def _prediction_layer(self, inputs, task_idx, no_samples):
K = no_samples
act = tf.tile(tf.expand_dims(inputs, 0), [K, 1, 1])
weight_share = self.weights[0]
weight_head = self.weights[1]
for i in range(self.no_layers - 1):
din = self.size[i]
dout = self.size[i + 1]
#get w
list_mean_w = []
list_variance_w = []
list_coff = []
for mixture in range(self.gauss_mixture):
list_mean_w.append(weight_share[mixture][0][0][i])
list_variance_w.append(
tf.exp(
self.compute_weightfromfactorize_variance(
weight_share[mixture][1][0][i])))
list_coff.append(weight_share[mixture][2][0][i])
weights = sample_gumbel_trick.sample_from_gumbel_softmax_trick(
list_mean_w, list_variance_w, list_coff, self.tau,
self.gauss_mixture, K, din, dout, False)
list_mean_w = []
list_variance_w = []
list_coff = []
for mixture in range(self.gauss_mixture):
list_mean_w.append(weight_share[mixture][0][1][i])
list_variance_w.append(tf.exp(weight_share[mixture][1][1][i]))
list_coff.append(weight_share[mixture][2][1][i])
biases = sample_gumbel_trick.sample_from_gumbel_softmax_trick(
list_mean_w, list_variance_w, list_coff, self.tau,
self.gauss_mixture, K, din, dout, True)
#feed vao mang
pre = tf.add(tf.einsum('mni,mio->mno', act, weights), biases)
act = tf.nn.relu(pre)
#tai layer cuoi cung
din = self.size[-2]
dout = self.size[-1]
eps_w = tf.random_normal((K, din, dout), 0, 1, dtype=tf.float32)
eps_b = tf.random_normal((K, 1, dout), 0, 1, dtype=tf.float32)
Wtask_m = tf.gather(weight_head[0][0], task_idx)
Wtask_v = tf.gather(self.Wlast_v, task_idx)
btask_m = tf.gather(weight_head[0][1], task_idx)
btask_v = tf.gather(weight_head[1][0], task_idx)
weights = tf.add(tf.multiply(eps_w, tf.exp(0.5 * Wtask_v)), Wtask_m)
biases = tf.add(tf.multiply(eps_b, tf.exp(0.5 * btask_v)), btask_m)
act = tf.expand_dims(act, 3)
weights = tf.expand_dims(weights, 1)
pre = tf.add(tf.reduce_sum(act * weights, 2), biases)
return pre
def _prediction_layer(self, inputs, task_idx, no_samples):
K = no_samples
act = tf.tile(tf.expand_dims(inputs, 0), [K, 1, 1])
for i in range(self.no_layers - 1):
din = self.size[i]
dout = self.size[i + 1]
#get w
list_mean_w = []
list_variance_w = []
list_coff = []
for mixture in range(self.gauss_mixture):
list_mean_w.append(self.weights[mixture][0][0][i])
list_variance_w.append(tf.exp(self.weights[mixture][1][0][i]))
list_coff.append(self.weights[mixture][2][0][i])
weights = sample_gumbel_trick.sample_from_gumbel_softmax_trick(
list_mean_w, list_variance_w, list_coff, self.tau,
self.gauss_mixture, K, din, dout, False)
list_mean_w = []
list_variance_w = []
list_coff = []
for mixture in range(self.gauss_mixture):
list_mean_w.append(self.weights[mixture][0][1][i])
list_variance_w.append(tf.exp(self.weights[mixture][1][1][i]))
list_coff.append(self.weights[mixture][2][1][i])
biases = sample_gumbel_trick.sample_from_gumbel_softmax_trick(
list_mean_w, list_variance_w, list_coff, self.tau,
self.gauss_mixture, K, din, dout, True)
#feed vao mang
pre = tf.add(tf.einsum('mni,mio->mno', act, weights), biases)
act = tf.nn.relu(pre)
#tai layer cuoi cung
din = self.size[-2]
dout = self.size[-1]
Wtask_m , Wtask_v , btask_m , btask_v , coff_m , coff_b = [] , [] , [] , [] , [] , []
for mixture in range(self.gauss_mixture):
Wtask_m.append(self.weights[mixture][0][2][0])
Wtask_v.append(tf.exp(self.weights[mixture][1][2][0]))
btask_m.append(self.weights[mixture][0][3][0])
btask_v.append(tf.exp(self.weights[mixture][1][3][0]))
coff_m.append(self.weights[mixture][2][2][0])
coff_b.append(self.weights[mixture][2][3][0])
weights = sample_gumbel_trick.sample_from_gumbel_softmax_trick(
Wtask_m, Wtask_v, coff_m, self.tau, self.gauss_mixture, K, din,
dout, False)
biases = sample_gumbel_trick.sample_from_gumbel_softmax_trick(
btask_m, btask_v, coff_b, self.tau, self.gauss_mixture, K, din,
dout, True)
act = tf.expand_dims(act, 3)
weights = tf.expand_dims(weights, 1)
pre = tf.add(tf.reduce_sum(act * weights, 2), biases)
return pre