def test_simple_network():
searcher = Searcher()
searcher.register_trial('hidden_size', [64, 128, 256, 512, 1024])
searcher.register_trial('batchsize', [32, 64, 128, 256, 512, 1024])
searcher.register_trial('lr', [0.05, 0.1, 0.2, 0.3, 0.4, 0.5])
n_trials = 30
model_kwargs = dict(
input_dim=3,
n_train_epochs=200,
)
_ = searcher.search(objectve,
n_trials=n_trials,
deep_surrogate_model='tfdbonas.deep_surrogate_models:SimpleNetwork',
n_random_trials=10,
model_kwargs=model_kwargs)
assert len(searcher.result) == n_trials
warnings.warn('results = {}'.format(searcher.result))
warnings.warn('best_trial {}'.format(searcher.best_trial))
warnings.warn('best_value {}'.format(searcher.best_value))
])
return m
if __name__ == '__main__':
m1 = create_modal()
m2 = create_modal()
m = m1 + m2
m.append_lazy(keras.layers.Dense, [
dict(units=10, activation='softmax'),
])
g = Generator(m, dump_nn_graph=True)
g.draw_graph('/home/ono/Dropbox/test.png')
num_nodes = 12
num_layer_type = 4
searcher = Searcher()
searcher.register_trial('graph', g)
n_trials = 30
model_kwargs = dict(
num_nodes=num_nodes,
input_channels=num_layer_type,
n_train_epochs=400,
)
_ = searcher.search(objectve,
n_trials=n_trials,
deep_surrogate_model=
f'tfdbonas.deep_surrogate_models:GCNSurrogateModel',
n_random_trials=10,
model_kwargs=model_kwargs)
print(searcher.result)
print('best_trial', searcher.best_trial)
#!/usr/bin/env python
import math
from tfdbonas import Searcher, Trial
import numpy as np
def objectve(trial: Trial):
# (x*, y*) = [1, 1]
# f(x*, y*) = 0
o = 100 * (trial.y - trial.x**2)**2 + (trial.x - 1)**2
return -o
if __name__ == '__main__':
searcher = Searcher()
searcher.register_trial('x', np.arange(-2.048, 2.048, 0.02))
searcher.register_trial('y', np.arange(-2.048, 2.048, 0.02))
model_kwargs = dict(
input_dim=2, # coresponding to the number of register_trial
n_train_epochs=200,
)
n_trials = 20
_ = searcher.search(
objectve,
n_trials=n_trials,
deep_surrogate_model='tfdbonas.deep_surrogate_models:SimpleNetwork',
n_random_trials=10,
model_kwargs=model_kwargs)
assert len(searcher.result) == n_trials
print('results = {}'.format(searcher.result))
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0
opt = tf.keras.optimizers.SGD(learning_rate=0.01) #trial.lr)
model.compile(optimizer=opt,
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)
model.fit(x_train, y_train, epochs=1, batch_size=128) #trial.batchsize)
out = model.evaluate(x_test, y_test, verbose=2)
accuracy = out[1]
return accuracy
if __name__ == '__main__':
searcher = Searcher()
searcher.register_trial('h1', [16, 32, 64, 128, 256, 512, 1024])
searcher.register_trial('h2', [16, 32, 64, 128, 256, 512, 1024])
searcher.register_trial('h3', [16, 32, 64, 128, 256, 512, 1024])
searcher.register_trial('h4', [16, 32, 64, 128, 256, 512, 1024])
# searcher.register_trial('batchsize', [32, 64, 128, 256, 512, 1024])
# searcher.register_trial('lr', [0.05, 0.1, 0.2, 0.3, 0.4, 0.5])
n_trials = 30
model_kwargs = dict(
input_dim=4,
n_train_epochs=200,
)
_ = searcher.search(
objectve,
n_trials=n_trials,
def test_simple_network_with_categorical_features():
# (False, reason="categorical feature is not supported yet. [idea] I shoud apply embedding layers for categorical inputs.")
searcher = Searcher()
searcher.register_trial('hidden_size', [64, 128, 256, 512, 1024])
searcher.register_trial('batchsize', [32, 64, 128, 256, 512, 1024])
searcher.register_trial('lr', [0.05, 0.1, 0.2, 0.3, 0.4, 0.5])
searcher.register_trial('activation', [None, 'relu', 'tanh'])
#
n_trials = 30
model_kwargs = dict(
input_dim=3,
n_train_epochs=200,
)
_ = searcher.search(objectve,
n_trials=n_trials,
deep_surrogate_model='tfdbonas.deep_surrogate_models:SimpleNetwork',
n_random_trials=10,
model_kwargs=model_kwargs)
assert len(searcher.result) == n_trials
warnings.warn('results = {}'.format(searcher.result))
warnings.warn('best_trial {}'.format(searcher.best_trial))
warnings.warn('best_value {}'.format(searcher.best_value))
def test_cnn_model_with_few_trials():
searcher = Searcher()
searcher.register_trial('cnn_h1', [16, 64])
searcher.register_trial('cnn_k1', [1, 3])
searcher.register_trial('cnn_s1', [1])
searcher.register_trial('pool_k1', [1, 3])
searcher.register_trial('cnn_h2', [32, 64])
searcher.register_trial('cnn_k2', [1])
searcher.register_trial('cnn_s2', [1])
searcher.register_trial('pool_k2', [2, 3])
searcher.register_trial('cnn_h3', [64, 128])
searcher.register_trial('cnn_k3', [1, 3])
searcher.register_trial('cnn_s3', [1])
searcher.register_trial('pool_k3', [1, 3])
searcher.register_trial('fc1', [64, 128, 256])
searcher.register_trial('batchsize', [32, 64, 128])
searcher.register_trial('lr', [0.05, 0.1])
n_trials = 30
model_kwargs = dict(
input_dim=15,
n_train_epochs=200,
)
warnings.warn('CNN: the number of trials = {}'.format(len(searcher)))
_ = searcher.search(cnn_objectve,
n_trials=n_trials,
deep_surrogate_model='tfdbonas.deep_surrogate_models:SimpleNetwork',
n_random_trials=10,
model_kwargs=model_kwargs)
assert len(searcher.result) == n_trials
warnings.warn('CNN: results = {}'.format(searcher.result))
warnings.warn('CNN: best_trial {}'.format(searcher.best_trial))
warnings.warn('CNN: best_value {}'.format(searcher.best_value))
def test_cnn_model_with_few_trials():
searcher = Searcher()
searcher.register_trial('cnn_h1', [16, 64])
searcher.register_trial('cnn_k1', [1, 3])
searcher.register_trial('cnn_s1', [1, 3])
searcher.register_trial('pool_k1', [1, 3])
searcher.register_trial('cnn_h2', [32, 64])
searcher.register_trial('cnn_k2', [1, 2])
searcher.register_trial('cnn_s2', [1, 2])
searcher.register_trial('pool_k2', [2, 3])
searcher.register_trial('cnn_h3', [64, 128])
searcher.register_trial('cnn_k3', [1, 3])
searcher.register_trial('cnn_s3', [1, 2])
searcher.register_trial('pool_k3', [1, 3])
searcher.register_trial('fc1', [64, 128, 256])
searcher.register_trial('batchsize', [32, 64, 128])
searcher.register_trial('lr', [0.05, 0.1])
n_trials = 30
model_kwargs = dict(
input_dim=15,
n_train_epochs=200,
)
warnings.warn('CNN: the number of trials = {}'.format(len(searcher)))
# CNN: the number of trials = 73728
# var = np.diag(np.matmul(np.matmul(predicted_bases, self.k_inv), predicted_bases.transpose()) + 1 / beta)
# MemoryError: Unable to allocate array with shape (73718, 73718) and data type float64
_ = searcher.search(cnn_objectve,
n_trials=n_trials,
deep_surrogate_model='tfdbonas.deep_surrogate_models:SimpleNetwork',
n_random_trials=10,
model_kwargs=model_kwargs)
assert len(searcher.result) == n_trials
warnings.warn('CNN: results = {}'.format(searcher.result))
warnings.warn('CNN: best_trial {}'.format(searcher.best_trial))
warnings.warn('CNN: best_value {}'.format(searcher.best_value))