def __init__(self, dim_v, dim_h, samples
, init_w= tf.random_uniform_initializer(minval= -0.01, maxval= 0.01)
, ftype= tf.float32
, scope= 'rbm'):
self.dim_v, self.dim_h, self.ftype, self.scope = dim_v, dim_h, ftype, scope
with tf.variable_scope(scope):
# todo add bias
self.w = tf.get_variable(name= 'w', shape= (self.dim_v, self.dim_h), initializer= init_w)
# positive stage: inference
self.v_ = tf.placeholder(name= 'v_', dtype= self.ftype, shape= (None, self.dim_v))
with tf.name_scope('hgv'):
self.hgv = tf.sigmoid(tf.matmul(self.v_, self.w))
# self.act_h = binary(self.hgv, transform= False, threshold= None)
# self.h_ = tf.placeholder(name= 'h_', dtype= self.ftype, shape= (None, self.dim_h))
# self.vgh = tf.matmul(self.h_, self.w, transpose_b= True)
# self.act_v = binary(self.vgh, transform= False, threshold= None)
with tf.name_scope('pos'):
self.pos = tf.matmul(self.v_, self.hgv, transpose_a= True)
self.pos /= tf.cast(tf.shape(self.v_)[0], dtype= self.ftype)
# negative stage: stochastic approximation
self.v = binary_variable(name= 'v', shape= (samples, self.dim_v), dtype= self.ftype)
self.h = binary_variable(name= 'h', shape= (samples, self.dim_h), dtype= self.ftype)
self.k_ = tf.placeholder(name= 'k_', dtype= tf.int32, shape= ())
def gibbs(v, _h):
h = binary(tf.matmul(v, self.w))
v = binary(tf.matmul(h, self.w, transpose_b= True))
# todo real valued v
# v = tf.sigmoid(tf.matmul(h, self.w, transpose_b= True))
return v, h
with tf.name_scope('gibbs'):
vh = self.v, self.h
v, h = self.gibbs = tuple(
tf.assign(x, x2, validate_shape= False) for x, x2 in zip(
vh, tf.while_loop(
loop_vars= (self.k_, vh)
, cond= lambda k, vh: (0 < k)
, body= lambda k, vh: (k - 1, gibbs(*vh)))[1]))
with tf.name_scope('neg'):
# todo update with real probabilities instead of binaries
h = tf.sigmoid(tf.matmul(v, self.w))
v = tf.sigmoid(tf.matmul(h, self.w, transpose_b= True))
self.neg = tf.matmul(v, h, transpose_a= True)
self.neg /= tf.cast(tf.shape(self.v)[0], dtype= self.ftype)
self.lr_ = tf.placeholder(name= 'lr_', dtype= self.ftype, shape= ())
with tf.name_scope('up'):
self.up = self.w.assign_add((self.pos - self.neg) * self.lr_).op
self.step = 0
def __init__(self, dim, samples
, init_w= tf.random_uniform_initializer(minval= -0.01, maxval= 0.01)
, ftype= tf.float32, scope= 'sbn'):
self.dim, self.ftype, self.scope = dim, ftype, scope
with tf.variable_scope(scope):
self.wr = tuple(
tf.get_variable(name= "wr{}".format(i), shape= (dim_d, dim_a), initializer= init_w)
for i, (dim_d, dim_a) in enumerate(zip(self.dim, self.dim[1:]), 1))
self.wg = tuple(
tf.get_variable(name= "wg{}".format(i), shape= (dim_a, dim_d), initializer= init_w)
for i, (dim_d, dim_a) in enumerate(zip(self.dim, self.dim[1:]), 1))[::-1]
self.lr_ = tf.placeholder(name= 'lr_', dtype= self.ftype, shape= ())
# wake
self.v_ = tf.placeholder(name= 'v_', dtype= self.ftype, shape= (None, self.dim[0]))
with tf.name_scope('wake'):
recogn = [self.v_]
for w in self.wr: recogn.append(binary(tf.matmul(recogn[-1], w)))
self.recogn = tuple(recogn)
recogn = recogn[::-1]
eps = self.lr_ / tf.cast(tf.shape(self.v_)[0], dtype= self.ftype)
self.wake = tuple(
w.assign_add(tf.matmul(sk, (sj - pj), transpose_a= True) * eps).op
for w, sk, sj, pj in zip(
self.wg, recogn, recogn[1:]
, (tf.sigmoid(tf.matmul(s, w))
for w, s in zip(self.wg, recogn))))
# sleep
with tf.name_scope('a'):
self.a = tf.round(tf.random_uniform(shape= (samples, self.dim[-1])))
with tf.name_scope('sleep'):
recons = [self.a]
for w in self.wg: recons.append(binary(tf.matmul(recons[-1], w)))
self.recons = tuple(recons)
recons = recons[::-1]
eps = self.lr_ / tf.cast(tf.shape(self.a)[0], dtype= self.ftype)
self.sleep = tuple(
w.assign_add(tf.matmul(sj, (sk - qk), transpose_a= True) * eps).op
for w, sj, sk, qk in zip(
self.wr, recons, recons[1:]
, (tf.sigmoid(tf.matmul(s, w))
for w, s in zip(self.wr, recons))))
# the waking world is the amnesia of dream.
self.v = self.recons[-1]
self.step = 0
def __init__(self, dim, samples
, init_w= tf.random_uniform_initializer(minval= -0.01, maxval= 0.01)
, ftype= tf.float32, scope= 'dbn'):
self.dim, self.ftype = dim, ftype
with tf.variable_scope(scope):
self.rbm = tuple(
Rbm(scope= "rbm{}".format(i)
, dim_v= dim_v
, dim_h= dim_h
, samples= samples
, init_w= init_w
, ftype= self.ftype)
for i, (dim_v, dim_h) in enumerate(zip(dim, dim[1:]), 1))
self.w = tuple(rbm.w for rbm in self.rbm[::-1])
self.wg = tuple(tf.transpose(w) for w in self.w)
self.wr = tuple(
tf.get_variable(name= "wr{}".format(i), shape= (dim_d, dim_a), initializer= init_w)
for i, (dim_d, dim_a) in enumerate(zip(self.dim, self.dim[1:]), 1))
self.lr_ = tf.placeholder(name= 'lr_', dtype= self.ftype, shape= ())
# wake
self.v_ = self.rbm[0].v_
with tf.name_scope('wake'):
recogn = [self.v_]
for w in self.wr: recogn.append(binary(tf.matmul(recogn[-1], w)))
self.recogn = tuple(recogn)
recogn = recogn[::-1]
eps = self.lr_ / tf.cast(tf.shape(self.v_)[0], dtype= self.ftype)
self.wake = tuple(
w.assign_add(tf.matmul((sj - pj), sk, transpose_a= True) * eps).op
for w, sk, sj, pj in zip(
self.w, recogn, recogn[1:]
, (tf.sigmoid(tf.matmul(s, w))
for w, s in zip(self.wg, recogn))))
# sleep
top = self.rbm[-1]
self.k_, (self.v, self.a) = top.k_, top.gibbs
with tf.name_scope('sleep'):
recons = [self.a, self.v]
for w in self.wg[1::]: recons.append(binary(tf.matmul(recons[-1], w)))
self.recons = tuple(recons)
recons = recons[::-1]
eps = self.lr_ / tf.cast(tf.shape(self.a)[0], dtype= self.ftype)
self.sleep = tuple(
w.assign_add(tf.matmul(sj, (sk - qk), transpose_a= True) * eps).op
for w, sj, sk, qk in zip(
self.wr, recons, recons[1:]
, (tf.sigmoid(tf.matmul(s, w))
for w, s in zip(self.wr, recons))))
# the waking world is the amnesia of dream.
self.v = self.recons[-1]
with tf.name_scope('ances'):
self.a = self.rbm[-1].h
ances = [self.a]
for w in self.wg: ances.append(binary(tf.matmul(ances[-1], w)))
self.ances = ances[-1]
self.step = 0
def __init__(self, *, use_dice=False):
self.graph = tf.Graph()
self.tensor_info = {}
self.use_dice = use_dice
with self.graph.as_default():
# Main Inputs
with tf.name_scope('Main_Inputs'):
self.target_ph = tf.placeholder(tf.float32, [None, None], name='target_ph')
self.lr_ph = tf.placeholder(tf.float32, [], name="lr_ph")
# with tf.name_scope("Teacher_Info"):
self.teacher_id_ph = tf.placeholder(tf.int32, [None, ], name="teacher_id_ph")
self.student_count_ph = tf.placeholder(tf.int32, [None, ], name="student_count_ph")
self.province_id_ph = tf.placeholder(tf.int32, shape=[None, ], name="province_id_ph")
self.city_id_ph = tf.placeholder(tf.int32, shape=[None, ], name="city_id_ph")
# TODO: binary 是否金牌讲师
self.core_type_ph = tf.placeholder(tf.int32, shape=[None, ], name="core_type_ph")
# with tf.name_scope("Class_Info"):
#今天 教的班级id
self.class_id_ph = tf.placeholder(tf.int32, [None, ], name="class_id_ph")
#课本版本
self.edition_id_ph = tf.placeholder(tf.int32, [None, ], name="edition_id_ph")
self.grade_id_ph = tf.placeholder(tf.int32, [None, ], name="grade_id_ph")
#老师教的所有班级 学生总人数， int连续特征
self.class_student_ph = tf.placeholder(tf.int32, [None, ], name="class_student_ph")
#kefei
#浮点连续特征
self.cap_avg_ph = tf.placeholder(tf.float32, [None, ], name="cap_avg_ph")
self.cap_max_ph = tf.placeholder(tf.float32, [None, ], name="cap_max_ph")
self.cap_min_ph = tf.placeholder(tf.float32, [None, ], name="cap_min_ph")
# with tf.name_scope("Homework_Info"):
#候选， 召回集 首先天宇会给一个初步刷选的作业集，好几组，每组进去很多题目，暂且不管
#粒度暂且放在一组，上，看做一个 作业， 特征属性两个chapters sections
#这一组中，今天这个老师 id，，布置了某一组，则吧这个 label 为1 其他布置的 为0 这样就构造了样本
#另外chapters sections分别可能是几个数字®️的， 类似于多lable吧， 为了保持统一长度， 所以补零
self.today_chapters_ph = tf.placeholder(tf.int32, [None, None], name="today_chapters_ph")
self.today_sections_ph = tf.placeholder(tf.int32, [None, None], name="today_sections_ph")
#没用
self.today_chap_mask_ph = tf.placeholder(tf.float32, [None, None], name='today_chap_mask_ph')
self.today_chap_len_ph = tf.placeholder(tf.int32, [None, ], name='today_chap_len_ph')
self.today_sec_mask_ph = tf.placeholder(tf.float32, [None, None], name='today_sec_mask_ph')
self.today_sec_len_ph = tf.placeholder(tf.int32, [None, ], name='today_sec_len_ph')
#作业的风格 是这道题的 风格， 什么预习啊 什么深度啊，，
self.today_style_ph = tf.placeholder(tf.int32, [None, ], name='today_style_ph')
# TODO: use three dims to capture more history info
#这个是 这个班级前三天 给布置的作业， 比如，昨天的，仍是两个特征来表征，chap sec 每个都是多个数字，
#所以N N
for fir in ['one', 'two', 'three', 'four','five','six','seven','eight','nine','ten','eleven','twelve','thirteen','fourteen']:
key_s = "history_" + fir + "_sec_ph"
sty = "style_" + fir + "_ph"
key_c = "history_" + fir + "_chap_ph"
setattr(self, key_c,
tf.placeholder(tf.int32, [None, None], name=key_c))
setattr(self, key_s,
tf.placeholder(tf.int32, [None, None], name=key_s))
setattr(self, sty,
tf.placeholder(tf.int32, [None,], name=sty))
# TODO: All belows should consider the type and input
# with tf.name_scope("Study_Info"):
# TODO: study_vector_ph's type can change?
#kefei 这个班级的学习能力 类似于期中考试， 这个班级 表征为 20维的向量 int
self.study_vector_ph = tf.placeholder(tf.float32, [None, 20], name="study_vector_ph")
#上面的结果 什么时候评测的 连续值，隔的天数
self.gap_days_ph = tf.placeholder(tf.int32, [None, ], name="gap_days_ph")
self.analysis_avg_times_ph = tf.placeholder(tf.float32, [None, ], name="analysis_avg_times_ph")
self.analysis_avg_rate_ph = tf.placeholder(tf.float32, [None, ], name="analysis_avg_rate_ph")
self.analysis_avg_exp_score_ph = tf.placeholder(tf.float32, [None, ], name="analysis_avg_exp_score_ph")
self.analysis_avg_exp_level_ph = tf.placeholder(tf.float32, [None, ], name="analysis_avg_exp_level_ph")
# with tf.name_scope("Submit_Info"): 这个班级 一个月的app内 作业提交率 ，连续float特征
self.month_submit_rate_ph = tf.placeholder(tf.float32, [None, ], name="month_submit_rate_ph")
# with tf.name_scope("Capacity_Info"): 地区区域整体能力 也是 float连续特征
self.region_capacity_ph = tf.placeholder(tf.float32, [None, ], name="region_capacity_ph")
# with tf.name_scope("Prefer_Info"):
#老师 在这个班级 上，，喜欢布置作业的 难度 和时间 float连续值特征
self.prefer_assign_time_avg_ph = tf.placeholder(tf.float32, [None, ],
name="prefer_assign_time_avg_ph")
self.prefer_assign_time_var_ph = tf.placeholder(tf.float32, [None, ],
name="prefer_assign_time_var_ph")
self.prefer_assign_rank_avg_ph = tf.placeholder(tf.float32, [None, ],
name="prefer_assign_rank_avg_ph")
self.prefer_assign_rank_var_ph = tf.placeholder(tf.float32, [None, ],
name="prefer_assign_rank_var_ph")
# with tf.name_scope("Register_Info"): 老师 注册app的 时间，int连续值特征
self.register_diff_ph = tf.placeholder(tf.int32, [None, ], name="register_diff_ph")
# with tf.name_scope("HomeworkCount_Info"): 老师 布置了多少题目 int连续值特征
#是总共布置的吗 从注册app？？？
self.homework_count_ph = tf.placeholder(tf.int32, [None, ], name="homework_count_ph")
# with tf.name_scope("Style_Info"):
# TODO: use 3 dims 老师 作业 的风格 ？ 一个特征域
# for fir in ["1", "2", "3", "4"]:
# for sec in ["100", "010", "001", "110", "101", "011", "111"]:
# key = "style_" + fir + "0" + sec + "_ph"
# setattr(self, key,
# tf.placeholder(tf.int32, [None, ], name=key))
# with tf.name_scope("WeekHomeworkCount_Info"):
#这周 老师布置作业，，作业率，，怎么float ？？ 连续值特征
self.week_count_ph = tf.placeholder(tf.float32, [None, ], name="week_count_ph")
# with tf.name_scope("Reflect_Info"):
# TODO: explore more graceful 映射 作业类目
self.reflect_value_ph = tf.placeholder(tf.int32, [None, None], name="reflect_value_ph")
self.reflect_mask_ph = tf.placeholder(tf.float32, [None, None], name="reflect_mask_ph")
self.reflect_len_ph = tf.placeholder(tf.int32, [None, ], name="reflect_len_ph")
# with tf.name_scope("Lastdat_Info"): 昨天布置的 个数 int连续实值
self.lastday_count_ph = tf.placeholder(tf.int32, [None, ], name="lastday_count_ph")
# Embedding layer
with tf.name_scope('Main_Embedding_layer'):
# almost done
with tf.name_scope("Others"):
# teacher
with tf.name_scope("Teacher"):
self.teacher_id_embeddings_var = tf.get_variable("teacher_id_embeddings_var",
[N_TEACHER, EMBEDDING_DIM], )
# tf.summary.histogram('teacher_id_embeddings_var', self.teacher_id_embeddings_var)
self.teacher_id_embedded = tf.nn.embedding_lookup(self.teacher_id_embeddings_var,
self.teacher_id_ph, )
self.province_id_embeddings_var = tf.get_variable("province_id_embeddings_var",
[N_PROVINCE, EMBEDDING_DIM])
# tf.summary.histogram('province_id_embeddings_var', self.province_id_embeddings_var)
self.province_id_embedded = tf.nn.embedding_lookup(self.province_id_embeddings_var,
self.province_id_ph)
self.city_id_embeddings_var = tf.get_variable("city_id_embeddings_var",
[N_CITY, EMBEDDING_DIM])
# tf.summary.histogram('city_id_embeddings_var', self.city_id_embeddings_var)
self.city_id_embedded = tf.nn.embedding_lookup(self.city_id_embeddings_var,
self.city_id_ph)
self.core_type_embeddings_var = tf.get_variable("core_type_embeddings_var",
[2, EMBEDDING_DIM])
# tf.summary.histogram('core_type_embeddings_var', self.core_type_embeddings_var)
self.core_type_embedded = tf.nn.embedding_lookup(self.core_type_embeddings_var,
self.core_type_ph)
# just to use embedded for var,maybe tf.identify?
self.student_count_embedded = get_self_or_expand_dims(self.student_count_ph)
with tf.name_scope("Class"):
self.class_id_embeddings_var = tf.get_variable("class_id_embeddings_var",
[N_CLASS, EMBEDDING_DIM])
# tf.summary.histogram('class_id_embeddings_var', self.class_id_embeddings_var)
self.class_id_embedded = tf.nn.embedding_lookup(self.class_id_embeddings_var,
self.class_id_ph)
self.edition_id_embeddings_var = tf.get_variable("edition_id_embeddings_var",
[N_EDITION, EMBEDDING_DIM])
# tf.summary.histogram('edition_id_embeddings_var', self.edition_id_embeddings_var)
self.edition_id_embedded = tf.nn.embedding_lookup(self.edition_id_embeddings_var,
self.edition_id_ph)
self.grade_id_embeddings_var = tf.get_variable("grade_id_embeddings_var",
[N_GRADE, EMBEDDING_DIM])
# tf.summary.histogram('grade_id_embeddings_var', self.grade_id_embeddings_var)
self.grade_id_embedded = tf.nn.embedding_lookup(self.grade_id_embeddings_var,
self.grade_id_ph)
# just to use embedded for var,maybe tf.identify?
#连续值 dense 本身有意义的直接喂入
self.class_student_embedded = get_self_or_expand_dims(self.class_student_ph)
self.cap_avg_embedded = get_self_or_expand_dims(self.cap_avg_ph)
self.cap_max_embedded = get_self_or_expand_dims(self.cap_max_ph)
self.cap_min_embedded = get_self_or_expand_dims(self.cap_min_ph)
with tf.name_scope("Study"):
# just to use embedded for var,maybe tf.identify?
self.study_vector_embedded = self.study_vector_ph
self.gap_days_embedded = get_self_or_expand_dims(self.gap_days_ph)
with tf.name_scope("Study_analysis"):
self.analysis_avg_times_embedded = get_self_or_expand_dims(self.analysis_avg_times_ph)
self.analysis_avg_rate_embedded = get_self_or_expand_dims(self.analysis_avg_rate_ph)
self.analysis_avg_exp_score_embedded = get_self_or_expand_dims(self.analysis_avg_exp_score_ph)
self.analysis_avg_exp_level_embedded = get_self_or_expand_dims(self.analysis_avg_exp_level_ph)
with tf.name_scope("Submit"):
# just to use embedded for var,maybe tf.identify?
self.month_submit_rate_embedded = get_self_or_expand_dims(self.month_submit_rate_ph)
with tf.name_scope("Capacity"):
# just to use embedded for var,maybe tf.identify?
self.region_capacity_embedded = get_self_or_expand_dims(self.region_capacity_ph)
with tf.name_scope("Prefer"):
# just to use embedded for var,maybe tf.identify?
self.prefer_assign_time_avg_embedded = get_self_or_expand_dims(
self.prefer_assign_time_avg_ph)
self.prefer_assign_time_var_embedded = get_self_or_expand_dims(
self.prefer_assign_time_var_ph)
self.prefer_assign_rank_avg_embedded = get_self_or_expand_dims(
self.prefer_assign_rank_avg_ph)
self.prefer_assign_rank_var_embedded = get_self_or_expand_dims(
self.prefer_assign_rank_var_ph)
with tf.name_scope("Register"):
self.register_diff_embedded = get_self_or_expand_dims(self.register_diff_ph)
with tf.name_scope("HomeworkCount"):
self.homework_count_embedded = get_self_or_expand_dims(self.homework_count_ph)
with tf.name_scope("WeekHomeworkCount"):
self.week_count_embedded = get_self_or_expand_dims(self.week_count_ph)
with tf.name_scope("Lastday"):
self.lastday_count_embedded = get_self_or_expand_dims(self.lastday_count_ph)
# TODO: homework and reflect and style
# with tf.name_scope("Style"):
# for fir in ["1", "2", "3", "4"]:
# for sec in ["100", "010", "001", "110", "101", "011", "111"]:
# key = "style_" + fir + "0" + sec + "_ph"
# embed_key = "style_" + fir + "0" + sec + "_embedded"
# setattr(self, embed_key,
# get_self_or_expand_dims(getattr(self, key)))
# homework
with tf.name_scope("Homework"):
self.style_embeddings_var = tf.get_variable("style_embeddings_var",
[N_STYLE, EMBEDDING_DIM])
self.chapters_embeddings_var = tf.get_variable("chapters_embeddings_var",
[N_CHAPTER, EMBEDDING_DIM])
self.sections_embeddings_var = tf.get_variable("sections_embeddings_var",
[N_SECTION, EMBEDDING_DIM])
# tf.summary.histogram('homework_embeddings_var', self.homework_embeddings_var)
self.today_chapters_embedded = get_mask_zero_embedded(self.chapters_embeddings_var,
self.today_chapters_ph)
self.today_sections_embedded = get_mask_zero_embedded(self.sections_embeddings_var,
self.today_sections_ph)
self.history_chap_embedded = get_history_sum_embedded(self)
self.today_style_embedded = tf.nn.embedding_lookup(self.style_embeddings_var,
self.today_style_ph)
self.today_cha_rnn = get_rnn_sum(self.today_chapters_embedded,"rnncha")
self.today_sec_rnn = get_rnn_sum(self.today_sections_embedded,"rnnsec")
# reflect
with tf.name_scope("Reflect"):
self.reflect_embeddings_var = tf.get_variable("reflect_embeddings_var",
[N_REFLECT, EMBEDDING_DIM])
# tf.summary.histogram('reflect_embeddings_var', self.reflect_embeddings_var)
self.reflect_value_embedded = get_mask_zero_embedded(self.reflect_embeddings_var,
self.reflect_value_ph)
self.ref_rnn = get_rnn_sum(self.reflect_value_embedded,"ref")
def __init__(self, dim, samples
, init_w= tf.random_uniform_initializer(minval= -0.01, maxval= 0.01)
, ftype= tf.float32, scope= 'dbm'):
self.dim, self.ftype = dim, ftype
# todo pretraining
with tf.variable_scope(scope):
self.rbm = tuple(
Rbm(scope= "rbm{}".format(i)
, dim_v= dim_v
, dim_h= dim_h
, samples= samples
, init_w= init_w
, ftype= self.ftype)
for i, (dim_v, dim_h) in enumerate(zip(dim, dim[1:]), 1))
self.w = tuple(rbm.w for rbm in self.rbm)
# positive stage: variational inference
self.m = tuple(rbm.h for rbm in self.rbm)
self.v_ = self.rbm[0].v_
self.k_meanf_ = tf.placeholder(name= 'k_meanf_', dtype= tf.int32, shape= ())
def meanf(m):
mf, ml = [], self.v_
for wl, wr, mr in zip(self.w, self.w[1:], m[1:]):
mf.append(tf.sigmoid(tf.matmul(ml, wl) + tf.matmul(mr, wr, transpose_b= True)))
ml = mf[-1]
mf.append(tf.sigmoid(tf.matmul(ml, wr)))
return tuple(mf)
with tf.name_scope('meanf'):
self.meanf = tuple(
tf.assign(m, mf, validate_shape= False) for m, mf in zip(
self.m, tf.while_loop(
loop_vars= (self.k_meanf_, self.m)
, cond= lambda k, _: (0 < k)
, body= lambda k, m: (k - 1, meanf(m)))[1]))
with tf.name_scope('pos'):
bs = tf.cast(tf.shape(self.v_)[0], dtype= self.ftype)
vm = (self.v_,) + self.meanf
self.pos = tuple((tf.matmul(ml, mr, transpose_a= True) / bs) for ml, mr in zip(vm, vm[1:]))
# negative stage: stochastic approximation
self.x = tuple(rbm.v for rbm in self.rbm)
self.x += (binary_variable(name= 'x', shape= (samples, self.dim[-1]), dtype= self.ftype),)
self.v = self.x[0]
self.k_gibbs_ = tf.placeholder(name= 'k_gibbs_', dtype= tf.int32, shape= ())
def gibbs(x):
x = list(x)
# update odd layers
for i, (xl, xr, wl, wr) in enumerate(zip(x[::2], x[2::2], self.w, self.w[1:])):
x[1+(2*i)] = binary(tf.matmul(xl, wl) + tf.matmul(xr, wr, transpose_b= True))
# update first layer
x[0] = binary(tf.matmul(x[1], self.w[0], transpose_b= True))
# update even layers
for i, (xl, xr, wl, wr) in enumerate(zip(x[1::2], x[3::2], self.w[1:], self.w[2:])):
x[2+(2*i)] = binary(tf.matmul(xl, wl) + tf.matmul(xr, wr, transpose_b= True))
# update last layer
x[-1] = binary(tf.matmul(x[-2], self.w[-1]))
return tuple(x)
with tf.name_scope('gibbs'):
x = self.gibbs = tuple(
tf.assign(x, xg, validate_shape= False) for x, xg in zip(
self.x, tf.while_loop(
loop_vars= (self.k_gibbs_, self.x)
, cond= lambda k, x: (0 < k)
, body= lambda k, x: (k - 1, gibbs(x)))[1]))
with tf.name_scope('neg'):
bs = tf.cast(tf.shape(self.v)[0], dtype= self.ftype)
self.neg = tuple((tf.matmul(xl, xr, transpose_a= True) / bs) for xl, xr in zip(x, x[1:]))
# parameter update
self.lr_ = tf.placeholder(name= 'lr_', dtype= self.ftype, shape= ())
with tf.name_scope('up'):
self.up = tuple(
w.assign_add((pos - neg) * self.lr_).op
for w, pos, neg in zip(self.w, self.pos, self.neg))
self.step = 0
def __init__(self, dat, dim_rec, dim_z, dim_gen, scope='vae'):
assert 2 == dat.ndim
assert isinstance(dim_rec, tuple)
assert isinstance(dim_z, int)
assert isinstance(dim_gen, tuple)
init_w = tf.variance_scaling_initializer(scale=2.0,
mode='fan_in',
distribution='uniform')
init_b = tf.constant_initializer(0.01)
init_z = tf.zeros_initializer()
with tf.variable_scope(scope):
dat = self.dat = tf.constant(name='dat', value=dat)
bs_ = self.bs_ = tf.placeholder(name='bs_',
dtype=tf.int32,
shape=())
bat = self.bat = tf.random_uniform(name='bat',
shape=(bs_, ),
minval=0,
maxval=dat.shape[0],
dtype=tf.int32)
h = x = self.x = tf.nn.embedding_lookup(name='x',
params=dat,
ids=bat)
for i, dim in enumerate(dim_rec, 1):
name = "hr{}".format(i)
h = tf.layers.dense(name=name,
inputs=h,
units=dim,
activation=tf.nn.relu,
kernel_initializer=init_w,
bias_initializer=init_b)
setattr(self, name, h)
mu = self.mu = tf.layers.dense(name='mu',
inputs=h,
units=dim_z,
kernel_initializer=init_w,
bias_initializer=init_z)
lv = self.lv = tf.layers.dense(name='lv',
inputs=h,
units=dim_z,
kernel_initializer=init_w,
bias_initializer=init_z)
with tf.name_scope('z'):
h = z = self.z = mu + tf.exp(
0.5 * lv) * tf.random_normal(shape=tf.shape(lv))
for i, dim in enumerate(dim_gen, 1):
name = "hg{}".format(i)
h = tf.layers.dense(name=name,
inputs=h,
units=dim,
activation=tf.nn.relu,
kernel_initializer=init_w,
bias_initializer=init_b)
setattr(self, name, h)
logits = tf.layers.dense(
name='logits',
inputs=h,
units=dat.shape[1]
# , activation= tf.nn.sigmoid
,
kernel_initializer=init_w,
bias_initializer=init_z)
g = self.g = tf.sigmoid(logits)
with tf.name_scope('loss_recons'):
# loss_recons = self.loss_recons = tf.reduce_mean(
# tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(labels= x, logits= logits), axis= 1))
loss_recons = self.loss_recons = tf.reduce_mean(
tf.reduce_sum(tf.square(x - g), axis=1))
with tf.name_scope('loss_relent'):
# loss_relent = self.loss_relent = tf.reduce_mean(
# 0.5 * tf.reduce_sum((- 1.0 - lv + tf.exp(lv) + tf.square(mu)), axis= 1))
loss_relent = self.loss_relent = tf.reduce_mean(
tf.reduce_sum((-1.0 - lv + tf.exp(lv) + tf.square(mu)),
axis=1))
with tf.name_scope('loss'):
loss = self.loss = loss_relent + loss_recons
up = self.up = tf.train.AdamOptimizer().minimize(loss)
self.step = 0
def __init__(self, *, use_dice=False):
self.graph = tf.Graph()
self.tensor_info = {}
self.use_dice = use_dice
with self.graph.as_default():
# Main Inputs
with tf.name_scope('Main_Inputs'):
self.target_ph = tf.placeholder(tf.float32, [None, None], name='target_ph')
self.lr_ph = tf.placeholder(tf.float32, [], name="lr_ph")
# with tf.name_scope("Teacher_Info"):
self.teacher_id_ph = tf.placeholder(tf.int32, [None, ], name="teacher_id_ph")
self.student_count_ph = tf.placeholder(tf.int32, [None, ], name="student_count_ph")
self.province_id_ph = tf.placeholder(tf.int32, shape=[None, ], name="province_id_ph")
self.city_id_ph = tf.placeholder(tf.int32, shape=[None, ], name="city_id_ph")
# TODO: binary
self.core_type_ph = tf.placeholder(tf.int32, shape=[None, ], name="core_type_ph")
# with tf.name_scope("Class_Info"):
self.class_id_ph = tf.placeholder(tf.int32, [None, ], name="class_id_ph")
self.edition_id_ph = tf.placeholder(tf.int32, [None, ], name="edition_id_ph")
self.grade_id_ph = tf.placeholder(tf.int32, [None, ], name="grade_id_ph")
self.class_student_ph = tf.placeholder(tf.int32, [None, ], name="class_student_ph")
self.cap_avg_ph = tf.placeholder(tf.float32, [None, ], name="cap_avg_ph")
self.cap_max_ph = tf.placeholder(tf.float32, [None, ], name="cap_max_ph")
self.cap_min_ph = tf.placeholder(tf.float32, [None, ], name="cap_min_ph")
# with tf.name_scope("Homework_Info"):
self.today_chapters_ph = tf.placeholder(tf.int32, [None, None], name="today_chapters_ph")
self.today_sections_ph = tf.placeholder(tf.int32, [None, None], name="today_sections_ph")
self.today_chap_mask_ph = tf.placeholder(tf.float32, [None, None], name='today_chap_mask_ph')
self.today_chap_len_ph = tf.placeholder(tf.int32, [None, ], name='today_chap_len_ph')
self.today_sec_mask_ph = tf.placeholder(tf.float32, [None, None], name='today_sec_mask_ph')
self.today_sec_len_ph = tf.placeholder(tf.int32, [None, ], name='today_sec_len_ph')
self.today_style_ph = tf.placeholder(tf.int32, [None, ], name='today_style_ph')
# TODO: use three dims to capture more history info
for fir in ['one', 'two', 'three', 'four','five','six','seven','eight','nine','ten','eleven','twelve','thirteen','fourteen']:
key = "history_" + fir + "_chap_ph"
setattr(self, key,
tf.placeholder(tf.int32, [None, None], name=key))
for fir in ['one', 'two', 'three', 'four','five','six','seven','eight','nine','ten','eleven','twelve','thirteen','fourteen']:
key = "history_" + fir + "_sec_ph"
setattr(self, key,
tf.placeholder(tf.int32, [None, None], name=key))
# TODO: All belows should consider the type and input
# with tf.name_scope("Study_Info"):
# TODO: study_vector_ph's type can change?
self.study_vector_ph = tf.placeholder(tf.float32, [None, 20], name="study_vector_ph")
self.gap_days_ph = tf.placeholder(tf.int32, [None, ], name="gap_days_ph")
self.month_submit_rate_ph = tf.placeholder(tf.float32, [None, ], name="month_submit_rate_ph")
# with tf.name_scope("Capacity_Info"):
self.region_capacity_ph = tf.placeholder(tf.float32, [None, ], name="region_capacity_ph")
# with tf.name_scope("Prefer_Info"):
self.prefer_assign_time_avg_ph = tf.placeholder(tf.float32, [None, ],
name="prefer_assign_time_avg_ph")
self.prefer_assign_time_var_ph = tf.placeholder(tf.float32, [None, ],
name="prefer_assign_time_var_ph")
self.prefer_assign_rank_avg_ph = tf.placeholder(tf.float32, [None, ],
name="prefer_assign_rank_avg_ph")
self.prefer_assign_rank_var_ph = tf.placeholder(tf.float32, [None, ],
name="prefer_assign_rank_var_ph")
# with tf.name_scope("Register_Info"):
self.register_diff_ph = tf.placeholder(tf.int32, [None, ], name="register_diff_ph")
# with tf.name_scope("HomeworkCount_Info"):
self.homework_count_ph = tf.placeholder(tf.int32, [None, ], name="homework_count_ph")
for fir in ["1", "2", "3", "4"]:
for sec in ["100", "010", "001", "110", "101", "011", "111"]:
key = "style_" + fir + "0" + sec + "_ph"
setattr(self, key,
tf.placeholder(tf.int32, [None, ], name=key))
# with tf.name_scope("WeekHomeworkCount_Info"):
self.week_count_ph = tf.placeholder(tf.float32, [None, ], name="week_count_ph")
# with tf.name_scope("Reflect_Info"):
# TODO: explore more graceful
self.reflect_value_ph = tf.placeholder(tf.int32, [None, None], name="reflect_value_ph")
self.reflect_mask_ph = tf.placeholder(tf.float32, [None, None], name="reflect_mask_ph")
self.reflect_len_ph = tf.placeholder(tf.int32, [None, ], name="reflect_len_ph")
# with tf.name_scope("Lastdat_Info"):
self.lastday_count_ph = tf.placeholder(tf.int32, [None, ], name="lastday_count_ph")
# Embedding layer
with tf.name_scope('Main_Embedding_layer'):
# almost done
with tf.name_scope("Others"):
# teacher
with tf.name_scope("Teacher"):
self.teacher_id_embeddings_var = tf.get_variable("teacher_id_embeddings_var",
[N_TEACHER, EMBEDDING_DIM], )
# tf.summary.histogram('teacher_id_embeddings_var', self.teacher_id_embeddings_var)
self.teacher_id_embedded = tf.nn.embedding_lookup(self.teacher_id_embeddings_var,
self.teacher_id_ph, )
self.province_id_embeddings_var = tf.get_variable("province_id_embeddings_var",
[N_PROVINCE, EMBEDDING_DIM])
# tf.summary.histogram('province_id_embeddings_var', self.province_id_embeddings_var)
self.province_id_embedded = tf.nn.embedding_lookup(self.province_id_embeddings_var,
self.province_id_ph)
self.city_id_embeddings_var = tf.get_variable("city_id_embeddings_var",
[N_CITY, EMBEDDING_DIM])
# tf.summary.histogram('city_id_embeddings_var', self.city_id_embeddings_var)
self.city_id_embedded = tf.nn.embedding_lookup(self.city_id_embeddings_var,
self.city_id_ph)
self.core_type_embeddings_var = tf.get_variable("core_type_embeddings_var",
[2, EMBEDDING_DIM])
# tf.summary.histogram('core_type_embeddings_var', self.core_type_embeddings_var)
self.core_type_embedded = tf.nn.embedding_lookup(self.core_type_embeddings_var,
self.core_type_ph)
# just to use embedded for var,maybe tf.identify?
self.student_count_embedded = get_self_or_expand_dims(self.student_count_ph)
with tf.name_scope("Class"):
self.class_id_embeddings_var = tf.get_variable("class_id_embeddings_var",
[N_CLASS, EMBEDDING_DIM])
# tf.summary.histogram('class_id_embeddings_var', self.class_id_embeddings_var)
self.class_id_embedded = tf.nn.embedding_lookup(self.class_id_embeddings_var,
self.class_id_ph)
self.edition_id_embeddings_var = tf.get_variable("edition_id_embeddings_var",
[N_EDITION, EMBEDDING_DIM])
# tf.summary.histogram('edition_id_embeddings_var', self.edition_id_embeddings_var)
self.edition_id_embedded = tf.nn.embedding_lookup(self.edition_id_embeddings_var,
self.edition_id_ph)
self.grade_id_embeddings_var = tf.get_variable("grade_id_embeddings_var",
[N_GRADE, EMBEDDING_DIM])
# tf.summary.histogram('grade_id_embeddings_var', self.grade_id_embeddings_var)
self.grade_id_embedded = tf.nn.embedding_lookup(self.grade_id_embeddings_var,
self.grade_id_ph)
# just to use embedded for var,maybe tf.identify?
#连续值 dense 本身有意义的直接喂入
self.class_student_embedded = get_self_or_expand_dims(self.class_student_ph)
self.cap_avg_embedded = get_self_or_expand_dims(self.cap_avg_ph)
self.cap_max_embedded = get_self_or_expand_dims(self.cap_max_ph)
self.cap_min_embedded = get_self_or_expand_dims(self.cap_min_ph)
with tf.name_scope("Study"):
# just to use embedded for var,maybe tf.identify?
self.study_vector_embedded = self.study_vector_ph
self.gap_days_embedded = get_self_or_expand_dims(self.gap_days_ph)
with tf.name_scope("Submit"):
# just to use embedded for var,maybe tf.identify?
self.month_submit_rate_embedded = get_self_or_expand_dims(self.month_submit_rate_ph)
with tf.name_scope("Capacity"):
# just to use embedded for var,maybe tf.identify?
self.region_capacity_embedded = get_self_or_expand_dims(self.region_capacity_ph)
with tf.name_scope("Prefer"):
# just to use embedded for var,maybe tf.identify?
self.prefer_assign_time_avg_embedded = get_self_or_expand_dims(
self.prefer_assign_time_avg_ph)
self.prefer_assign_time_var_embedded = get_self_or_expand_dims(
self.prefer_assign_time_var_ph)
self.prefer_assign_rank_avg_embedded = get_self_or_expand_dims(
self.prefer_assign_rank_avg_ph)
self.prefer_assign_rank_var_embedded = get_self_or_expand_dims(
self.prefer_assign_rank_var_ph)
with tf.name_scope("Register"):
self.register_diff_embedded = get_self_or_expand_dims(self.register_diff_ph)
with tf.name_scope("HomeworkCount"):
self.homework_count_embedded = get_self_or_expand_dims(self.homework_count_ph)
with tf.name_scope("WeekHomeworkCount"):
self.week_count_embedded = get_self_or_expand_dims(self.week_count_ph)
with tf.name_scope("Lastday"):
self.lastday_count_embedded = get_self_or_expand_dims(self.lastday_count_ph)
# TODO: homework and reflect and style
with tf.name_scope("Style"):
for fir in ["1", "2", "3", "4"]:
for sec in ["100", "010", "001", "110", "101", "011", "111"]:
key = "style_" + fir + "0" + sec + "_ph"
embed_key = "style_" + fir + "0" + sec + "_embedded"
setattr(self, embed_key,
get_self_or_expand_dims(getattr(self, key)))
# homework
with tf.name_scope("Homework"):
self.style_embeddings_var = tf.get_variable("style_embeddings_var",
[N_STYLE, EMBEDDING_DIM])
self.chapters_embeddings_var = tf.get_variable("chapters_embeddings_var",
[N_CHAPTER, EMBEDDING_DIM])
self.sections_embeddings_var = tf.get_variable("sections_embeddings_var",
[N_SECTION, EMBEDDING_DIM])
# tf.summary.histogram('homework_embeddings_var', self.homework_embeddings_var)
self.today_chapters_embedded = get_mask_zero_embedded(self.chapters_embeddings_var,
self.today_chapters_ph)
self.today_sections_embedded = get_mask_zero_embedded(self.sections_embeddings_var,
self.today_sections_ph)
self.history_chap_embedded,self.history_sec_embedded = get_history_bgru_embedded(self)
self.today_style_embedded = tf.nn.embedding_lookup(self.style_embeddings_var,
self.today_style_ph)
# reflect
with tf.name_scope("Reflect"):
self.reflect_embeddings_var = tf.get_variable("reflect_embeddings_var",
[N_REFLECT, EMBEDDING_DIM])
# tf.summary.histogram('reflect_embeddings_var', self.reflect_embeddings_var)
self.reflect_value_embedded = get_mask_zero_embedded(self.reflect_embeddings_var,
self.reflect_value_ph)