def __init__(self,
hidden_units=None,
reduce_factor=None,
dnn_dropout=None,
use_bn=None,
dnn_layers=None,
activation=None,
space=None,
name=None,
**hyperparams):
if hidden_units is None:
hidden_units = Choice([100, 200, 300, 500, 800, 1000])
hyperparams['hidden_units'] = hidden_units
if reduce_factor is None:
reduce_factor = Choice([1, 0.8, 0.5])
hyperparams['reduce_factor'] = reduce_factor
if dnn_dropout is None:
dnn_dropout = Choice([0, 0.1, 0.3, 0.5])
hyperparams['dnn_dropout'] = dnn_dropout
if use_bn is None:
use_bn = Bool()
hyperparams['use_bn'] = use_bn
if dnn_layers is None:
dnn_layers = Choice([1, 2, 3])
hyperparams['dnn_layers'] = dnn_layers
if activation is None:
activation = 'relu'
hyperparams['activation'] = activation
ModuleSpace.__init__(self, space, name, **hyperparams)
def cnn_search_space(input_shape, output_units, output_activation='softmax', block_num_choices=[2, 3, 4, 5, 6],
activation_choices=['relu'], filters_choices=[32, 64], kernel_size_choices=[(1, 1), (3, 3)]):
space = HyperSpace()
with space.as_default():
hp_use_bn = Bool()
hp_pooling = Choice(list(range(2)))
hp_filters = Choice(filters_choices)
hp_kernel_size = Choice(kernel_size_choices)
hp_fc_units = Choice([1024, 2048, 4096])
if len(activation_choices) == 1:
hp_activation = activation_choices[0]
else:
hp_activation = Choice(activation_choices)
hp_bn_act = Choice([seq for seq in itertools.permutations(range(2))])
input = Input(shape=input_shape)
blocks = Repeat(
lambda step: conv_block(
block_no=step,
hp_pooling=hp_pooling,
hp_filters=hp_filters,
hp_kernel_size=hp_kernel_size,
hp_use_bn=hp_use_bn,
hp_activation=hp_activation,
hp_bn_act=hp_bn_act),
repeat_times=block_num_choices)(input)
x = Flatten()(blocks)
x = Dense(units=hp_fc_units, activation=hp_activation, name='fc1')(x)
x = Dense(units=hp_fc_units, activation=hp_activation, name='fc2')(x)
x = Dense(output_units, activation=output_activation, name='predictions')(x)
return space
def mini_dt_space():
space = HyperSpace()
with space.as_default():
p_nets = MultipleChoice(['dnn_nets', 'linear', 'fm_nets'],
num_chosen_most=2)
dt_module = DTModuleSpace(nets=p_nets,
auto_categorize=Bool(),
cat_remain_numeric=Bool(),
auto_discrete=Bool(),
apply_gbm_features=Bool(),
gbm_feature_type=Choice([
DT_consts.GBM_FEATURE_TYPE_DENSE,
DT_consts.GBM_FEATURE_TYPE_EMB
]),
embeddings_output_dim=Choice([4, 10]),
embedding_dropout=Choice([0, 0.5]),
stacking_op=Choice([
DT_consts.STACKING_OP_ADD,
DT_consts.STACKING_OP_CONCAT
]),
output_use_bias=Bool(),
apply_class_weight=Bool(),
earlystopping_patience=Choice([1, 3, 5]))
dnn = DnnModule(hidden_units=Choice([100, 200]),
reduce_factor=Choice([1, 0.8]),
dnn_dropout=Choice([0, 0.3]),
use_bn=Bool(),
dnn_layers=2,
activation='relu')(dt_module)
fit = DTFit(batch_size=Choice([128, 256]))(dt_module)
return space
def get_space():
space = HyperSpace()
with space.as_default():
p1 = Int(1, 100)
p2 = Choice(['a', 'b'])
p3 = Bool()
p4 = Real(0.0, 1.0)
id1 = Identity(p1=p1)
id2 = Identity(p2=p2)(id1)
id3 = Identity(p3=p3)(id2)
id4 = Identity(p4=p4)(id3)
return space
def default_dt_space():
space = HyperSpace()
with space.as_default():
p_nets = MultipleChoice([
'dnn_nets', 'linear', 'cin_nets', 'fm_nets', 'afm_nets',
'pnn_nets', 'cross_nets', 'cross_dnn_nets', 'dcn_nets',
'autoint_nets', 'fgcnn_dnn_nets', 'fibi_dnn_nets'
],
num_chosen_most=3)
dt_module = DTModuleSpace(
nets=p_nets,
auto_categorize=Bool(),
cat_remain_numeric=Bool(),
auto_discrete=Bool(),
apply_gbm_features=Bool(),
gbm_feature_type=Choice([
DT_consts.GBM_FEATURE_TYPE_DENSE,
DT_consts.GBM_FEATURE_TYPE_EMB
]),
embeddings_output_dim=Choice([4, 10, 20]),
embedding_dropout=Choice([0, 0.1, 0.2, 0.3, 0.4, 0.5]),
stacking_op=Choice(
[DT_consts.STACKING_OP_ADD, DT_consts.STACKING_OP_CONCAT]),
output_use_bias=Bool(),
apply_class_weight=Bool(),
earlystopping_patience=Choice([1, 3, 5]))
dnn = DnnModule()(dt_module)
fit = DTFit(batch_size=Choice([128, 256]))(dt_module)
return space
def tiny_dt_space(**hyperparams):
space = HyperSpace()
with space.as_default():
dt_module = DTModuleSpace(nets=['dnn_nets'],
auto_categorize=Bool(),
cat_remain_numeric=Bool(),
auto_discrete=False,
apply_gbm_features=False,
stacking_op=Choice([
DT_consts.STACKING_OP_ADD,
DT_consts.STACKING_OP_CONCAT
]),
output_use_bias=Bool(),
apply_class_weight=Bool(),
earlystopping_patience=Choice([1, 3, 5]))
dnn = DnnModule(hidden_units=Choice([10, 20]),
reduce_factor=1,
dnn_dropout=Choice([0, 0.3]),
use_bn=False,
dnn_layers=2,
activation='relu')(dt_module)
fit = DTFit(**hyperparams)(dt_module)
return space
def get_space():
space = HyperSpace()
with space.as_default():
id1 = Identity(p1=Int(0, 10), p2=Choice(['a', 'b']))
id2 = Identity(p3=Real(0., 1.), p4=Bool())(id1)
return space
