y = np.random.uniform(size=(batch_size, )).astype(np.float32)
return x, y
def __len__(self):
return 256
net_input = keras.layers.Input(shape=input_shape, dtype=np.float32)
net = keras.layers.Flatten()(net_input)
net = keras.layers.Dense(32, activation='relu')(net)
net = keras.layers.Dense(1, activation='sigmoid')(net)
model = keras.Model(net_input, net)
model.compile(optimizer='rmsprop', loss='mse')
generator = MyGenerator()
graph1 = hg.Graph(name="keras_exec_opt")
with graph1.as_default():
time_history = TimeHistory()
set_batch_size = hg.call(generator.set_batch_size) << [
hg.tweak(hg.QUniform(low=2, high=64), name='batch_size')
]
fit = hg.call(model.fit_generator) << {
'generator':
generator,
'epochs':
1,
'max_queue_size':
hg.tweak(hg.QUniform(low=2, high=64), name='max_queue_size'),
'workers':
hg.tweak(hg.QLogUniform(low=1, high=16), name='workers'),
'callbacks': [time_history],
import hypergraph as hg
def test1(x, y):
return x + y
graph1 = hg.Graph(name='g1')
with graph1.as_default():
hg.mark("x_value") << 5
hg.output() << (
hg.invoke() << [test1, {
'x': hg.node_ref("x_value"),
'y': 3
}])
# Note, in a more realistic scenario, the first param of the invoke, that is the function to be invoked,
# may be the result of another computation or tweak
print(hg.run(graph1))
import hypergraph as hg
import numpy as np
graph1 = hg.Graph()
with graph1.as_default():
hg.mark("abc1") << (hg.dump() << "** abc1 **")
n = hg.mark("abc2") << (hg.dump() << "** abc2 **")
idx = hg.node(lambda _: np.random.randint(0, 2))
hg.output() << (hg.select(idx) << ["abc1", "abc2"])
for _ in range(3):
ctx = hg.ExecutionContext()
with ctx.as_default():
print(graph1())
print("*** end of execution ***")
import hypergraph as hg
from hypergraph.genetic import GeneticOperators
graph1 = hg.Graph(name="g1")
with graph1.as_default():
hg.output() << (hg.permutation(size=3) << list('abcdef'))
# create a population from graph's phenotype
genetic = GeneticOperators(graph1)
print(genetic.phenotype)
population = genetic.create_population(3)
print(population[:2])
# crossover two parent to get a new individual
child = genetic.crossover_uniform_multi_parents(population[:2])
print(child)
genetic.mutations(child, prob=0.5)
# apply mutations to an individual
print(child)
# use an individual as graph's tweaks
print(hg.run(graph1, tweaks=child))