# outer objective (validation error) (not weighted)
val_loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=out))
accuracy = tf.reduce_mean(
tf.cast(tf.equal(tf.argmax(y, 1), tf.argmax(out, 1)), tf.float32))
# optimizers
# get an hyperparameter for the learning rate
lr = far.get_hyperparameter('lr', 0.01)
io_optim = far.GradientDescentOptimizer(
lr) # for training error minimization an optimizer from far_ho is needed
oo_optim = tf.train.AdamOptimizer(
) # for outer objective optimizer all optimizers from tf are valid
print('hyperparameters to optimize')
[print(h) for h in far.hyperparameters()]
# build hyperparameter optimizer
farho = far.HyperOptimizer()
run = farho.minimize(val_loss,
oo_optim,
tr_loss,
io_optim,
init_dynamics_dict={
v: h
for v, h in zip(tf.model_variables(),
far.utils.hyperparameters()[:4])
})
print(
'Variables (or tensors) that will store the values of the hypergradients')
self + residual_block(self.out, 256)
self + tcl.conv2d(
self.out, 2048, 1, variables_collections=self.var_coll)
self + tf.nn.avg_pool(self.out, [1, 6, 6, 1], [1, 6, 6, 1], 'SAME')
self + tcl.conv2d(
self.out, 512, 1, variables_collections=self.var_coll)
self + tf.reshape(self.out, (-1, 512))
def for_input(self, new_input):
return TCML_ResNet_Omniglot_v2(new_input, self.name,
self.deterministic_initialization, True)
def hr_res_net_tcml_Omniglot_builder_v2():
return lambda x, name: TCML_ResNet_Omniglot_v2(x, name=name)
def hr_res_net_tcml_v1_builder():
return lambda x, name: TCML_ResNet(x, name=name)
def hr_res_net_tcml_Omniglot_builder():
return lambda x, name: TCML_ResNet_Omniglot(x, name=name)
if __name__ == '__main__':
inp = tf.placeholder(tf.float32, (None, 84, 84, 3))
net = TCML_ResNet(inp)
print(net.out)
print(far.hyperparameters())
