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("gru", 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(biout, axis=-2)
return rnnoutput
def get_rnn_sum(input_seq,name='cha',nameln = 'sec'):
#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//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)
#这里的 name 和上面 一样吗，还是 每个不一样，再加个name2
# lnout = tf.contrib.layers.layer_norm(
# inputs=(biout + input_seq), begin_norm_axis=-1, begin_params_axis=-1, scope=nameln)
lnout = tf.contrib.layers.layer_norm(
inputs=(biout + input_seq), begin_norm_axis=-1, begin_params_axis=-1)
rnnoutput = tf.reduce_mean(lnout, axis=-2)
#attention_output = layer_norm(biout + input_seq)
return rnnoutput
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, 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= '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