def decode(self, q, name=None):
qpoly = atu.polyexpand(q, self.Np_dec)
N_coeff = atu.Npolyexpand( self.size_q, self.Np_dec )
We1 = self._var("dec_W", (N_coeff, self.size_x) )
be1 = self._var("dec_b", (self.size_x,) )
x = tf.matmul( qpoly, We1 ) + be1
return tf.identity(x,name=name)
def encode(self, x, name=None):
N_coeff = atu.Npolyexpand( self.size_x, self.Np_enc )
if self.encoder_init in encoder_init_options.keys():
We1 = self._var("enc_W", (N_coeff, self.size_q),
initial_value = encoder_init_options[self.encoder_init])
else:
We1 = self._var("enc_W", (N_coeff, self.size_q) )
be1 = self._var("enc_b", (self.size_q,) )
q = tf.matmul( atu.polyexpand(x, self.Np_enc), We1 ) + be1
return tf.identity(q,name=name)
def decode(self, q, name=None, phase_act="softmax"):
qpoly = atu.polyexpand(q, self.Np_dec)
N_coeff = atu.Npolyexpand( self.size_q, self.Np_dec )
W1 = self._var("dec_W_curve", (N_coeff, self.N_curve, self.size_x) )
b1 = self._var("dec_b_curve", (self.N_curve, self.size_x) )
h_curve = tf.tensordot(qpoly,W1,axes=[-1,0])+b1
h_select = self.classify(q,phase_act=phase_act)
x = tf.einsum('ijk,ij->ik',h_curve,h_select)
return tf.identity(x,name=name)
def classify(self, q, name=None, phase_act="softmax"):
qpoly = atu.polyexpand(q, self.Np_dec)
N_coeff = atu.Npolyexpand( self.size_q, self.Np_dec )
W2 = self._var("dec_W_bound", (N_coeff, self.N_bound) )
b2 = self._var("dec_b_bound", (self.N_bound,) )
act = self.activations[self.boundary_activation]
h_bound = act(tf.tensordot(qpoly,W2,axes=[-1,0])+b2)
W_select = self._var("dec_W_select", (self.N_bound,self.N_curve))
b_select = self._var("dec_b_select",(self.N_curve,))
h_select = tf.tensordot(h_bound,W_select, axes=[-1,0]) + b_select
if phase_act=="softmax":
h_select = tf.nn.softmax(self.softmax_beta*h_select
,name=name)
else:
h_select = tf.one_hot(
tf.argmax(h_select,axis=-1),
depth=h_select.shape[-1],
dtype=h_select.dtype,name=name)
print("Did this.")
return h_select
def encode(self, x, name=None):
N_coeff = atu.Npolyexpand( self.size_x, self.Np_enc )
W = self._var("enc_W", (N_coeff, self.size_q),
initial_value=encoder_init_options[self.encoder_init])
b = self._var("enc_b", (self.size_q,) )
return tf.add(tf.matmul( atu.polyexpand(x, self.Np_enc), W ), b, name=name)
def decode(self, q, name=None):
N_coeff = atu.Npolyexpand( self.size_q, self.Np_dec )
nxt = atu.polyexpand(q, self.Np_dec)
nxt = self._layers(nxt, self.dec_layers + [self.size_x], prefix="dec")
return tf.identity(nxt, name=name)
def encode(self, x, name=None):
N_coeff = atu.Npolyexpand( self.size_x, self.Np_enc )
nxt = atu.polyexpand(x, self.Np_enc)
nxt = self._layers(nxt, self.enc_layers + [self.size_q], prefix="enc")
return tf.identity(nxt, name=name)