espnet2.enh.layers.ncsnpp_utils.layers.ddpm_conv3x3

source

espnet2.enh.layers.ncsnpp_utils.layers.ddpm_conv3x3(in_planes, out_planes, stride=1, bias=True, dilation=1, init_scale=1.0, padding=1)

3x3 convolution with DDPM initialization.

This function creates a 3x3 convolutional layer initialized using the DDPM (Denoising Diffusion Probabilistic Models) method. The weights are initialized with a scale factor and the biases are set to zero.

• Parameters:
  • in_planes (int) – Number of input channels.
  • out_planes (int) – Number of output channels.
  • stride (int , optional) – Stride of the convolution. Defaults to 1.
  • bias (bool , optional) – Whether to include a bias term. Defaults to True.
  • dilation (int , optional) – Dilation rate for the convolution. Defaults to 1.
  • init_scale (float , optional) – Scale for weight initialization. Defaults to 1.0.
  • padding (int , optional) – Padding added to both sides of the input. Defaults to 1.
• Returns: A 3x3 convolutional layer with DDPM initialization.
• Return type: nn.Conv2d

Examples

>>> conv_layer = ddpm_conv3x3(16, 32)
>>> output = conv_layer(torch.randn(1, 16, 64, 64))
>>> print(output.shape)
torch.Size([1, 32, 64, 64])

NOTE

The weight initialization is performed using the default_init function defined elsewhere in the codebase, which applies variance scaling to the weights based on the given scale factor.