def __call__(self, inputs, state, scope=None):
"""
:param inputs: [N, d + JQ + JQ * d]
:param state: [N, d]
:param scope:
:return:
"""
with tf.variable_scope(scope or self.__class__.__name__):
c_prev, h_prev = state
x = tf.slice(inputs, [0, 0], [-1, self._input_size])
q_mask = tf.slice(inputs, [0, self._input_size],
[-1, self._q_len]) # [N, JQ]
qs = tf.slice(inputs, [0, self._input_size + self._q_len],
[-1, -1])
qs = tf.reshape(qs,
[-1, self._q_len, self._input_size]) # [N, JQ, d]
x_tiled = tf.tile(tf.expand_dims(x, 1),
[1, self._q_len, 1]) # [N, JQ, d]
h_prev_tiled = tf.tile(tf.expand_dims(h_prev, 1),
[1, self._q_len, 1]) # [N, JQ, d]
f = tf.tanh(
linear([qs, x_tiled, h_prev_tiled],
self._input_size,
True,
scope='f')) # [N, JQ, d]
a = tf.nn.softmax(
exp_mask(linear(f, 1, True, squeeze=True, scope='a'),
q_mask)) # [N, JQ]
q = tf.reduce_sum(qs * tf.expand_dims(a, -1), 1)
z = tf.concat(axis=1, values=[x, q]) # [N, 2d]
return self._cell(z, state)
def attention(query,
keys,
key_mask,
dim_query,
dim_key,
dtype=None,
scope=None):
with ops.variable_scope(scope or "attention", dtype=dtype):
# content-based addressing
# e_i = v_a^T tanh(W query + key_i)
# alpha = softmax({e_i})
# (n_query, dim_query) -> (n_query, dim_key)
mapped_query = nn.linear(query, [dim_query, dim_key],
False,
scope="map-query")
# (n_key, n_query, dim_key)
act = T.tanh(mapped_query[None, :, :] + keys)
# (n_key, n_query, 1)
e = nn.linear(act, [dim_key, 1], False,
scope="pre-alpha") # (n_key, n_query, 1)
# (n_key, n_query)
e = T.reshape(e, e.shape[:2])
e = e.T # (n_query, n_key)
# match dimension
key_mask = key_mask.T
alpha = nn.masked_softmax(e, key_mask) # (n_query, n_key)
alpha = alpha.T # (n_key, n_query)
return alpha
def coarseattention(query,
keys,
key_mask,
dim_query,
dim_key,
dtype=None,
scope=None):
with ops.variable_scope(scope or "coarseattention", dtype=dtype):
# content-based addressing
# e_i = v_a^T tanh(W query + key_i)
# alpha = softmax({e_i})
# (n_query, dim_query) -> (n_query, dim_key)
e = []
for i in range(len(keys)):
mapped_query = nn.linear(query, [dim_query, dim_key[i]],
False,
scope="map-query_{}".format(i))
# (n_key, n_query, dim_key)
act = T.tanh(mapped_query[None, :, :] + keys[i])
# (n_key, n_query, 1)
em = nn.linear(
act, [dim_key[i], 1], False,
scope="pre-alpha_{}".format(i)) # (n_key, n_query, 1)
e.append(em)
e = reduce(T.add, e)
# (n_key, n_query)
e = T.reshape(e, e.shape[:2])
e = e.T # (n_query, n_key)
# match dimension
key_mask = key_mask.T
alpha = nn.masked_softmax(e, key_mask) # (n_query, n_key)
alpha = alpha.T # (n_key, n_query)
return alpha
def attention(query,
mapped_states,
state_size,
attn_size,
attention_mask=None,
scope=None):
with ops.variable_scope(scope or "attention"):
mapped_query = nn.linear(query, [state_size, attn_size],
False,
scope="query_w")
mapped_query = mapped_query[None, :, :]
hidden = theano.tensor.tanh(mapped_query + mapped_states)
score = nn.linear(hidden, [attn_size, 1], False, scope="attention_v")
score = score.reshape([score.shape[0], score.shape[1]])
exp_score = theano.tensor.exp(score)
if attention_mask is not None:
exp_score = exp_score * attention_mask
alpha = exp_score / theano.tensor.sum(exp_score, 0)
return alpha
def prediction(self, y_emb, state, context, keep_prob=1.0):
"""
maxout -> readout -> softmax
p(y_j) \propto f(y_{j-1}, s_{j-1}, c_{j})
:param y_emb:
:param state:
:param context:
:param keep_prob:
:return:
"""
features = [state, y_emb, context]
maxhid = nn.maxout(
features,
[[self.dim_hid, self.dim_y, self.dim_value], self.dim_maxout],
self.max_part, True)
readout = nn.linear(maxhid, [self.dim_maxout, self.dim_readout],
False,
scope="readout")
if keep_prob < 1.0:
readout = nn.dropout(readout, keep_prob=keep_prob)
logits = nn.linear(readout, [self.dim_readout, self.n_y_vocab],
True,
scope="logits")
if logits.ndim == 3:
new_shape = [logits.shape[0] * logits.shape[1], -1]
logits = logits.reshape(new_shape)
probs = T.nnet.softmax(logits)
return probs
def __build_decoder(self, dec_inps, total_attn_states, total_mask,
global_trace, initial_state, genre, topic_trace):
with variable_scope.variable_scope("seq2seq_Decoder"):
dec_cell = tf.nn.rnn_cell.GRUCell(self.hps.hidden_size)
dec_cell = tf.nn.rnn_cell.DropoutWrapper(
dec_cell,
output_keep_prob=self.keep_prob,
input_keep_prob=self.keep_prob)
output_size = self.hps.vocab_size # num_decoder_symbols is vocabulary size
if not total_attn_states.get_shape()[1:3].is_fully_defined():
raise ValueError(
"Shape[1] and [2] of attn_states must be known: %s" %
total_attn_states.get_shape())
dec_outs, dec_states, attn_weights = [], [], []
normed_dec_outs = []
# build attn_mask
imask = 1.0 - total_mask
attn_mask = tf.where(tf.equal(imask, 1),
tf.ones_like(imask) * (-float('inf')), imask)
calcu_attention = lambda query: self.__attention_calcu(
total_attn_states, query, attn_mask, "calcu_attention")
state = initial_state
decoder_input_size = initial_state.get_shape()[1].value
for i, inp in enumerate(dec_inps):
if i > 0:
variable_scope.get_variable_scope().reuse_variables()
input_size = inp.get_shape().with_rank(2)[1]
if input_size.value is None:
raise ValueError(
"Could not infer input size from input: %s" % inp.name)
# Read memory
attns, align = calcu_attention(
[flatten_query(state), global_trace, topic_trace])
with variable_scope.variable_scope("input_merge"):
x = linear([inp, attns, genre[i], global_trace],
decoder_input_size, True)
# Run the GRU
cell_out, state = dec_cell(x, state)
with variable_scope.variable_scope("OutputProjection"):
output = linear(cell_out, output_size, True)
dec_outs.append(tf.identity(output))
normed_dec_outs.append(tf.identity(tf.nn.softmax(output)))
attn_weights.append([align])
dec_states.append(tf.identity(state))
return normed_dec_outs, dec_outs, dec_states, attn_weights
def fuse_gate_v2(lhs, rhs, is_train=None, wd=0.0, input_keep_prob=1.0, scope=None):
with tf.variable_scope(scope or "fuse_gate"):
dim = lhs.get_shape().as_list()[-1]
lhs_2 = linear(lhs, dim, True, bias_start=0.0, scope="lhs_2", squeeze=False, wd=wd, input_keep_prob=input_keep_prob, is_train=is_train)
rhs_2 = linear(rhs, dim, True, bias_start=0.0, scope="rhs_2", squeeze=False, wd=wd, input_keep_prob=input_keep_prob, is_train=is_train)
f = tf.sigmoid(lhs_2 + rhs_2)
out = f * lhs + (1 - f) * rhs
return out
def fuse_gate(config, is_train, lhs, rhs, scope=None):
with tf.variable_scope(scope or "fuse_gate"):
dim = lhs.get_shape().as_list()[-1]
lhs_1 = linear(lhs, dim ,True, bias_start=0.0, scope="lhs_1", squeeze=False, wd=config.wd, input_keep_prob=config.input_keep_prob, is_train=is_train)
rhs_1 = linear(rhs, dim ,True, bias_start=0.0, scope="rhs_1", squeeze=False, wd=0.0, input_keep_prob=config.input_keep_prob, is_train=is_train)
z = tf.tanh(lhs_1 + rhs_1)
lhs_2 = linear(lhs, dim ,True, bias_start=0.0, scope="lhs_2", squeeze=False, wd=config.wd, input_keep_prob=config.input_keep_prob, is_train=is_train)
rhs_2 = linear(rhs, dim ,True, bias_start=0.0, scope="rhs_2", squeeze=False, wd=config.wd, input_keep_prob=config.input_keep_prob, is_train=is_train)
f = tf.sigmoid(lhs_2 + rhs_2)
out = f * lhs + (1 - f) * z
return out
def __call__(self, inputs, is_training):
with tf.variable_scope('decoder'):
dec_l1 = nn.linear(inputs, (self.z_dim, 1000), 'dec_l1')
dec_b1 = nn.batch_normalization(dec_l1, 1000, 'dec_b1', is_training)
dec_r1 = nn.leaky_relu(dec_b1)
dec_l2 = nn.linear(dec_r1, (1000, 1000), 'dec_l2')
dec_b2 = nn.batch_normalization(dec_l2, 1000, 'dec_b2', is_training)
dec_r2 = nn.leaky_relu(dec_b2)
dec_l3 = nn.linear(dec_r2, (1000, 784), 'dec_l3')
return tf.sigmoid(dec_l3)
def domain_sensitive_attention(keys, key_mask, dim_key, dim_domain, dtype=None, scope=None):
with ops.variable_scope(scope or "domain_sensitive_attention", dtype=dtype):
mapped_keys = nn.linear(keys, [dim_key, dim_domain], True, scope="map-key")
act = T.tanh(mapped_keys)
# (n_key, n_query, 1)
e = nn.linear(act, [dim_domain, 1], False, scope="pre-alpha") # (n_key, n_query, 1)
# (n_key, n_query)
e = T.reshape(e, e.shape[:2])
e = e.T # (n_query, n_key)
# match dimension
key_mask = key_mask.T
alpha = nn.masked_softmax(e, key_mask) # (n_query, n_key)
alpha = alpha.T # (n_key, n_query)
return alpha
def __call__(self, inputs, is_training):
with tf.variable_scope('encoder'):
enc_l1 = nn.linear(inputs, (784, 1000), 'enc_l1')
enc_b1 = nn.batch_normalization(enc_l1, 1000, 'enc_b1', is_training)
enc_r1 = nn.leaky_relu(enc_b1)
enc_l2 = nn.linear(enc_r1, (1000, 1000), 'enc_l2')
enc_b2 = nn.batch_normalization(enc_l2, 1000, 'enc_b2', is_training)
enc_r2 = nn.leaky_relu(enc_b2)
enc_l3 = nn.linear(enc_r2, (1000, self.z_dim), 'enc_l3')
encoded = nn.batch_normalization(enc_l3, self.z_dim, 'enc_b3', is_training)
return encoded
def __call__(self, input1, input2, scope=None, input_keep_prob=1.0, wd=0.0, is_train=None):
N = tf.shape(input1)[0]
d1 = input1.get_shape().as_list()[-1]
d2 = input2.get_shape().as_list()[-1]
with tf.variable_scope(scope or 'ntn'):
out1 = tf.reshape(tf.matmul(tf.reshape(linear(input1, self.rank * d2, False, scope='out1',
input_keep_prob=input_keep_prob, wd=wd, is_train=is_train),
[N, self.rank, d2]),
tf.expand_dims(input2, -1)),
[N, self.rank]) # [N, rank]
out2 = linear(tf.concat([input1, input2], -1), self.rank, True,
scope='out2', input_keep_prob=input_keep_prob, wd=wd, is_train=is_train)
out = tf.nn.tanh(out1 + out2)
out = linear(out, 1, False, scope='out', input_keep_prob=input_keep_prob, wd=wd, is_train=is_train)
return out
def __build_key_memory(self):
#print ("memory")
key_states = []
with variable_scope.variable_scope("EncoderRNN"):
for i in xrange(0, self.hps.key_slots):
if i > 0:
variable_scope.get_variable_scope().reuse_variables()
(outputs, state_fw,
state_bw) = rnn.static_bidirectional_rnn(self.enc_cell_fw,
self.enc_cell_bw,
self.emb_key_inps[i],
dtype=tf.float32)
key_state = array_ops.concat([state_fw, state_bw], 1)
key_states.append(key_state)
with variable_scope.variable_scope("key_memory"):
key_states = [
array_ops.reshape(e, [-1, 1, self.enc_cell_fw.output_size * 2])
for e in key_states
]
key_states = array_ops.concat(key_states, 1)
key_states = tf.multiply(self.key_mask, key_states)
final_state = math_ops.reduce_mean(key_states, axis=1)
final_state = linear(final_state,
self.hps.hidden_size,
True,
scope="key_initial")
final_state = tf.tanh(final_state)
return final_state, key_states
def __build_encoder(self, step):
with variable_scope.variable_scope("EncoderRNN", reuse=True):
(outputs, enc_state_fw,
enc_state_bw) = rnn.static_bidirectional_rnn(
self.enc_cell_fw,
self.enc_cell_bw,
self.emb_enc_inps[step][:self.enc_len],
dtype=tf.float32)
enc_outs = outputs
with variable_scope.variable_scope("seq2seq_Encoder"):
enc_state = enc_state_bw
final_state = linear(enc_state,
self.hps.hidden_size,
True,
scope="enc_initial")
final_state = tf.tanh(final_state)
top_states = [
array_ops.reshape(e, [-1, 1, self.enc_cell_fw.output_size * 2])
for e in enc_outs
]
attention_states = array_ops.concat(top_states, 1)
final_attn_states = tf.multiply(self.enc_mask[step],
attention_states)
return final_state, final_attn_states, enc_outs
def prediction(self, y_emb, state, context, y_pos, keep_prob=1.0):
"""
readout -> softmax
p(y_j) \propto f(y_{j-1}, s_{j}, c_{j})
:param y_pos:
:param y_emb:
:param state:
:param context:
:param keep_prob:
:return:
"""
state = nn.feedforward([state, y_pos], [[self.dim_hid, self.poshdim], self.dim_hid], True,
activation=T.tanh, scope="enhancedstate")
features = [state, y_emb, context, y_pos]
readout = nn.feedforward(features, [[self.dim_hid, self.dim_y, self.dim_value, self.poshdim], self.dim_readout], True,
activation=T.tanh,
scope="readout")
if keep_prob < 1.0:
readout = nn.dropout(readout, keep_prob=keep_prob)
logits = nn.linear(readout, [self.dim_readout, self.n_y_vocab], True,
scope="logits")
if logits.ndim == 3:
new_shape = [logits.shape[0] * logits.shape[1], -1]
logits = logits.reshape(new_shape)
probs = T.nnet.softmax(logits)
return probs
def __call__(self, inputs, state, scope=None):
if not isinstance(inputs, (list, tuple)):
inputs = [inputs]
input_size = self.input_size
output_size = self.output_size
if len(inputs) != len(input_size):
raise RuntimeError("unmatched elements: inputs and input_size")
size = [list(input_size) + [output_size], 4 * output_size]
with variable_scope(scope or "lstm"):
c, h = state
new_inputs = list(inputs[:]) + [h]
concat = linear(new_inputs, size, True, concat=True, scope="gates")
i, j, f, o = theano.tensor.split(concat, [output_size] * 4, 4, -1)
j = theano.tensor.tanh(j)
# input, forget, output gate
i = theano.tensor.nnet.sigmoid(i)
f = theano.tensor.nnet.sigmoid(f)
o = theano.tensor.nnet.sigmoid(o)
new_c = c * f + i * j
# no output activation
new_h = new_c * o
new_state = (new_c, new_h)
return new_h, new_state
def map_attention_states(attention_states, input_size, attn_size, scope=None):
with ops.variable_scope(scope or "attention"):
mapped_states = nn.linear(attention_states, [input_size, attn_size],
False,
scope="attention_w")
return mapped_states
def sampleP_theta(self,h_dec,share=None):
# H_DEC->X
with tf.variable_scope("P_theta",reuse=share):
p=nn.linear(h_dec,self.x_dim_flat)
x= tf.sigmoid(p) # mean of bernoulli distribution
x=tf.reshape(x,(-1,)+self.x_dim) # reshape to original image dimensions
return x
def __global_trace_update(self, global_trace, enc_states): #enc_states [batch_size,time,self.enc_cell_fw.output_size*2]
with variable_scope.variable_scope("global_trace_update", reuse=None):
state = math_ops.reduce_mean(enc_states, axis=1) #[batch_size,self.enc_cell_fw.output_size*2]
new_global_trace = math_ops.tanh(linear([global_trace, state],
self.hps.global_trace_size, True)) #[batch_size,global_trace_size]
new_global_trace = array_ops.reshape(new_global_trace, [-1, self.hps.global_trace_size])
return new_global_trace
def __call__(self, input, scope=None, input_keep_prob=1.0, wd=0.0, is_train=None):
N = tf.shape(input)[0]
with tf.variable_scope(scope or "maxout"):
out = linear(input, self.num_features * self.dim, False, scope='out')
out = tf.reshape(out, [N, self.num_features, self.dim])
out = tf.reduce_max(out, 1) # [N, dim]
return out
def __attention_calcu(self, attentions, query, attn_mask, scope):
with variable_scope.variable_scope(scope):
attn_length = attentions.get_shape(
)[1].value # the length of a input sentence
attn_size = attentions.get_shape(
)[2].value # hidden state size of encoder, that is 2*size
# To calculate W1 * h_t we use a 1-by-1 convolution, need to reshape before.
# Remember we use bidirectional RNN
hidden = array_ops.reshape(attentions,
[-1, attn_length, 1, attn_size])
# Size of query vectors for attention
# query vector is decoder state
attention_vec_size = attn_size
k = variable_scope.get_variable(
"AttnW", [1, 1, attn_size, attention_vec_size])
hidden_features = nn_ops.conv2d(hidden, k, [1, 1, 1, 1], "SAME")
v = variable_scope.get_variable("AttnV", [attention_vec_size])
y = linear(query, attention_vec_size, True)
y = array_ops.reshape(y, [-1, 1, 1, attention_vec_size])
# Attention mask is a softmax of v^T * tanh(...).
s = math_ops.reduce_sum(v * math_ops.tanh(hidden_features + y),
[2, 3])
a = nn_ops.softmax(s + attn_mask)
# Now calculate the attention-weighted vector d.
d = math_ops.reduce_sum(
array_ops.reshape(a, [-1, attn_length, 1, 1]) * hidden, [1, 2])
d = array_ops.reshape(d, [-1, attn_size]) #remember this size
return d, a
def __call__(self, inputs, state, scope=None):
with tf.variable_scope(scope or "gru_cell"):
with tf.variable_scope("gates"):
r_u = linear([inputs, state],
2 * self._num_units,
True,
scope=scope)
r, u = tf.split(r_u, 2, 1)
r, u = tf.nn.sigmoid(r), tf.nn.sigmoid(u)
with tf.variable_scope("candidate"):
c = self._activation(
linear([inputs, r * state],
self._num_units,
True,
scope=scope))
new_h = u * state + (1 - u) * c
return new_h, new_h
def __call__(self, inputs, state, scope=None):
with tf.variable_scope(scope or "basic_rnn_cell"):
output = linear([inputs, state],
self._num_units,
True,
scope=scope)
output = self._activation(output)
return output, output
def __global_trace_update(self, global_trace, enc_states):
with variable_scope.variable_scope("global_trace_update", reuse=None):
state = math_ops.reduce_mean(enc_states, axis=1)
new_global_trace = math_ops.tanh(
linear([global_trace, state], self.hps.global_trace_size,
True))
new_global_trace = array_ops.reshape(
new_global_trace, [-1, self.hps.global_trace_size])
return new_global_trace
def transition(self, z):
# compute A,B,o linearization matrices
# Z -> H_TRANS -> (A,B,o)
z_dim=self.z_dim
u_dim=self.u_dim
with tf.variable_scope("trans"):
h=self.dynamics(z)
with tf.variable_scope("A"):
v,r=tf.split(1,2,nn.linear(h,z_dim*2))
v1=tf.expand_dims(v,-1) # (batch, z_dim, 1)
rT=tf.expand_dims(r,1) # batch, 1, z_dim
I=tf.diag([1.]*z_dim)
A=(I+tf.batch_matmul(v1,rT)) # (z_dim, z_dim) + (batch, z_dim, 1)*(batch, 1, z_dim) (I is broadcasted)
with tf.variable_scope("B"):
B=nn.linear(h,z_dim*u_dim)
B=tf.reshape(B,[-1,z_dim,u_dim])
with tf.variable_scope("o"):
o=nn.linear(h,z_dim)
return A,B,o,v,r
def __call__(self, input1, input2, scope=None, input_keep_prob=1.0, wd=0.0, is_train=None, epsilon=1e-8):
N = tf.shape(input1)[0]
d1 = input1.get_shape().as_list()[-1]
d2 = input2.get_shape().as_list()[-1]
with tf.variable_scope(scope or "hyperbolic_distance"):
x = 1 + 2 * tf.divide(tf.norm(input1 - input2, axis=-1), \
(1 - tf.norm(input1, axis=-1)) * (1 - tf.norm(input2, axis=-1)) + epsilon)
out = tf.log(x + tf.sqrt(tf.square(x) - 1) + epsilon)
out = linear(tf.expand_dims(out, 1), 1, True, scope='out')
return tf.reshape(out, [N])
def prediction(prev_inputs, prev_state, context, keep_prob=1.0):
features = [prev_state, prev_inputs, context]
maxhid = nn.maxout(features, [[thdim, tedim, 2 * shdim], maxdim],
maxpart, True)
readout = nn.linear(maxhid, [maxdim, deephid], False,
scope="deepout")
if keep_prob < 1.0:
readout = nn.dropout(readout, keep_prob=keep_prob)
logits = nn.linear(readout, [deephid, tvsize], True,
scope="logits")
if logits.ndim == 3:
new_shape = [logits.shape[0] * logits.shape[1], -1]
logits = logits.reshape(new_shape)
probs = theano.tensor.nnet.softmax(logits)
return probs
def __topic_trace_update(self, topic_trace, key_align, key_states): #topic_trace [batch_size,topic_trace_size+key_slots] key_align [batch_size,key_slots] key_states [batch_size,key_slots,hidden_size*2]
with variable_scope.variable_scope("topic_trace_update"):
key_used = math_ops.reduce_mean(tf.multiply(key_states,
tf.expand_dims(key_align, axis=2)), axis=1) #[batch_size,key_slots,1] => [batch_size,key_slots,2*hidden_size] => [batch_size,2*hidden_size]
new_topic_trace = linear([topic_trace[:, 0:self.hps.topic_trace_size], key_used],
self.hps.topic_trace_size, True)
new_topic_trace = tf.tanh(new_topic_trace) #[batch_size,topic_trace_size]
attn_los = topic_trace[:, self.hps.topic_trace_size:] + key_align #[batch_size,key_slots]
fin_topic_trace = array_ops.concat([new_topic_trace, attn_los], 1)
return fin_topic_trace
def __topic_trace_update(self, topic_trace, key_align, key_states):
with variable_scope.variable_scope("topic_trace_update"):
key_used = math_ops.reduce_mean(tf.multiply(
key_states, tf.expand_dims(key_align, axis=2)),
axis=1)
new_topic_trace = linear(
[topic_trace[:, 0:self.hps.topic_trace_size], key_used],
self.hps.topic_trace_size, True)
new_topic_trace = tf.tanh(new_topic_trace)
attn_los = topic_trace[:, self.hps.topic_trace_size:] + key_align
fin_topic_trace = array_ops.concat([new_topic_trace, attn_los], 1)
return fin_topic_trace
def __call__(self, inputs, y, is_training):
"""
INPUTS:
z: latent space
y: one-hot vectors with shpae [num_class + 1] (+1: for unlabelled data)
RETURN:
`logits` (i.e. no activation function like sigmoid, softmax, ...)
"""
h = tf.concat([inputs, y], axis=1) # inputs's shape: (batch_size, z_dim + num_classes + 1)
with tf.variable_scope('discriminator'):
h = nn.linear(h, (self.z_dim + self.num_classes + 1, 500), 'disc_l1')
h = nn.batch_normalization(h, 500, 'disc_b1')
h = nn.leaky_relu(h)
h = nn.linear(h, (500, 500), 'disc_l2')
h = nn.batch_normalization(h, 500, 'disc_b2')
h = nn.leaky_relu(h)
logits = nn.linear(h, (500, 1), 'disc_l3')
return logits
def encode(self,x,share=None):
# X -> H_ENC
with tf.variable_scope("encoder",reuse=share):
for l in range(3):
x=nn.ReLU(x,5,"l"+str(l))
return nn.linear(x,2*self.z_dim)
def encode(self,x,share=None):
with tf.variable_scope("encoder",reuse=share):
l1=nn.ReLU(x,150,"l1")
l2=nn.ReLU(l1,150,"l2")
h_enc=nn.linear(l2,2*self.z_dim)
return h_enc
def decode(self, z, share=None):
with tf.variable_scope("decoder",reuse=share):
l1=nn.ReLU(z,200,"l1")
l2=nn.ReLU(l1,200,"l2")
h_dec=nn.linear(l2,self.x_dim_flat)
return h_dec