def create_model(argv, loader):
if argv.model == "nb":
model = NB_Model(argv)
elif argv.model == "dan":
model = DAN(argv)
elif argv.model == "blstm":
model = BLSTM(argv, loader.desired_len)
elif argv.model == "blstm_2dcnn":
model = BLSTM_2DCNN(argv, loader.desired_len)
return model
def create_model_hierarchy(cls,
bottom_item_list,
emb_wgts_bottom_items_dict,
layer_nums=3,
rnn_state_size=[],
bottom_emb_item_len=3,
flag_embedding_trainable=1,
seq_len=39,
batch_size=20,
mode_attention=1,
drop_out_r=0.,
att_layer_cnt=2,
bhDwellAtt=0,
rnn_type="WGRU",
RNN_norm="GRU",
flagCid3RNNNorm=False):
c_mask_value = 0.
att_zero_value = -2 ^ 31
def slice(x):
return x[:, -1, :]
flag_concate_sku_cid = True
RNN = rnn_type
MODE_BHDWELLATT = True if bhDwellAtt == 1 else False
ATT_NET_LAYER_CNT = att_layer_cnt
bottom_item_len = len(bottom_item_list)
input = [None] * bottom_item_len
word_num = [None] * bottom_item_len
emb_len = [None] * bottom_item_len
embedding_bottom_item = [None] * bottom_item_len
embed = [None] * bottom_item_len
layer_nums_max = 3
rnn_embed = [None] * layer_nums_max
rnn = [None] * layer_nums_max
rnn_output = [None] * layer_nums_max
flag_embedding_trainable = True if flag_embedding_trainable == 1 else False
##Embedding layer
# Embedding sku, bh, cid3, gap, dwell: 0, 1, 2, 3, 4
for i in range(bottom_item_len):
bottom_item = bottom_item_list[i]
###input
input[i] = Input(batch_shape=(
batch_size,
seq_len,
),
dtype='int32')
###Embedding
# load embedding weights
# emb_wgts[i] = np.loadtxt(init_wgts_file_emb[i])
word_num[i], emb_len[i] = emb_wgts_bottom_items_dict[
bottom_item].shape
print
word_num[i], emb_len[i]
# get embedding
cur_flag_embedding_trainable = flag_embedding_trainable
if (i == 0):
cur_flag_embedding_trainable = False
embedding_bottom_item[i] = Embedding(
word_num[i],
emb_len[i],
input_length=seq_len,
trainable=cur_flag_embedding_trainable)
embed[i] = embedding_bottom_item[i](input[i]) # drop_out=0.2
embedding_bottom_item[i].set_weights(
[emb_wgts_bottom_items_dict[bottom_item]])
# cal mask
mask_sku = np.zeros((batch_size, seq_len))
mask_cid3 = np.zeros((batch_size, seq_len))
for j in range(batch_size):
sku = input[0][j, :]
cid3 = input[2][j, :]
for k in range(seq_len - 1):
if (sku[k] == 0 or sku[k] == sku[k + 1]):
mask_sku[j][k] = 1
if (sku[k] == 0 or cid3[k] == cid3[k + 1]):
mask_cid3[j][k] = 1
# f mask
def f_mask_sku(x):
x_new = x
for j in range(batch_size):
for k in range(seq_len):
if (mask_sku[j][k] == 1):
x_new = T.set_subtensor(x_new[j, k, :], c_mask_value)
return x_new
def f_mask_cid3(x):
x_new = x
for j in range(batch_size):
for k in range(seq_len):
if (mask_cid3[j][k] == 1):
x_new = T.set_subtensor(x_new[j, k, :], c_mask_value)
return x_new
def f_mask_att_sku(x):
x_new = x
for j in range(batch_size):
for k in range(seq_len):
if (mask_sku[j][k] == 1):
x_new = T.set_subtensor(x_new[j, k], att_zero_value)
return x_new
def f_mask_att_cid3(x):
x_new = x
for j in range(batch_size):
for k in range(seq_len):
if (mask_cid3[j][k] == 1):
x_new = T.set_subtensor(x_new[j, k], att_zero_value)
return x_new
def K_dot(arr):
axes = [1, 1]
x, y = arr[0], arr[1]
return K.batch_dot(x, y, axes=axes)
def K_squeeze(x):
return K.squeeze(x, axis=-1)
Lambda_sequeeze = Lambda(lambda x: K_squeeze(x))
##RNN layer
if (RNN == "BLSTM"):
rnn[0] = BLSTM(rnn_state_size[0],
interval_dim=emb_len[3],
weight_dim=emb_len[1],
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_micro")
rnn[1] = BLSTM(rnn_state_size[1],
interval_dim=emb_len[3],
weight_dim=emb_len[4],
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_sku")
if (not flagCid3RNNNorm):
rnn[2] = BLSTM(rnn_state_size[2],
interval_dim=emb_len[3],
weight_dim=0,
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_cid3")
elif (RNN == "BLSTM2"):
rnn[0] = BLSTM2(rnn_state_size[0],
interval_dim=emb_len[3],
weight_dim=emb_len[1],
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_micro")
rnn[1] = BLSTM2(rnn_state_size[1],
interval_dim=emb_len[3],
weight_dim=emb_len[4],
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_sku")
if (not flagCid3RNNNorm):
rnn[2] = BLSTM2(rnn_state_size[2],
interval_dim=emb_len[3],
weight_dim=0,
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_cid3")
elif (RNN == "TimeLSTM"):
rnn[0] = BLSTM(rnn_state_size[0],
interval_dim=emb_len[3],
weight_dim=0,
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_micro")
rnn[1] = BLSTM(rnn_state_size[1],
interval_dim=emb_len[3],
weight_dim=0,
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_sku")
if (not flagCid3RNNNorm):
rnn[2] = BLSTM(rnn_state_size[2],
interval_dim=emb_len[3],
weight_dim=0,
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_cid3")
elif (RNN == "WGRU"):
rnn[0] = WGRU(rnn_state_size[0],
weight_dim=emb_len[1],
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_micro")
rnn[1] = WGRU(rnn_state_size[1],
weight_dim=emb_len[3],
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_sku")
if (not flagCid3RNNNorm):
rnn[2] = WGRU(rnn_state_size[2],
weight_dim=emb_len[3],
tateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_cid3")
elif (RNN == "LSTM" or RNN == "GRU"):
RNN = LSTM if RNN == "LSTM" else GRU
rnn[0] = RNN(rnn_state_size[0],
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_micro")
rnn[1] = RNN(rnn_state_size[1],
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_sku")
else:
print
"%s is not valid RNN!" % RNN
if (RNN_norm == "LSTM"):
rnn_cid3 = LSTM
else:
rnn_cid3 = GRU
if (flagCid3RNNNorm):
rnn[2] = rnn_cid3(rnn_state_size[2],
stateful=False,
return_sequences=True,
dropout=drop_out_r,
name="rnn_out_cid3")
# rnn embed 0
if (bottom_emb_item_len == 5):
rnn_embed[0] = Concatenate(axis=-1)(
[embed[0], embed[1], embed[2], embed[3], embed[4]])
elif (bottom_emb_item_len == 4):
rnn_embed[0] = Concatenate(axis=-1)(
[embed[0], embed[1], embed[2], embed[3]])
elif (bottom_emb_item_len == 3):
rnn_embed[0] = Concatenate(axis=-1)([embed[0], embed[1], embed[2]])
elif (bottom_emb_item_len == 1):
rnn_embed[0] = embed[0]
elif (bottom_emb_item_len == 2):
rnn_embed[0] = Concatenate(axis=-1)([embed[0], embed[2]])
else:
rnn_embed[0] = Concatenate(axis=-1)(
[embed[0], embed[1], embed[2], embed[3]])
# add interval, wei
if (RNN == "WGRU"):
rnn_embed[0] = Concatenate(axis=-1)([rnn_embed[0], embed[1]])
if (RNN == "BLSTM" or RNN == "BLSTM2"):
rnn_embed[0] = Concatenate(axis=-1)(
[rnn_embed[0], embed[3], embed[1]])
if (RNN == "TimeLSTM"):
rnn_embed[0] = Concatenate(axis=-1)([rnn_embed[0], embed[3]])
# rnn micro
rnn_output[0] = rnn[0](rnn_embed[0])
# rnn sku
if (flag_concate_sku_cid):
rnn_embed[1] = Concatenate(axis=-1)([embed[0], rnn_output[0]])
else:
rnn_embed[1] = rnn_output[0]
# mask sku
# rnn embed 1
# rnn_embed[1] = Lambda(f_mask_sku, output_shape=(seq_len, rnn_state_size[1]))(rnn_embed[1])
if (RNN == "WGRU"):
rnn_embed[1] = Concatenate(axis=-1)([rnn_embed[1], embed[4]])
if (RNN == "BLSTM" or RNN == "BLSTM2"):
rnn_embed[1] = Concatenate(axis=-1)(
[rnn_embed[1], embed[3], embed[4]])
if (RNN == "TimeLSTM"):
rnn_embed[1] = Concatenate(axis=-1)([rnn_embed[1], embed[3]])
rnn_embed[1] = Lambda(f_mask_sku)(rnn_embed[1])
rnn_embed[1] = Masking(mask_value=c_mask_value)(rnn_embed[1])
rnn_output[1] = rnn[1](rnn_embed[1])
# rnn cid3
if (flag_concate_sku_cid):
rnn_embed[2] = Concatenate()([embed[2], rnn_output[1]])
else:
rnn_embed[2] = rnn_output[1]
if (not flagCid3RNNNorm):
rnn_embed[2] = Concatenate(axis=-1)([rnn_embed[2], embed[3]])
# mask cid3
# rnn_embed[2] = Lambda(f_mask_cid3, output_shape=(seq_len, rnn_state_size[2]))(rnn_embed[2])
rnn_embed[2] = Lambda(f_mask_cid3)(rnn_embed[2])
rnn_embed[2] = Masking(mask_value=c_mask_value)(rnn_embed[2])
rnn_output[2] = rnn[2](rnn_embed[2])
# rnn final output
rnn_out_final = rnn_output[layer_nums - 1]
rnn_out_micro = rnn_output[0]
rnn_out_sku = rnn_output[1]
rnn_out_cid3 = rnn_output[2]
# predict sku, cid3
if (mode_attention == 0):
# micro
att_out_micro = Lambda(
slice, output_shape=(rnn_state_size[0], ))(rnn_out_micro)
# trans to sku emb len
out_micro_sku_emb = Dense(emb_len[0],
activation="tanh")(att_out_micro)
out_micro = out_micro_sku_emb
# sku
att_out_sku = Lambda(
slice, output_shape=(rnn_state_size[1], ))(rnn_out_sku)
# trans to sku emb len
out_sku_emb = Dense(emb_len[0], activation="tanh")(att_out_sku)
out_sku = out_sku_emb
# cid3
att_out_cid3 = Lambda(
slice, output_shape=(rnn_state_size[2], ))(rnn_out_cid3)
out_cid3_emb = Dense(emb_len[2], activation="tanh")(att_out_cid3)
out_cid3 = out_cid3_emb
# out_cid3 = Dense(word_num[2], activation="softmax")(out_cid3_emb)
if (mode_attention == 2):
# atten micro
m_h = rnn_out_micro
m_h_last = Lambda(slice,
output_shape=(rnn_state_size[0], ),
name="rnn_out_micro_last")(m_h)
m_h_r = RepeatVector(seq_len)(m_h_last)
if (MODE_BHDWELLATT):
m_h_c = Concatenate(axis=-1)([m_h, m_h_r, embed[1]])
else:
m_h_c = Concatenate(axis=-1)([m_h, m_h_r])
if (ATT_NET_LAYER_CNT == 2):
m_h_a_1 = TimeDistributed(
Dense(ATT_NET_HIDSIZE, activation='tanh'))(m_h_c)
m_h_a = TimeDistributed(Dense(1, activation='tanh'))(m_h_a_1)
else:
m_h_a = TimeDistributed(Dense(1, activation='tanh'))(m_h_c)
m_h_a = Lambda(lambda x: x, output_shape=lambda s: s)(m_h_a)
m_att = Flatten()(m_h_a)
m_att_micro = Softmax(name="att_micro")(m_att)
m_att_out = Lambda(K_dot,
output_shape=(rnn_state_size[0], ),
name="out_micro_pre")([m_h, m_att_micro])
# trans to sku emb len
out_micro = Dense(emb_len[0], activation="tanh")(m_att_out)
# attenion sku
s_h = rnn_out_sku
s_h_last = Lambda(slice,
output_shape=(rnn_state_size[1], ),
name="rnn_out_sku_last")(s_h)
s_h_r = RepeatVector(seq_len)(s_h_last)
if (MODE_BHDWELLATT):
s_h_c = Concatenate(axis=-1)([s_h, s_h_r, embed[4]])
else:
s_h_c = Concatenate(axis=-1)([s_h, s_h_r])
if (ATT_NET_LAYER_CNT == 2):
s_h_a_1 = TimeDistributed(
Dense(ATT_NET_HIDSIZE, activation='tanh'))(s_h_c)
s_h_a = TimeDistributed(Dense(1, activation='tanh'))(s_h_a_1)
else:
s_h_a = TimeDistributed(Dense(1, activation='tanh'))(s_h_c)
s_h_a = Lambda(lambda x: x, output_shape=lambda s: s)(s_h_a)
s_att = Flatten()(s_h_a)
s_att = Lambda(f_mask_att_sku)(s_att)
s_att_sku = Softmax(axis=-1, name="att_sku")(s_att)
s_att_out = Lambda(K_dot,
output_shape=(rnn_state_size[1], ),
name="out_sku_pre")([s_h, s_att_sku])
# attention cid3
c_h = rnn_out_cid3
c_h_last = Lambda(slice,
output_shape=(rnn_state_size[2], ),
name="rnn_out_cid3_last")(c_h)
c_h_r = RepeatVector(seq_len)(c_h_last)
c_h_c = Concatenate(axis=-1)([c_h, c_h_r])
if (ATT_NET_LAYER_CNT == 2):
c_h_a_1 = TimeDistributed(
Dense(ATT_NET_HIDSIZE, activation='tanh'))(c_h_c)
c_h_a = TimeDistributed(Dense(1, activation='tanh'))(c_h_a_1)
else:
c_h_a = TimeDistributed(Dense(1, activation='tanh'))(c_h_c)
c_h_a = Lambda(lambda x: x, output_shape=lambda s: s)(c_h_a)
c_att = Flatten()(c_h_a)
c_att = Lambda(f_mask_att_cid3)(c_att)
c_att_cid3 = Softmax(axis=-1, name="att_cid3")(c_att)
c_att_out = Lambda(K_dot,
output_shape=(rnn_state_size[2], ),
name="out_cid3_pre")([c_h, c_att_cid3])
out_cid3 = Dense(emb_len[2], activation="tanh")(c_att_out)
out_sku = Dense(emb_len[0], activation="tanh")(s_att_out)
# model
model = Model(
inputs=[input[0], input[1], input[2], input[3], input[4]],
outputs=[out_micro, out_sku, out_cid3])
# return embedding, rnn, ret_with_target, input, out
return model
#index2answer = {1: '采用“流量模块+语音模块+短信模块”自由定制形式，每个模块提供阶梯单价，用户自由定制，任意组合，每月承诺消费不低于19元。', 2: '个人定制套餐为4G套餐，默认开通4G功能。', 3: '若您为个人定制套餐用户，您在手机端进入手机版电信营业厅，完成登录后，点击底部导航“服务”，切换到“办理”页面中进行办理。', 4: '很抱歉，目前只支持新入网用户办理此套餐。', 5: '您可以到本地营业厅办理销户。', 6: '号卡激活后当月月基本功能费和套餐内容均按天折算计扣，按照过渡期资费处理。', 7: '自激活日起，根据订购套餐月基本费按天折算计扣（入网当日到月底），四舍五入到分，套餐内容（如手机上网流量、通话时长、短信条数等）按天数折算，向上取整，业务转换、业务转赠功能次月开放。', 8: '套餐外资费：国内流量0.0001元/KB，国内通话0.15元/分钟，国内短信0.1元/条，其他按标准资费收取。', 9: '具体可变更套餐规则请咨询当地营业厅。', 10: '个人定制套餐暂不支持副卡办理。', 11: ' 个人定制套餐暂不支持办理流量包、语音包、短信包业务。', 12: '个人定制套餐暂不支持流量包业务。', 13: '个人定制套餐暂不支持语音包业务。', 14: '个人定制套餐暂不支持短信包业务。', 15: '转换业务是指仅套餐内订购三种业务量（含套餐内被赠的业务量）可在当月内按照转换规则进行自由互转，套餐外优惠叠加及其他活动转赠或者充值的业务量（例如流量卡充值流量）不在转换范围内，例如：使用剩余语音业务量可按照转换规则转换为流量。转换规则为1分钟语音=2MB流量=2条短信。每月可用于转换的语音上限值为1000分钟（加和值，即多次使用语音进行转换的语音总量不超过1000分钟），可用于转换的流量上限值为1000MB（加和值），可用于转换的短信上限值为100条（加和值）每月最多转换3次，每次转换1种业务。个人定制当月转赠、转换及套餐内剩余流量均可保留到次月，但这些流量在次月不可被再次转赠和转换。套餐业务转换在您账户正常状态下可使用，如您账户存在欠费、停机、挂失等问题，则无法使用该项业务。', 16: '转赠业务是指仅套餐内订购流量的剩余可用量（含套餐内转换业务量，不能二次转赠）可在当月内向同时正在使用本套餐本省的其他用户进行转赠，套餐外优惠叠加或充值的业务量（例如流量卡充值流量）不在转赠范围内。每月可用于转赠的流量上限值为1000MB（加和值，即多次使用流量进行转赠的总量不超过1000MB）每月最多转赠3次，单次转赠1人。每月获赠不受次数限制。个人定制当月转赠、转换及套餐内剩余流量均可保留到次月，但这些流量在次月不可被再次转赠和转换。套餐业务转换在您账户正常状态下可使用，如您账户存在欠费、停机、挂失等问题，则无法使用该项业务。', 17: '您在手机端进入手机版电信营业厅，完成登录后，点击底部导航“服务”，切换到“办理”页面即可进行“套餐变更”办理。', 18: '您在手机端进入手机版电信营业厅，完成登录后，点击底部导航“服务”，切换到“办理”页面即可进行“套餐转换”办理。', 19: '您在手机端进入手机版电信营业厅，完成登录后，点击底部导航“服务”，\t切换到“办理”页面即可进行“套餐转赠”办理。', 20: '您在手机端进入手机版电信营业厅，完成登录后，点击底部导航“服务”，切换到“查询”页面即可查询套餐变更记录。', 21: '您好，有的。当您在网上营业厅使用转赠功能时，可勾选短信提醒被赠人，并写下您对被赠人的留言。当您在手机版电信营业厅使用转赠功能时，会默认给被赠人发送短信提醒，您还可以写下您给被赠人的留言。', 22: '您好，个人定制套餐不激活保留时间因全国各省规则不同，请您咨询本省10000号或到本地营业厅进行咨询。', 23: '1.登录网上营业厅。2.登录手机营业厅客户端。3.联系当地人工客服或者前往营业厅进行查询。', 24: '您可登录网上营业厅http://www.189.cn/ 首页点击费用＞我的话费＞余额查询，即可查询可用余额情况。'}
# 读取词向量
embedding_matrix = get_embedding(emb_path, 'word', 400)
print("building model")
config = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=config) as sess:
with tf.device(gpu_config):
initializer = tf.random_uniform_initializer(-0.02, 0.02)
with tf.variable_scope("model", reuse=None, initializer=initializer):
# 模型实例
model = BLSTM(400,
20,
embedding_matrix,
attention=True,
num_epochs=100,
dropout=0.3,
is_training=False)
saver = tf.train.Saver()
# 读取训练好的模型参数
saver.restore(sess, 'models/model')
# Demo主体部分
# 预测，string为要预测的字符串
while True:
string = str(input("Please input the string:"))
# 输出预测结果
label, prob = model.predict_label(sess, string)
if prob[0][label] > 0.4:
id2word = pickle.load(open('id2word', 'rb'))
X_train = np.array([[word2id[word] for word in line] for line in X_train])
# 正在读取词向量
embedding_matrix = get_embedding(emb_path, 'word', 400)
print("building model")
config = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=config) as sess:
with tf.device(gpu_config):
initializer = tf.random_uniform_initializer(-0.02, 0.02)
with tf.variable_scope("model", reuse=None, initializer=initializer):
model = BLSTM(400,
20,
embedding_matrix,
attention=True,
num_epochs=100,
dropout=0.3)
print("training model")
tf.global_variables_initializer().run()
# 训练的迭代次数
num = 10
model.train(sess,
model_path,
X_train,
y_train,
X_train,
y_train,
0,
num_epochs=num)
print("success!")