def hop(scope, last, sentence, sentence_bkg, bkg_iter, bkg_fix,
doc_len, real_max_len, convert_flag,
biases_initializer=tf.initializers.zeros(),
weights_initializer=tf.contrib.layers.xavier_initializer()):
if bkg_iter is None:
bkg_iter = []
if bkg_fix is None:
bkg_fix = []
if not isinstance(bkg_fix, Iterable):
bkg_fix = [bkg_fix]
bkg_fix = list(bkg_fix)
hidden_size = sentence_bkg.shape[2]
with tf.variable_scope(scope):
sentence = tf.stop_gradient(sentence) \
if not last else sentence
sentence_bkg = tf.stop_gradient(sentence_bkg) \
if not last else sentence_bkg
alphas = attention(sentence_bkg, [bkg_iter] + bkg_fix, doc_len, real_max_len,
biases_initializer=biases_initializer,
weights_initializer=weights_initializer)
new_bkg = tf.matmul(alphas, sentence)
new_bkg = tf.reshape(new_bkg, [-1, hidden_size], name='new_bkg')
if 'o' in convert_flag:
new_bkg = bkg_iter + new_bkg
return new_bkg
def __init__(self, embedding_count_dict, embedding_dim_dict, embedding_features_list, user_behavior_features,
activation='PReLU'):
super(DIN, self).__init__()
# Init Embedding Layer
self.embedding_count_dict = embedding_count_dict
self.embedding_dim_dict = embedding_dim_dict
self.embedding_layers = dict()
for feature in embedding_features_list:
self.embedding_layers[feature] = layers.Embedding(input_dim=embedding_count_dict[feature],
output_dim=embedding_dim_dict[feature],
embeddings_initializer='random_uniform')
# DIN Attention + Sum Pooling
self.hist_at = attention(alibaba_utils.get_input_dim(embedding_dim_dict, user_behavior_features))
# Init Fully Connection Layer
self.fc = tf.keras.Sequential()
self.fc.add(layers.BatchNormalization())
self.fc.add(layers.Dense(200, activation="relu"))
if activation == "Dice":
self.fc.add(Dice())
elif activation == "dice":
self.fc.add(dice(200))
elif activation == "PReLU":
self.fc.add(layers.PReLU(alpha_initializer='zeros', weights=None))
self.fc.add(layers.Dense(80, activation="relu"))
if activation == "Dice":
self.fc.add(Dice())
elif activation == "dice":
self.fc.add(dice(80))
elif activation == "PReLU":
self.fc.add(layers.PReLU(alpha_initializer='zeros', weights=None))
self.fc.add(layers.Dense(2, activation=None))
def attention_flow(self):
"""
Attention Flow Layer
contains Context-to-query Attention and Query-to-context Attention
:return:
"""
self.g = attention(self.hidden_size, self.h, self.u)
if self.use_dropout:
self.g = tf.nn.dropout(self.g, self.dropout_keep_prob)
def nsc(self, x, max_sen_len, max_doc_len, sen_len, doc_len):
def lstm(inputs, sequence_length, hidden_size, scope):
cell_fw = tf.nn.rnn_cell.LSTMCell(hidden_size // 2, forget_bias=0.,
initializer=xavier())
cell_bw = tf.nn.rnn_cell.LSTMCell(hidden_size // 2, forget_bias=0.,
initializer=xavier())
outputs, state = tf.nn.bidirectional_dynamic_rnn(
cell_fw=cell_fw, cell_bw=cell_bw, inputs=inputs,
sequence_length=sequence_length, dtype=tf.float32, scope=scope)
outputs = tf.concat(outputs, axis=2)
return outputs, state
with tf.variable_scope('sentence_layer'):
# lstm_outputs, _state = lstm(x, sen_len, self.hidden_size, 'lstm')
# lstm_outputs = tf.reshape(lstm_outputs, [-1, max_sen_len, self.hidden_size])
lstm_bkg, _state = lstm(x, sen_len, self.hidden_size, 'lstm_bkg')
lstm_bkg = tf.reshape(lstm_bkg, [-1, max_sen_len, self.hidden_size])
lstm_outputs = lstm_bkg
alphas = attention(lstm_bkg, [], sen_len, max_sen_len,
biases_initializer=self.biases_initializer,
weights_initializer=self.weights_initializer)
sen_bkg = tf.matmul(alphas, lstm_outputs)
sen_bkg = tf.reshape(sen_bkg, [-1, self.hidden_size], name='new_bkg')
outputs = tf.reshape(sen_bkg, [-1, max_doc_len, self.hidden_size])
with tf.variable_scope('document_layer'):
# lstm_outputs, _state = lstm(outputs, doc_len, self.hidden_size, 'lstm')
lstm_bkg, _state = lstm(outputs, doc_len, self.hidden_size, 'lstm_bkg')
lstm_outputs = lstm_bkg
alphas = attention(lstm_bkg, [], doc_len, max_doc_len,
biases_initializer=self.biases_initializer,
weights_initializer=self.weights_initializer)
doc_bkg = tf.matmul(alphas, lstm_outputs)
doc_bkg = tf.reshape(doc_bkg, [-1, self.hidden_size], name='new_bkg')
outputs = doc_bkg
with tf.variable_scope('result'):
d_hats = tf.layers.dense(tf.concat([outputs, self.usr, self.prd], axis=1), self.cls_cnt,
kernel_initializer=self.weights_initializer,
bias_initializer=self.biases_initializer)
return d_hats
def dnsc(self, x, max_sen_len, max_doc_len, sen_len, doc_len, task_label):
x = tf.reshape(x, [-1, max_sen_len, self.emb_dim])
sen_len = tf.reshape(sen_len, [-1])
def lstm(inputs, sequence_length, hidden_size, scope, init_state):
init_state_fw, init_state_bw = init_state
cell_fw = tf.nn.rnn_cell.LSTMCell(hidden_size // 2,
forget_bias=0.,
initializer=xavier())
cell_bw = tf.nn.rnn_cell.LSTMCell(hidden_size // 2,
forget_bias=0.,
initializer=xavier())
outputs, state = tf.nn.bidirectional_dynamic_rnn(
cell_fw=cell_fw,
cell_bw=cell_bw,
initial_state_fw=init_state_fw,
initial_state_bw=init_state_bw,
inputs=inputs,
sequence_length=sequence_length,
dtype=tf.float32,
scope=scope)
outputs = tf.concat(outputs, axis=2)
return outputs, state
with tf.variable_scope('sentence_layer'):
# lstm_outputs, _state = lstm(x, sen_len, self.hidden_size, 'lstm')
# lstm_outputs = tf.reshape(lstm_outputs, [-1, max_sen_len, self.hidden_size])
sen_task_label = tf.reshape(
tf.tile(task_label[:, None], [1, max_doc_len]), [-1])
sen_init_state = tf.get_variable(
'sen_init_state', [self.task_cnt, 2 * self.hidden_size])
sen_init_state = tf.nn.embedding_lookup(sen_init_state,
sen_task_label)
sen_init_state_fw = tf.nn.rnn_cell.LSTMStateTuple(
sen_init_state[:, :self.hidden_size // 2],
sen_init_state[:, self.hidden_size // 2:self.hidden_size])
sen_init_state_bw = tf.nn.rnn_cell.LSTMStateTuple(
sen_init_state[:, self.hidden_size:self.hidden_size * 3 // 2],
sen_init_state[:, self.hidden_size * 3 // 2:])
sen_init_state = (sen_init_state_fw, sen_init_state_bw)
lstm_bkg, _state = lstm(x, sen_len, self.hidden_size, 'lstm_bkg',
sen_init_state)
lstm_bkg = tf.reshape(lstm_bkg,
[-1, max_sen_len, self.hidden_size])
lstm_outputs = lstm_bkg
alphas = attention(lstm_bkg, [],
sen_len,
max_sen_len,
biases_initializer=self.b_init,
weights_initializer=self.w_init)
sen_bkg = alphas @ lstm_outputs
sen_bkg = tf.reshape(sen_bkg, [-1, self.hidden_size],
name='new_bkg')
outputs = tf.reshape(sen_bkg, [-1, max_doc_len, self.hidden_size])
self.alphas = alphas
with tf.variable_scope('document_layer'):
# lstm_outputs, _state = lstm(outputs, doc_len, self.hidden_size, 'lstm')
doc_task_label = task_label
doc_init_state = tf.get_variable(
'doc_init_state', [self.task_cnt, 2 * self.hidden_size])
doc_init_state = tf.nn.embedding_lookup(doc_init_state,
doc_task_label)
doc_init_state_fw = tf.nn.rnn_cell.LSTMStateTuple(
doc_init_state[:, :self.hidden_size // 2],
doc_init_state[:, self.hidden_size // 2:self.hidden_size])
doc_init_state_bw = tf.nn.rnn_cell.LSTMStateTuple(
doc_init_state[:, self.hidden_size:self.hidden_size * 3 // 2],
doc_init_state[:, self.hidden_size * 3 // 2:])
doc_init_state = (doc_init_state_fw, doc_init_state_bw)
lstm_bkg, _state = lstm(outputs, doc_len, self.hidden_size,
'lstm_bkg', doc_init_state)
lstm_outputs = lstm_bkg
alphas = attention(lstm_bkg, [],
doc_len,
max_doc_len,
biases_initializer=self.b_init,
weights_initializer=self.w_init)
doc_bkg = alphas @ lstm_outputs
doc_bkg = tf.reshape(doc_bkg, [-1, self.hidden_size],
name='new_bkg')
outputs = doc_bkg
return outputs
def build(self, data_iter, bert_config_file):
# get the inputs
with tf.variable_scope('inputs'):
input_map = data_iter.get_next()
usrid, prdid, input_x, input_y, doc_len = \
(input_map['usr'], input_map['prd'],
input_map['content'], input_map['rating'],
input_map['doc_len'])
input_x = tf.cast(input_x, tf.int32)
self.usr = lookup(self.embeddings['usr_emb'],
usrid,
name='cur_usr_embedding')
self.prd = lookup(self.embeddings['prd_emb'],
prdid,
name='cur_prd_embedding')
# input_x = lookup(self.embeddings['wrd_emb'], input_x, name='cur_wrd_embedding')
input_x = tf.reshape(input_x, [-1, self.max_doc_len])
input_mask = tf.sequence_mask(doc_len, self.max_doc_len)
input_mask = tf.cast(input_mask, tf.int32)
bert_config = BertConfig.from_json_file(bert_config_file)
bert = BertModel(bert_config,
is_training=True,
input_ids=input_x,
input_mask=input_mask,
token_type_ids=None,
use_one_hot_embeddings=False)
pooled_output = bert.get_pooled_output()
sequence_output = bert.get_sequence_output()
alphas = attention(sequence_output, None, self.max_doc_len,
self.max_doc_len)
sequence_output = tf.matmul(alphas, sequence_output)
sequence_output = tf.squeeze(sequence_output, axis=1)
bert_output = tf.concat([pooled_output, sequence_output], axis=1)
logits = tf.layers.dense(
bert_output,
self.cls_cnt,
kernel_initializer=tf.truncated_normal_initializer(stddev=0.02))
self.bert_output = bert_output
self.logits = logits
# build the process of model
prediction = tf.argmax(logits, 1, name='prediction')
self.prediction = prediction
with tf.variable_scope("loss"):
sce = tf.nn.softmax_cross_entropy_with_logits_v2
log_probs = tf.nn.log_softmax(logits)
self.probs = tf.nn.softmax(logits)
loss = -tf.reduce_sum(tf.one_hot(
input_y, self.cls_cnt, dtype=tf.float32) * log_probs,
axis=-1)
self.loss = tf.reduce_mean(loss)
# self.loss = sce(logits=logits, labels=tf.one_hot(input_y, self.cls_cnt))
# self.loss = tf.reduce_mean(self.loss)
self.total_loss = tf.reduce_sum(loss)
prediction = tf.argmax(logits, 1, name='prediction')
with tf.variable_scope("metrics"):
correct_prediction = tf.equal(prediction, input_y)
self.correct = correct_prediction
mse = tf.reduce_sum(tf.square(prediction - input_y), name="mse")
correct_num = tf.reduce_sum(tf.cast(correct_prediction,
dtype=tf.int32),
name="correct_num")
accuracy = tf.reduce_sum(tf.cast(correct_prediction, "float"),
name="accuracy")
return self.total_loss, mse, correct_num, accuracy
