MultiClassNet Class Reference

Inheritance diagram for MultiClassNet:

Collaboration diagram for MultiClassNet:

Public Member Functions
	__init__ (self, input_size, num_labels=3)
	forward (self, x)

Public Attributes
	activation = nn.ReLU()
	Activation function applied between all linear layers.
	network
	Fully connected network layers.

Protected Member Functions
	_init_weights (self)

Detailed Description

Multi-class classification network with fully connected layers.

Two architectures selected automatically by num_labels:
- Deep (num_labels <= 10, e.g. category network with 3 classes):
  256 -> 128 (dropout 0.2) -> 64 (dropout 0.15) -> 32 (dropout 0.1) -> 16 -> num_labels
- Shallow (num_labels > 10, e.g. main network with 139 classes):
  256 -> 128 (dropout 0.1) -> 128 (dropout 0.1) -> 64 (dropout 0.05) -> num_labels
- ReLU activation, Xavier initialization (gain=0.5, bias=0.01)

Definition at line 710 of file train.py.

Constructor & Destructor Documentation

__init__()

__init__	(		self,
			input_size,
			num_labels = 3 )

Build the network graph for the given input size and number of output classes.

Definition at line 721 of file train.py.

    def __init__(self, input_size, num_labels=3):
        """Build the network graph for the given input size and number of output classes."""
        super().__init__()
 
        
        self.activation = nn.ReLU()
 
        if num_labels <= 10:
            
            self.network = nn.Sequential(
                nn.Linear(input_size, 128),
                self.activation,
                nn.Linear(128, 128),
                nn.Dropout(0.2),
                self.activation,
                nn.Linear(128, 64),
                nn.Dropout(0.15),
                self.activation,
                nn.Linear(64, 32),
                nn.Dropout(0.1),
                self.activation,
                nn.Linear(32, 16),
                self.activation,
                nn.Linear(16, num_labels)
            )
        else:
            self.network = nn.Sequential(
                nn.Linear(input_size, 256),
                self.activation,
                nn.Linear(256, 128),
                nn.Dropout(0.1),
                self.activation,
                nn.Linear(128, 128),
                nn.Dropout(0.1),
                self.activation,
                nn.Linear(128, 64),
                nn.Dropout(0.05),
                self.activation,
                nn.Linear(64, num_labels)
            )
 
        self._init_weights()
 

Member Function Documentation

_init_weights()

_init_weights ( self )

protected

Initialize weights with Xavier uniform (gain=0.5) and bias=0.01.

Definition at line 764 of file train.py.

    def _init_weights(self):
        """Initialize weights with Xavier uniform (gain=0.5) and bias=0.01."""
        for m in self.modules():
            if isinstance(m, nn.Linear):
                nn.init.xavier_uniform_(m.weight, gain=0.5)
                nn.init.constant_(m.bias, 0.01)
 

forward()

forward	(		self,
			x )

Run a forward pass through the network.

Definition at line 771 of file train.py.

    def forward(self, x):
        """Run a forward pass through the network."""
        return self.network(x)
 
 

Member Data Documentation

activation

activation = nn.ReLU()

Activation function applied between all linear layers.

Definition at line 726 of file train.py.

network

network

Initial value:

=  nn.Sequential(
                nn.Linear(input_size, 128),
                self.activation,
                nn.Linear(128, 128),
                nn.Dropout(0.2),
                self.activation,
                nn.Linear(128, 64),
                nn.Dropout(0.15),
                self.activation,
                nn.Linear(64, 32),
                nn.Dropout(0.1),
                self.activation,
                nn.Linear(32, 16),
                self.activation,
                nn.Linear(16, num_labels)
            )

Fully connected network layers.

Definition at line 730 of file train.py.

---

The documentation for this class was generated from the following file:

• analysis/scripts/modeSelector/training/train.py