Class MappingNetwork

tools/utils/layers.py:23–104

@persistence.persistent_class
class MappingNetwork(torch.nn.Module):
    def __init__(self,
        z_dim,                      # Input latent (Z) dimensionality, 0 = no latent.
        c_dim,                      # Conditioning label (C) dimensionality, 0 = no label.
        w_dim,                      # Intermediate latent (W) dimensionality.
        num_ws,                     # Number of intermediate latents to output, None = do not broadcast.
        num_layers      = 8,        # Number of mapping layers.
        embed_features  = None,     # Label embedding dimensionality, None = same as w_dim.
        layer_features  = None,     # Number of intermediate features in the mapping layers, None = same as w_dim.
        activation      = 'lrelu',  # Activation function: 'relu', 'lrelu', etc.
        lr_multiplier   = 0.01,     # Learning rate multiplier for the mapping layers.
        w_avg_beta      = 0.995,    # Decay for tracking the moving average of W during training, None = do not track.
        cfg             = {},       # Additional config
    ):
        super().__init__()

        self.cfg = cfg
        self.z_dim = z_dim
        self.c_dim = c_dim
        self.w_dim = w_dim
        self.num_ws = num_ws
        self.num_layers = num_layers
        self.w_avg_beta = w_avg_beta

        if embed_features is None:
            embed_features = w_dim
        if c_dim == 0:
            embed_features = 0
        if layer_features is None:
            layer_features = w_dim

        features_list = [z_dim + embed_features] + [layer_features] * (num_layers - 1) + [w_dim]

        if c_dim > 0:
            self.embed = FullyConnectedLayer(c_dim, embed_features)

        for idx in range(num_layers):
            in_features = features_list[idx]
            out_features = features_list[idx + 1]
            layer = FullyConnectedLayer(in_features, out_features, activation=activation, lr_multiplier=lr_multiplier)
            setattr(self, f'fc{idx}', layer)

        if num_ws is not None and w_avg_beta is not None:
            self.register_buffer('w_avg', torch.zeros([w_dim]))

    def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, skip_w_avg_update=False):
        # Embed, normalize, and concat inputs.
        x = None
        with torch.autograd.profiler.record_function('input'):
            if self.z_dim > 0:
                misc.assert_shape(z, [None, self.z_dim])
                x = normalize_2nd_moment(z.to(torch.float32))

            if self.c_dim > 0:
                misc.assert_shape(c, [None, self.c_dim])
                y = normalize_2nd_moment(self.embed(c.to(torch.float32)))
                x = torch.cat([x, y], dim=1) if x is not None else y