def get_adapter_kwargs(self, models, optimizers, before_training_starts,
lr_multiplier, **kwargs):
feature_combiner = models.pop("feature_combiner")
models = Models(models)
optimizers = Optimizers(optimizers, multipliers={"D": lr_multiplier})
misc = Misc({"feature_combiner": feature_combiner})
d_weight = self.optuna_trial.suggest_float("d_weight", 0, 1)
g_weight = self.optuna_trial.suggest_float("g_weight", 0, 1)
label_weight = self.optuna_trial.suggest_float("label_weight", 0, 1)
d_weighter = MeanWeighter(scale=d_weight)
g_weighter = MeanWeighter(weights={
"g_src_domain_loss": g_weight,
"g_target_domain_loss": g_weight,
"c_loss": label_weight,
}, )
hook_kwargs = {
"d_weighter": d_weighter,
"g_weighter": g_weighter,
}
return {
"models": models,
"optimizers": optimizers,
"misc": misc,
"before_training_starts": before_training_starts,
"hook_kwargs": hook_kwargs,
}
def get_adapter_kwargs(self, models, optimizers, before_training_starts,
lr_multiplier, **kwargs):
models = Models(models)
optimizers = Optimizers(optimizers, multipliers={"D": lr_multiplier})
d_weight = self.optuna_trial.suggest_float("d_weight", 0, 1)
g_weight = self.optuna_trial.suggest_float("g_weight", 0, 1)
src_weight = self.optuna_trial.suggest_float("src_weight", 0, 1)
target_weight = self.optuna_trial.suggest_float("target_weight", 0, 1)
d_loss_weighter = MeanWeighter(scale=d_weight)
g_loss_weighter = MeanWeighter(
weights={
"src_vat_loss": src_weight,
"target_vat_loss": target_weight,
"entropy_loss": target_weight,
"g_src_domain_loss": g_weight,
"g_target_domain_loss": g_weight,
})
vat_loss_fn = VATLoss(epsilon=2)
hook_kwargs = {
"d_weighter": d_loss_weighter,
"g_weighter": g_loss_weighter,
"vat_loss_fn": vat_loss_fn,
}
return {
"models": models,
"optimizers": optimizers,
"misc": None,
"before_training_starts": before_training_starts,
"hook_kwargs": hook_kwargs,
}
def get_adapter_kwargs(
self, models, optimizers, before_training_starts, lr_multiplier, **kwargs
):
models = Models(models)
optimizers = Optimizers(
optimizers,
multipliers={
"C": lr_multiplier,
"D": lr_multiplier,
},
)
domain_weight = self.optuna_trial.suggest_float("domain_weight", 0, 1)
bridge_G_weight = self.optuna_trial.suggest_float("bridge_G_weight", 0, 1)
bridge_D_weight = self.optuna_trial.suggest_float("bridge_D_weight", 0, 1)
weighter = MeanWeighter(
weights={
"src_domain_loss": domain_weight,
"target_domain_loss": domain_weight,
"g_src_bridge_loss": bridge_G_weight,
"d_src_bridge_loss": bridge_D_weight,
"g_target_bridge_loss": bridge_G_weight,
"d_target_bridge_loss": bridge_D_weight,
}
)
grl_weight = self.optuna_trial.suggest_float("grl_weight", 0.1, 10, log=True)
hook_kwargs = {"weighter": weighter, "gradient_reversal_weight": grl_weight}
return {
"models": models,
"optimizers": optimizers,
"misc": None,
"before_training_starts": before_training_starts,
"hook_kwargs": hook_kwargs,
}
def get_adapter_kwargs(self, models, optimizers, before_training_starts,
lr_multiplier, **kwargs):
models = Models(models)
optimizers = Optimizers(
optimizers,
multipliers={"D": lr_multiplier},
)
d_scale = self.optuna_trial.suggest_float("d_scale", 0, 1)
g_scale = self.optuna_trial.suggest_float("g_scale", 0, 1)
d_weighter = MeanWeighter(scale=d_scale)
g_weighter = MeanWeighter(scale=g_scale)
d_accuracy_threshold = self.optuna_trial.suggest_float(
"d_accuracy_threshold", 0, 1)
hook_kwargs = {
"threshold": d_accuracy_threshold,
"d_weighter": d_weighter,
"g_weighter": g_weighter,
}
return {
"models": models,
"optimizers": optimizers,
"misc": None,
"before_training_starts": before_training_starts,
"hook_kwargs": hook_kwargs,
}
def get_adapter_kwargs(self, models, optimizers, before_training_starts,
lr_multiplier, **kwargs):
del models["D"]
models = Models(models)
optimizers = Optimizers(optimizers, multipliers={"C": lr_multiplier})
return {
"models": models,
"optimizers": optimizers,
"misc": None,
"before_training_starts": before_training_starts,
"hook_kwargs": {},
}
def get_adapter_kwargs(self, models, optimizers, before_training_starts,
lr_multiplier, **kwargs):
feature_combiner = models.pop("feature_combiner")
models = Models(models)
optimizers = Optimizers(optimizers,
multipliers={"residual_model": lr_multiplier})
misc = Misc({"feature_combiner": feature_combiner})
label_weight = self.optuna_trial.suggest_float("label_weight", 0, 1)
confusion_weight = self.optuna_trial.suggest_float(
"confusion_weight", 0, 1)
entropy_weight = self.optuna_trial.suggest_float(
"entropy_weight", 0, 1)
weighter = MeanWeighter(
weights={
"c_loss": label_weight,
"features_confusion_loss": confusion_weight,
"entropy_loss": entropy_weight,
})
exponent = self.optuna_trial.suggest_int("exponent", 1, 8)
num_kernels = (exponent * 2) + 1
kernel_scales = get_kernel_scales(low=-exponent,
high=exponent,
num_kernels=num_kernels)
hook_kwargs = {
"weighter": weighter,
"aligner_loss_fn": MMDLoss(kernel_scales=kernel_scales),
}
return {
"models": models,
"optimizers": optimizers,
"misc": misc,
"before_training_starts": before_training_starts,
"hook_kwargs": hook_kwargs,
}
def get_adapter_kwargs(self, models, optimizers, before_training_starts,
lr_multiplier, **kwargs):
models = Models(models)
optimizers = Optimizers(
optimizers,
multipliers={"C": lr_multiplier},
)
domain_weight = self.optuna_trial.suggest_float("domain_weight", 0, 1)
category_weight = self.optuna_trial.suggest_float(
"category_weight", 0, 1)
confusion_weight = self.optuna_trial.suggest_float(
"confusion_weight", 0, 1)
entropy_weight = self.optuna_trial.suggest_float(
"entropy_weight", 0, 1)
c_weighter = MeanWeighter(
weights={
"c_symnets_src_domain_loss_0": domain_weight,
"c_symnets_target_domain_loss_1": domain_weight,
})
g_weighter = MeanWeighter(
weights={
"symnets_category_loss": category_weight,
"g_symnets_target_domain_loss_0": confusion_weight,
"g_symnets_target_domain_loss_1": confusion_weight,
"symnets_entropy_loss": entropy_weight,
})
hook_kwargs = {"c_weighter": c_weighter, "g_weighter": g_weighter}
return {
"models": models,
"optimizers": optimizers,
"misc": None,
"before_training_starts": before_training_starts,
"hook_kwargs": hook_kwargs,
}
def get_adapter_kwargs(
self, models, optimizers, before_training_starts, lr_multiplier, **kwargs
):
models = Models(models)
optimizers = Optimizers(optimizers, multipliers={"C": lr_multiplier})
confusion_weight = self.optuna_trial.suggest_float("confusion_weight", 0, 1)
label_weight = self.optuna_trial.suggest_float("label_weight", 0, 1)
weighter = MeanWeighter(
weights={
"features_confusion_loss": confusion_weight,
"logits_confusion_loss": confusion_weight,
"c_loss": label_weight,
}
)
hook_kwargs = {"weighter": weighter}
return {
"models": models,
"optimizers": optimizers,
"misc": None,
"before_training_starts": before_training_starts,
"hook_kwargs": hook_kwargs,
}
def get_adapter_kwargs(
self, models, optimizers, before_training_starts, lr_multiplier, **kwargs
):
models = Models(models)
optimizers = Optimizers(
optimizers,
multipliers={"C": lr_multiplier},
)
num_repeat = self.optuna_trial.suggest_int("num_repeat", 1, 10)
label_weight = self.optuna_trial.suggest_float("label_weight", 0, 1)
discrepancy_weight = self.optuna_trial.suggest_float("discrepancy_weight", 0, 1)
x_weighter = MeanWeighter(scale=label_weight)
y_weighter = MeanWeighter(
weights={
"c_loss0": label_weight,
"c_loss1": label_weight,
"discrepancy_loss": discrepancy_weight,
}
)
z_weighter = MeanWeighter(scale=discrepancy_weight)
hook_kwargs = {
"repeat": num_repeat,
"x_weighter": x_weighter,
"y_weighter": y_weighter,
"z_weighter": z_weighter,
}
return {
"models": models,
"optimizers": optimizers,
"misc": None,
"before_training_starts": before_training_starts,
"hook_kwargs": hook_kwargs,
}