def get_cfg(key=None):
if key is None:
cfg = CfgNode()
cfg.feature_units = -1 # -1 means not given and we will use the units of BERT
# TODO(sxjscience) Use a class to store the TextNet
cfg.text_net = CfgNode()
cfg.text_net.use_segment_id = True
cfg.text_net.pool_type = 'cls'
cfg.agg_net = FeatureAggregator.get_cfg()
cfg.categorical_net = CategoricalFeatureNet.get_cfg()
cfg.numerical_net = NumericalFeatureNet.get_cfg()
cfg.initializer = CfgNode()
cfg.initializer.weight = ['truncnorm', 0, 0.02]
cfg.initializer.bias = ['zeros']
return cfg
else:
raise NotImplementedError
def get_cfg(key=None):
if key is None:
cfg = CfgNode()
cfg.base_feature_units = -1 # -1 means not given and we will use the units of BERT
cfg.text_net = CfgNode()
cfg.text_net.use_segment_id = True
cfg.text_net.pool_type = 'cls'
cfg.aggregate_categorical = True # Whether to use one network to aggregate the categorical columns.
cfg.categorical_agg = CfgNode()
cfg.categorical_agg.activation = 'leaky'
cfg.categorical_agg.mid_units = 128
cfg.categorical_agg.num_layers = 1
cfg.categorical_agg.dropout = 0.1
cfg.categorical_agg.gated_activation = False
cfg.agg_net = FeatureAggregator.get_cfg()
cfg.categorical_net = CategoricalFeatureNet.get_cfg()
cfg.numerical_net = NumericalFeatureNet.get_cfg()
cfg.initializer = CfgNode()
cfg.initializer.weight = ['xavier', 'uniform', 'avg', 3.0]
cfg.initializer.bias = ['zeros']
return cfg
else:
raise NotImplementedError