import torch.nn as nn

from ...models.utils import trunc_normal_


class DINOHead(nn.Module):
    def __init__(
        self,
        in_dim,
        out_dim,
        use_bn=False,
        norm_last_layer=True,
        n_layers=3,
        hidden_dim=2048,
        bottleneck_dim=256,
    ):
        super().__init__()

        n_layers = max(n_layers, 1)
        if n_layers == 1:
            self.mlp = nn.Linear(in_dim, bottleneck_dim)
        else:
            layers = [nn.Linear(in_dim, hidden_dim)]
            if use_bn:
                layers.append(nn.BatchNorm1d(hidden_dim))
            layers.append(nn.GELU())
            for _ in range(n_layers - 2):
                layers.append(nn.Linear(hidden_dim, hidden_dim))
                if use_bn:
                    layers.append(nn.BatchNorm1d(hidden_dim))
                layers.append(nn.GELU())
            layers.append(nn.Linear(hidden_dim, bottleneck_dim))
            self.mlp = nn.Sequential(*layers)
        self.apply(self._init_weights)
        self.last_layer = nn.utils.weight_norm(
            nn.Linear(bottleneck_dim, out_dim, bias=False)
        )
        self.last_layer.weight_g.data.fill_(1)
        if norm_last_layer:
            self.last_layer.weight_g.requires_grad = False

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=0.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        x = self.mlp(x)
        x = nn.functional.normalize(x, dim=-1, p=2)
        x = self.last_layer(x)
        return x