ncdia.model

ncdia.models.models

get_network(config)

load model.

Parameters:

• trainer(config): model config

ncdia.models.net.inc_net

BaseNet

BaseNet for incremental learning.

__init__(self, network, base_classes, num_classes, att_classes, net_alice, mode)

The constructor method that initializes an instance of BaseNet.

Parameters:

• network (config): The config of the network.
• base_classes(int): The number of base classes.
• num_classes(int): The total class number.
• att_classes(int): The attribute class number.
• mode(str): classifier mode.

feature_dim(self)

The feature dimension of the network.

Returns:

• out_dim(int) feature dimension of the network.

extractor_vector(self, x)

get features of input x.

Parameters:

• x(tensor): input data.

Returns:

• out_features(tensor) features of the input.

forward(self, x)

forworad pass of the network.

Parameters:

• x(tensor): input data.

Returns:

• results (dict): forward pass results. Contains the following keys:
  • "fmaps": [x_1, x_2, ..., x_n],
  • "features": features
  • "logits": logits

copy(self)

copy.

Returns:

• copy function.

freeze(self)

freeze parameters.

IncrementalNet

Incremental Network which follows BaseNet.

__init__(self, network, base_classes, num_classes, att_classes, net_alice, mode)

The constructor method that initializes an instance of BaseNet.

Parameters:

• network (config): The config of the network.
• base_classes(int): The number of base classes.
• num_classes(int): The total class number.
• att_classes(int): The attribute class number.
• mode(str): classifier mode.

update_fc(self, nb_classes)

Update fc parameter, generate new fc and copy old parameter.

Parameters:

• network (int): New class number.

Returns:

• fc: updated fc layers.

generate_fc(self, in_dim, out_dim)

Parameters:

• in_dim (int): new fc in dimension.
• out_dim (int): new fc out dimension.

Returns:

• fc: new fc layers.

forward(self, x)

forworad pass of the network.

Parameters:

• x(tensor): input data.

Returns:

• results (dict): forward pass results. Contains the following keys:
  • "fmaps": [x_1, x_2, ..., x_n],
  • "features": features
  • "logits": logits

weight_align(self, increment)

Normalize classifer parameters.

Parameters:

• increment(int): incremental classes.