def get_history_sum_embedded(self):
# TODO: add mask info for this operation
his_days =['one', 'two', 'three', 'four','five','six','seven','eight','nine',
'ten','eleven','twelve','thirteen','fourteen']
for fir in his_days:
key = "history_" + fir + "_chap_ph"
embed_key = "history_" + fir + "_chap_embedded"
setattr(self, embed_key,get_mask_zero_embedded(self.chapters_embeddings_var,
getattr(self, key)))
for fir in his_days:
key = "history_" + fir + "_sec_ph"
embed_key = "history_" + fir + "_sec_embedded"
setattr(self, embed_key,get_mask_zero_embedded(self.sections_embeddings_var,
getattr(self, key)))
chap = tf.reduce_mean(self.history_one_chap_embedded, axis=-2)
#b*x*128 b*128 b*(128*14)
sec = tf.reduce_mean(self.history_one_sec_embedded, axis=-2)
for fir in his_days[:0:-1]:
key_c = "history_" + fir + "_chap_embedded"
chap = tf.concat([chap, tf.reduce_mean(getattr(self, key_c), axis=-2)], axis=-1)
key_s = "history_" + fir + "_sec_embedded"
sec = tf.concat([sec, tf.reduce_mean(getattr(self, key_s), axis=-2)], axis=-1)
history_chap_emb = tf.reshape(chap, [-1, HIS_DAYS, EMBEDDING_DIM])
history_sec_emb = tf.reshape(sec, [-1, HIS_DAYS, EMBEDDING_DIM])
chap_mean = tf.reduce_mean(history_chap_emb, axis=-2)
sec_mean = tf.reduce_mean(history_sec_emb, axis=-2)
return chap_mean, sec_mean
def get_history_din_embedded(self):
# TODO: add mask info for this operation
his_days = [
'one', 'two', 'three', 'four', 'five', 'six', 'seven', 'eight', 'nine',
'ten', 'eleven', 'twelve', 'thirteen', 'fourteen'
]
for fir in his_days:
key = "history_" + fir + "_chap_ph"
embed_key = "history_" + fir + "_chap_embedded"
setattr(
self, embed_key,
get_mask_zero_embedded(self.chapters_embeddings_var,
getattr(self, key)))
for fir in his_days:
key = "history_" + fir + "_sec_ph"
embed_key = "history_" + fir + "_sec_embedded"
setattr(
self, embed_key,
get_mask_zero_embedded(self.sections_embeddings_var,
getattr(self, key)))
history_cha_embedded = []
history_sec_embedded = []
for fir in his_days[::-1]:
key_c = "history_" + fir + "_chap_embedded"
key_s = "history_" + fir + "_sec_embedded"
history_cha_embedded.append(get_rnn_sum(getattr(self, key_c),
"rnncha"))
history_sec_embedded.append(get_rnn_sum(getattr(self, key_s),
"rnnsec"))
#self.history_all_embedded = tf.reshape(, [None,len(history_all_embedded),EMBEDDING_DIM])
# T*B*N -。 B*T*N
history_cha_emb = tf.transpose(history_cha_embedded, [1, 0, 2])
history_sec_emb = tf.transpose(history_sec_embedded, [1, 0, 2])
attention_cha_output = din_attention(
tf.concat([
get_rnn_sum(self.today_chapters_embedded, "rnncha"),
get_rnn_sum(self.today_sections_embedded, "rnnsec")
],
axis=-1),
tf.concat([history_cha_emb, history_sec_emb], axis=-1),
ATTENTION_SIZE,
self.history_mask_cha_ph,
stag="cha")
att_fea1 = tf.reduce_sum(attention_cha_output, -2)
#atte_out = tf.concat([att_fea1,att_fea2],axis=-1)
return att_fea1
def get_rnn_sum(input_seq, name='cha'):
#with tf.name_scope("GRU"):
num_layers = 2
HIDDEN_DIM = 128
KEEP_PROB = 0.8
with tf.name_scope('cell'), tf.variable_scope("cell", reuse=tf.AUTO_REUSE):
def build_cell(n, m):
cell = tf.nn.rnn_cell.GRUCell(n)
cell = tf.nn.rnn_cell.DropoutWrapper(cell, output_keep_prob=m)
return cell
num_units = [HIDDEN_DIM // 2, HIDDEN_DIM // 2]
cell_fw = [build_cell(n, KEEP_PROB) for n in num_units]
cell_bw = [build_cell(n, KEEP_PROB) for n in num_units]
with tf.name_scope('gru'), tf.variable_scope("gru", reuse=tf.AUTO_REUSE):
biout, output_fw, output_bw = tf.contrib.rnn.stack_bidirectional_dynamic_rnn(
cell_fw, cell_bw, inputs=input_seq, dtype=tf.float32, scope=name)
rnnoutput = tf.reduce_sum(tf.concat([biout, input_seq], axis=-1),
axis=-2)
return rnnoutput
def __call__(self, i_to_s, state, scope="DiagonalBiLSTMCell"):
c_prev = tf.slice(state, [0, 0], [-1, self._num_units])
h_prev = tf.slice(state, [0, self._num_units], [-1, self._num_units])
with tf.compat.v1.variable_scope(scope):
conv1d_inputs = tf.reshape(
h_prev, [-1, self._height, 1, self._hidden_dims],
name='conv1d_inputs')
conv_s_to_s = conv1d(conv1d_inputs,
4 * self._hidden_dims,
2,
scope='s_to_s')
s_to_s = tf.reshape(conv_s_to_s,
[-1, self._height * self._hidden_dims * 4])
lstm_matrix = tf.sigmoid(s_to_s + i_to_s)
i, g, f, o = tf.split(lstm_matrix, 4, 1)
c = f * c_prev + i * g
h = tf.multiply(o, tf.tanh(c), name='hid')
new_state = tf.concat([c, h], 1)
return h, new_state
def get_history_sum_embedded(self):
# TODO: add mask info for this operation
his_days =['one', 'two', 'three', 'four','five','six','seven','eight','nine',
'ten','eleven','twelve','thirteen','fourteen']
for fir in his_days:
key_c = "history_" + fir + "_chap_ph"
embed_key_c = "history_" + fir + "_chap_embedded"
setattr(self, embed_key_c,get_mask_zero_embedded(self.chapters_embeddings_var,
getattr(self, key_c)))
key_s = "history_" + fir + "_sec_ph"
embed_key_s = "history_" + fir + "_sec_embedded"
setattr(self, embed_key_s,get_mask_zero_embedded(self.sections_embeddings_var,
getattr(self, key_s)))
key_st = "style_" + fir + "_ph"
embed_key_st = "history_" + fir + "_sty_embedded"
setattr(self, embed_key_st,tf.nn.embedding_lookup(self.style_embeddings_var,
getattr(self, key_st)))
chap = get_rnn_sum(self.history_one_chap_embedded,"rnncha")
#b*x*128 b*128 b*(128*14)
sec = get_rnn_sum(self.history_one_sec_embedded, "rnnsec")
sty = self.history_one_sty_embedded
for fir in his_days[:0:-1]:
key_c = "history_" + fir + "_chap_embedded"
chap = tf.concat([chap, get_rnn_sum(getattr(self, key_c), "rnncha")], axis=-1)
key_s = "history_" + fir + "_sec_embedded"
sec = tf.concat([sec, get_rnn_sum(getattr(self, key_s), "rnnsec")], axis=-1)
key_st = "history_" + fir + "_sty_embedded"
sty = tf.concat([sty, getattr(self, key_st)], axis=-1)
history_chap_emb = tf.reshape(chap, [-1, HIS_DAYS, EMBEDDING_DIM])
history_sec_emb = tf.reshape(sec, [-1, HIS_DAYS, EMBEDDING_DIM])
history_sty_emb = tf.reshape(sty, [-1, HIS_DAYS, EMBEDDING_DIM])
chap_mean = tf.reduce_mean(history_chap_emb, axis=-2)
sec_mean = tf.reduce_mean(history_sec_emb, axis=-2)
sty_mean = tf.reduce_mean(history_sty_emb, axis=-2)
#return chap_mean, sec_mean
return tf.concat([chap_mean, sec_mean,sty_mean], axis=-1)
def get_mask_zero_embedded(var_em, var_ph):
mask = tf.equal(var_ph, 0)
mask2 = tf.concat(
[tf.expand_dims(~mask, -1) for i in range(EMBEDDING_DIM)], -1)
rst = tf.where(
mask2, tf.nn.embedding_lookup(var_em, var_ph),
tf.zeros([tf.shape(var_ph)[0],
tf.shape(var_ph)[1], EMBEDDING_DIM]))
return rst
def unskew(inputs, width=None, scope="unskew"):
with tf.compat.v1.name_scope(scope):
batch, height, skewed_width, channel = inputs.get_shape().as_list()
width = width if width else height
new_rows = []
rows = tf.split(inputs, height, 1)
for idx, row in enumerate(rows):
new_rows.append(tf.slice(row, [0, 0, idx, 0], [-1, -1, width, -1]))
outputs = tf.concat(new_rows, 1, name="output")
return outputs
def build_fcn_net(self, inp, use_dice=False):
with self.graph.as_default():
self.saver = tf.train.Saver(max_to_keep=1)
with tf.name_scope("Out"):
bn1 = tf.layers.batch_normalization(inputs=inp, name='bn1')
dnn1 = tf.layers.dense(bn1, 200, activation=None, name='f1')
if use_dice:
dnn1 = dice(dnn1, name='dice_1')
else:
dnn1 = prelu(dnn1, 'prelu1')
dnn2 = tf.layers.dense(dnn1, 80, activation=None, name='f2')
if use_dice:
dnn2 = dice(dnn2, name='dice_2')
else:
dnn2 = prelu(dnn2, 'prelu2')
dnn3 = tf.layers.dense(dnn2, 2, activation=None, name='f3')
self.y_hat = tf.nn.softmax(dnn3) + 0.00000001
with tf.name_scope('Metrics'):
# Cross-entropy loss and optimizer initialization
coe = tf.constant([1.2, 1.2])
coe_mask = tf.equal(self.core_type_ph, 1)
coe_mask2 = tf.concat(
[tf.expand_dims(coe_mask, -1) for i in range(2)], -1)
self.target_ph_coe = tf.where(coe_mask2, self.target_ph * coe,
self.target_ph)
ctr_loss = -tf.reduce_mean(tf.log(self.y_hat) * self.target_ph)
self.loss = ctr_loss
# tf.summary.scalar('loss', self.loss)
self.optimizer = tf.train.AdamOptimizer(
learning_rate=self.lr_ph).minimize(self.loss)
# self.optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.lr_ph).minimize(self.loss)
# Accuracy metric
self.accuracy = tf.reduce_mean(
tf.cast(tf.equal(tf.round(self.y_hat), self.target_ph),
tf.float32))
# tf.summary.scalar('accuracy', self.accuracy)
self.merged = tf.summary.merge_all()
def diagonal_bilstm(inputs, scope='diagonal_bilstm'):
with tf.compat.v1.variable_scope(scope):
def reverse(inputs):
return tf.reverse(inputs, [2])
output_state_fw = diagonal_lstm(inputs, scope='output_state_fw')
output_state_bw = reverse(
diagonal_lstm(reverse(inputs), scope='output_state_bw'))
batch, height, width, channel = output_state_bw.get_shape().as_list()
output_state_bw_except_last = tf.slice(output_state_bw, [0, 0, 0, 0],
[-1, height - 1, -1, -1])
output_state_bw_only_last = tf.slice(output_state_bw,
[0, height - 1, 0, 0],
[-1, 1, -1, -1])
dummy_zeros = tf.zeros_like(output_state_bw_only_last)
output_state_bw_with_last_zeros = tf.concat(
[output_state_bw_except_last, dummy_zeros], 1)
return output_state_fw + output_state_bw_with_last_zeros
def __init__(self, *, use_dice=False):
super().__init__(use_dice=use_dice)
self.other_inputs()
teacher = [
#self.teacher_id_embedded,
self.province_id_embedded,
self.city_id_embedded,
self.core_type_embedded,
self.student_count_embedded, ]
# 0-4
clazz = [
#self.class_id_embedded,
self.edition_id_embedded,
self.grade_id_embedded,
self.class_student_embedded,
self.cap_avg_embedded,
self.cap_max_embedded,
self.cap_min_embedded, ]
# 5-11
study = [
self.study_vector_embedded,
self.gap_days_embedded, ]
# 12-13
submit = [self.month_submit_rate_embedded,
]
# 14
capacity = [self.region_capacity_embedded,
]
# 15
prefer = [self.prefer_assign_time_avg_embedded,
self.prefer_assign_time_var_embedded,
self.prefer_assign_rank_avg_embedded,
self.prefer_assign_rank_var_embedded, ]
# 16-19
register = [self.register_diff_embedded,
]
# 20
homeworkcount = [self.homework_count_embedded,
]
# 21
weekcount = [self.week_count_embedded,
]
# 22
lastday = [self.lastday_count_embedded,
]
# 23
study_analysis = [self.analysis_avg_times_embedded,
self.analysis_avg_exp_score_embedded,
self.analysis_avg_rate_embedded,
self.analysis_avg_exp_level_embedded
]
o = teacher + clazz + study + submit + \
capacity + prefer + register + homeworkcount + weekcount + lastday+study_analysis
# here, we do like this, not tf.concat(o,axis=-1),
# because, issue:https://github.com/tensorflow/tensorflow/issues/24816
# ps: style,homework,reflect don't need to do like this, proved.
others = o[0]
for i in o[1:]:
others = tf.concat([others, i], axis=-1)
others = [others]
# style = []
# for fir in ["1", "2", "3", "4"]:
# for sec in ["100", "010", "001", "110", "101", "011", "111"]:
# embed_key = "style_" + fir + "0" + sec + "_embedded"
# style.append(getattr(self, embed_key))
homework = []
homework.append(self.today_style_embedded)
homework.append(tf.concat([self.today_cha_rnn,self.today_sec_rnn],axis=-1))
homework.append(self.history_chap_embedded)
homework.append(self.history_chap_embedded *tf.concat([self.today_cha_rnn,self.today_sec_rnn,self.today_style_embedded],axis=-1))
reflect = []
reflect.append(self.ref_rnn )
with self.graph.as_default():
with tf.name_scope("Concat"):
inps = tf.concat(others + homework + reflect, -1)
self.build_fcn_net(inps, self.use_dice)
def get_history_bgru_embedded(self):
# TODO: add mask info for this operation
his_days = [
'one', 'two', 'three', 'four', 'five', 'six', 'seven', 'eight', 'nine',
'ten', 'eleven', 'twelve', 'thirteen', 'fourteen'
]
for fir in his_days:
key = "history_" + fir + "_chap_ph"
embed_key = "history_" + fir + "_chap_embedded"
setattr(
self, embed_key,
get_mask_zero_embedded(self.chapters_embeddings_var,
getattr(self, key)))
for fir in his_days:
key = "history_" + fir + "_sec_ph"
embed_key = "history_" + fir + "_sec_embedded"
setattr(
self, embed_key,
get_mask_zero_embedded(self.sections_embeddings_var,
getattr(self, key)))
history_cha_embedded = []
history_sec_embedded = []
for fir in his_days[::-1]:
key_c = "history_" + fir + "_chap_embedded"
key_s = "history_" + fir + "_sec_embedded"
# B 3 128 B 128 14 B 128 B 14 128
history_cha_embedded.append(get_rnn_sum(getattr(self, key_c),
"rnncha"))
history_sec_embedded.append(get_rnn_sum(getattr(self, key_s),
"rnnsec"))
history_cha_emb = tf.transpose(history_cha_embedded, [1, 0, 2])
history_sec_emb = tf.transpose(history_sec_embedded, [1, 0, 2])
with tf.name_scope("GRU"):
num_layers = 2
HIDDEN_DIM = 128
KEEP_PROB = 0.8
with tf.name_scope('cell'):
def build_cell(n, m):
cell = tf.nn.rnn_cell.GRUCell(n)
cell = tf.nn.rnn_cell.DropoutWrapper(cell, output_keep_prob=m)
return cell
num_units = [HIDDEN_DIM // 2, HIDDEN_DIM]
cell_fw = [build_cell(n, KEEP_PROB) for n in num_units]
cell_bw = [build_cell(n, KEEP_PROB) for n in num_units]
with tf.name_scope('gru'):
biout, output_fw, output_bw = tf.contrib.rnn.stack_bidirectional_dynamic_rnn(
cell_fw,
cell_bw,
inputs=tf.concat([history_cha_emb, history_sec_emb], axis=-1),
dtype=tf.float32,
scope='cha')
rnnoutput = tf.reduce_sum(biout, axis=-2)
return rnnoutput
def get_history_din_embedded(self):
# TODO: add mask info for this operation
his_days = [
'one', 'two', 'three', 'four', 'five', 'six', 'seven', 'eight', 'nine',
'ten', 'eleven', 'twelve', 'thirteen', 'fourteen'
]
for fir in his_days:
key = "history_" + fir + "_chap_ph"
embed_key = "history_" + fir + "_chap_embedded"
setattr(
self, embed_key,
get_mask_zero_embedded(self.chapters_embeddings_var,
getattr(self, key)))
for fir in his_days:
key = "history_" + fir + "_sec_ph"
embed_key = "history_" + fir + "_sec_embedded"
setattr(
self, embed_key,
get_mask_zero_embedded(self.sections_embeddings_var,
getattr(self, key)))
history_cha_embedded = []
history_sec_embedded = []
for fir in his_days[::-1]:
key_c = "history_" + fir + "_chap_embedded"
key_s = "history_" + fir + "_sec_embedded"
history_cha_embedded.append(get_rnn_sum(getattr(self, key_c),
"rnncha"))
history_sec_embedded.append(get_rnn_sum(getattr(self, key_s),
"rnnsec"))
#self.history_all_embedded = tf.reshape(, [None,len(history_all_embedded),EMBEDDING_DIM])
# T*B*N -。 B*T*N
history_cha_emb = tf.transpose(history_cha_embedded, [1, 0, 2])
history_sec_emb = tf.transpose(history_sec_embedded, [1, 0, 2])
#dien
with tf.name_scope('rnn_1'):
rnn_outputs, _ = dynamic_rnn(GRUCell(HIDDEN_SIZE * 2),
inputs=tf.concat(
[history_cha_emb, history_sec_emb],
axis=-1),
sequence_length=self.history_len_ph,
dtype=tf.float32,
scope="gru1")
with tf.name_scope('Attention_layer_1'):
att_outputs, alphas = din_fcn_attention(tf.concat([
get_rnn_sum(self.today_chapters_embedded, "rnncha"),
get_rnn_sum(self.today_sections_embedded, "rnnsec")
],
axis=-1),
rnn_outputs,
ATTENTION_SIZE,
self.history_mask_cha_ph,
scope="1_1",
softmax_stag=1,
stag='1_1',
mode='LIST',
return_alphas=True)
with tf.name_scope('rnn_2'):
rnn_outputs2, final_state2 = dynamic_rnn(
VecAttGRUCell(HIDDEN_SIZE * 2),
inputs=rnn_outputs,
att_scores=tf.expand_dims(alphas, -1),
sequence_length=self.history_len_ph,
dtype=tf.float32,
scope="gru2")
return final_state2
