DenseNet

class cv.backbones.DenseNet.model.DenseNet(num_classes, dense_block_num_layers=[6, 12, 24, 16], in_channels=3, growth_rate=32, compression_factor=0.5, dense_block_dropouts=None)

Bases: Module

Implements the DenseNet model as described in the paper. DenseNet is a type of convolutional neural network where each layer is connected to every other layer in a feed-forward fashion. It uses dense connections to improve gradient flow and enhance feature reuse.

Parameters:

• num_classes (int) – Number of output classes for classification.

• dense_block_num_layers (list of int) – List specifying the number of layers in each dense block (default: [6, 12, 24, 16]).

• in_channels (int) – Number of input channels in the image, typically 3 for RGB (default: 3).

• growth_rate (int) – Growth rate of the network. Determines the number of output channels of each dense layer (default:32).

• compression_factor (float) – Compression factor for transition layers to reduce feature map size (default: 0.5).

• dense_block_dropouts (list of float or None) – Dropout rates for dense blocks. If None, dropout is not applied (default: None).

Example

>>> model = DenseNet(num_classes=1000)

initializeConv()

Initialize convolutional layers with specific weight and bias initialization.

This method initializes the weights and biases of convolutional layers using a specific strategy: - Weights: Initialized with a normal distribution where the mean is 0.0 and the standard deviation is computed based on the number of input units and output channels. - Biases: Initialized to 0.0.

The standard deviation for weight initialization is calculated using the formula:

\[\text{std} = \sqrt{\frac{2}{n_{\text{in}}}}\]

where:

\[n_{\text{in}} = \text{kernel_size[0]}^2 \times \text{out_channels}\]

The weight initialization is performed as follows:

\[\text{weight} \sim \mathcal{N}(\text{mean}=0.0, \text{std})\]

Biases are initialized with:

\[\text{bias} = 0.0\]

This initialization helps in stabilizing the learning process and improving the convergence rate.

forward(x)

Forward pass through the DenseNet model.

The input tensor is processed through the initial feature extraction layers, followed by a series of DenseBlocks, TransitionLayers, and finally a classification layer.

Parameters:

x (Tensor) – Input tensor of shape (batch_size, in_channels, height, width).

Returns:

Output logits of shape (batch_size, num_classes).

Return type:

torch.Tensor

Example

>>> output = model(torch.randn(1, 3, 224, 224))  # Example input tensor of shape (batch_size, channels, height, width)