'].shape[
0] - hidden_size - 1
rnnmodel = lstm_decoder_tanh(diaact_input_size, input_size,
hidden_size, output_size)
rnnmodel.model = copy.deepcopy(
model_params['model']) # rnn模型的w、b等trainable参数
model_params['params']['beam_size'] = dialog_config.nlg_beam_size
self.model = rnnmodel
self.word_dict = copy.deepcopy(model_params['word_dict'])
self.template_word_dict = copy.deepcopy(
model_params['template_word_dict'])
self.slot_dict = copy.deepcopy(
model_params['slot_dict']) # 与 slot_set.txt 相同
self.act_dict = copy.deepcopy(
model_params['act_dict']) # 与 dia_acts.txt 相同
self.inverse_word_dict = {
self.template_word_dict[k]: k
for k in self.template_word_dict.keys()
}
self.params = copy.deepcopy(model_params['params']) # 模型超参数
def load_nlg_model(self, model_path):
""" load the trained NLG model """
with open(model_path, "rb") as f:
model_params = pickle.load(f, encoding="latin1")
# model_params = pickle.load(open(model_path, 'rb'))
# hidden_size:100 output_size:1047
hidden_size = model_params['model']['Wd'].shape[0]
output_size = model_params['model']['Wd'].shape[1]
if model_params['params']['model'] == 'lstm_tanh': # lstm_tanh
diaact_input_size = model_params['model']['Wah'].shape[0]
input_size = model_params['model']['WLSTM'].shape[
0] - hidden_size - 1
rnnmodel = lstm_decoder_tanh(diaact_input_size, input_size,
hidden_size, output_size)
rnnmodel.model = copy.deepcopy(model_params['model'])
model_params['params']['beam_size'] = dialog_config.nlg_beam_size
self.model = rnnmodel
self.word_dict = copy.deepcopy(model_params['word_dict'])
self.template_word_dict = copy.deepcopy(
model_params['template_word_dict'])
self.slot_dict = copy.deepcopy(model_params['slot_dict'])
self.act_dict = copy.deepcopy(model_params['act_dict'])
self.inverse_word_dict = {
self.template_word_dict[k]: k
for k in self.template_word_dict.keys()
}
self.params = copy.deepcopy(model_params['params'])
def load_nlg_model(self, model_path):
""" load the trained NLG model """
import sys
if 'win' in sys.platform:
model_params = pickle.load(open(model_path, 'r'))
else:
model_params = pickle.load(open(model_path, 'rb'))
hidden_size = model_params['model']['Wd'].shape[0]
output_size = model_params['model']['Wd'].shape[1]
if model_params['params']['model'] == 'lstm_tanh': # lstm_tanh
diaact_input_size = model_params['model']['Wah'].shape[0]
input_size = model_params['model']['WLSTM'].shape[0] - hidden_size - 1
rnnmodel = lstm_decoder_tanh(diaact_input_size, input_size, hidden_size, output_size)
rnnmodel.model = copy.deepcopy(model_params['model'])
model_params['params']['beam_size'] = dialog_config.nlg_beam_size
self.model = rnnmodel
self.word_dict = copy.deepcopy(model_params['word_dict'])
self.template_word_dict = copy.deepcopy(model_params['template_word_dict'])
self.slot_dict = copy.deepcopy(model_params['slot_dict'])
self.act_dict = copy.deepcopy(model_params['act_dict'])
self.inverse_word_dict = {self.template_word_dict[k]:k for k in self.template_word_dict.keys()}
self.params = copy.deepcopy(model_params['params'])
