def SagCWL2GCN_Binary_quick_max(
dsm,
cwl2_local_act="sigmoid",
cwl2_layer_widths=[64],
cwl2_layer_depths=[3],
cwl2_stack_tfs=["keep_input"]):
in_dim = dsm.dim_wl2_features()
hidden = [
[b] * l
for b, l in cart(cwl2_layer_widths, cwl2_layer_depths)]
hps = cart(
conv_layer_dims=[[in_dim, *h, 1] for h in hidden],
conv_act=["sigmoid"],
conv_local_act=[cwl2_local_act],
conv_stack_tf=cwl2_stack_tfs,
conv_bias=[True],
learning_rate=[0.001],
squeeze_output=[True]
)
duplicate_settings = {
"conv_layer_dims": ["att_conv_layer_dims"],
"conv_act": ["att_conv_act"],
"conv_local_act": ["att_conv_local_act"],
"conv_stack_tf": ["att_conv_stack_tf"],
"conv_bias": ["att_conv_bias"],
}
return fy.map(entry_duplicator(duplicate_settings), hps)
def AvgCWL2GCN_FC_Binary(
dsm,
cwl2_local_act="sigmoid",
cwl2_layer_widths=[32, 64],
cwl2_layer_depths=[1, 3],
cwl2_stack_tfs=[None, "keep_input"]):
in_dim = dsm.dim_wl2_features()
hidden = [
([b] * l, [b, b])
for b, l in cart(cwl2_layer_widths, cwl2_layer_depths)]
hidden_hp = [dict(
conv_layer_dims=[in_dim, *ch],
fc_layer_dims=[*fh, 1]
) for ch, fh in hidden]
return cart_merge(cart(
conv_act=["sigmoid"],
conv_local_act=[cwl2_local_act],
conv_stack_tf=cwl2_stack_tfs,
conv_bias=[True],
fc_bias=[True],
learning_rate=[0.01, 0.001, 0.0001],
squeeze_output=[True]
), hidden_hp)
def AvgCWL2GCN_Binary(
dsm,
cwl2_local_act="sigmoid",
cwl2_layer_widths=[32, 64],
cwl2_layer_depths=[1, 3],
cwl2_stack_tfs=[None, "keep_input"]):
in_dim = dsm.dim_wl2_features()
hidden = [
[b] * l
for b, l in cart(cwl2_layer_widths, cwl2_layer_depths)]
return cart(
conv_layer_dims=[[in_dim, *h, 1] for h in hidden],
conv_act=["sigmoid"],
conv_local_act=[cwl2_local_act],
conv_stack_tf=cwl2_stack_tfs,
conv_bias=[True],
learning_rate=[0.01, 0.001, 0.0001],
squeeze_output=[True]
)
def AvgWL2GCN_Binary_3x32(dsm):
"Like AvgWL2GCN_Binary but only one hyperparam config."
in_dim = dsm.dim_wl2_features()
hidden = [
[32, 32, 32]
]
return cart(
layer_dims=[[in_dim, *h, 1] for h in hidden],
bias=[True],
learning_rate=[0.0001]
)
def SagK2GNN_FC_Binary(
dsm,
cwl2_local_act="sigmoid",
cwl2_layer_widths=[32, 64],
cwl2_layer_depths=[2, 4],
cwl2_stack_tfs=[None, "keep_input"]):
in_dim = dsm.dim_wl2_features()
hidden = [
([b] * l, [b, b])
for b, l in cart(cwl2_layer_widths, cwl2_layer_depths)]
hidden_hp = [dict(
conv_layer_dims=[in_dim, *ch],
fc_layer_dims=[*fh, 1]
) for ch, fh in hidden]
base_hp = cart(
conv_act=["sigmoid"],
conv_local_act=[cwl2_local_act],
conv_stack_tf=cwl2_stack_tfs,
conv_bias=[True],
fc_bias=[True],
learning_rate=[0.01, 0.001, 0.0001],
squeeze_output=[True]
)
hps = cart_merge(base_hp, hidden_hp)
duplicate_settings = {
"conv_layer_dims": ["att_conv_layer_dims"],
"conv_act": ["att_conv_act"],
"conv_local_act": ["att_conv_local_act"],
"conv_stack_tf": ["att_conv_stack_tf"],
"conv_bias": ["att_conv_bias"],
}
return fy.map(entry_duplicator(duplicate_settings), hps)
def AvgWL2GCN_Binary(dsm):
"Small hyperparam space for averaging WL2GCNs + binary classification."
in_dim = dsm.dim_wl2_features()
hidden = [
[8], [8, 8, 8],
[32], [32, 32]
]
return cart(
layer_dims=[[in_dim, *h, 1] for h in hidden],
bias=[True],
learning_rate=[0.0001]
)
def AvgGIN_Timing_Binary(
dsm,
wl1_local_act="sigmoid",
wl1_stack_tfs=[None]):
in_dim = dsm.dim_wl1_features()
return cart(
conv_act=["sigmoid"],
conv_stack_tf=wl1_stack_tfs,
conv_layer_dims=[[in_dim, 32, 1]],
conv_bias=[True],
fc_bias=[True],
learning_rate=[0.01],
squeeze_output=[True]
)
def AvgCWL2GCN_Timing_Binary(
dsm,
cwl2_local_act="sigmoid",
cwl2_stack_tfs=[None]):
in_dim = dsm.dim_wl2_features()
return cart(
conv_act=["sigmoid"],
conv_local_act=[cwl2_local_act],
conv_stack_tf=cwl2_stack_tfs,
conv_layer_dims=[[in_dim, 24, 24, 1]],
conv_bias=[True],
fc_bias=[True],
learning_rate=[0.01],
squeeze_output=[True]
)
def AvgWL2GCN_FC_Binary_Small(dsm):
"Small hyperparam space for averaging FC WL2GCNs + binary classification."
in_dim = dsm.dim_wl2_features()
base = cart(
squeeze_output=[False],
bias=[True],
learning_rate=[0.0001])
hidden = [
([8, 8, 8], [8, 8]),
([32, 32], [32, 32])
]
hidden_hp = [dict(
conv_layer_dims=[in_dim, *ch],
fc_layer_dims=[*fh, 1]
) for ch, fh in hidden]
return cart_merge(base, hidden_hp)
from __future__ import absolute_import, division, print_function,\ unicode_literals import ltag.models.kernel as kernel_models from ltag.evaluation.model_factories import kernel_classifier from ltag.utils import cart kernel_hps = lambda dsm: cart(C=[1, 0.1, 0.01, 0.001, 0.0001]) WL_st = kernel_classifier(kernel_models.WL_st)(kernel_hps) WL_st_1 = kernel_classifier(kernel_models.WL_st_1)(kernel_hps) WL_st_2 = kernel_classifier(kernel_models.WL_st_2)(kernel_hps) WL_st_3 = kernel_classifier(kernel_models.WL_st_3)(kernel_hps) WL_st_4 = kernel_classifier(kernel_models.WL_st_4)(kernel_hps) WL_sp = kernel_classifier(kernel_models.WL_sp)(kernel_hps) WL_sp_3 = kernel_classifier(kernel_models.WL_sp_3)(kernel_hps) LWL2 = kernel_classifier(kernel_models.LWL2)(kernel_hps) LWL2_1 = kernel_classifier(kernel_models.LWL2_1)(kernel_hps) GWL2 = kernel_classifier(kernel_models.GWL2)(kernel_hps)
