spar3d/models/transformers/backbone.py

from dataclasses import dataclass
from typing import Optional

import torch
from torch import nn

from spar3d.models.transformers.attention import FeedForward
from spar3d.models.utils import BaseModule


class CrossAttention(nn.Module):
    def __init__(
        self,
        dim,
        kv_dim=None,
        num_heads=16,
        qkv_bias=False,
        attn_drop=0.0,
        proj_drop=0.0,
    ):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = head_dim**-0.5
        kv_dim = dim if not kv_dim else kv_dim
        self.wq = nn.Linear(dim, dim, bias=qkv_bias)
        self.wk = nn.Linear(kv_dim, dim, bias=qkv_bias)
        self.wv = nn.Linear(kv_dim, dim, bias=qkv_bias)
        self.attn_drop = attn_drop
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, x_q, x_kv):
        B, N_q, C = x_q.shape
        B, N_kv, _ = x_kv.shape
        # [B, N_q, C] -> [B, N_q, H, C/H]
        q = self.wq(x_q).reshape(B, N_q, self.num_heads, C // self.num_heads)
        # [B, N_kv, C] -> [B, N_kv, H, C/H]
        k = self.wk(x_kv).reshape(B, N_kv, self.num_heads, C // self.num_heads)
        v = self.wv(x_kv).reshape(B, N_kv, self.num_heads, C // self.num_heads)

        #  attention
        x = torch.nn.functional.scaled_dot_product_attention(
            q.transpose(1, 2),
            k.transpose(1, 2),
            v.transpose(1, 2),
            attn_mask=None,
            dropout_p=self.attn_drop,
            scale=self.scale,
        ).transpose(1, 2)

        # [B, N_q, H, C/H] -> [B, N_q, C]
        x = x.reshape(B, N_q, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class BasicBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        kv_dim: Optional[int] = None,
        num_heads: int = 16,
        qkv_bias: bool = False,
        attn_drop: float = 0.0,
        proj_drop: float = 0.0,
        ff_drop: float = 0.0,
    ):
        super().__init__()
        self.norm1 = nn.LayerNorm(dim)
        self.attn1 = CrossAttention(
            dim,
            kv_dim=dim,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=proj_drop,
        )
        self.norm2 = nn.LayerNorm(dim)
        self.attn2 = CrossAttention(
            dim,
            kv_dim=kv_dim,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=proj_drop,
        )
        self.norm3 = nn.LayerNorm(dim)
        self.ff = FeedForward(dim, dropout=ff_drop)

    def forward(self, z, x):
        z_norm = self.norm1(z)
        z = z + self.attn1(z_norm, z_norm)
        # TODO: do we need to have the second attention when x is None?
        z_norm = self.norm2(z)
        z = z + self.attn2(z_norm, x if x is not None else z_norm)
        z_norm = self.norm3(z)
        z = z + self.ff(z_norm)
        return z


class SingleStreamTransformer(BaseModule):
    @dataclass
    class Config(BaseModule.Config):
        num_attention_heads: int = 16
        attention_head_dim: int = 88
        in_channels: Optional[int] = None
        out_channels: Optional[int] = None
        num_layers: int = 16
        dropout: float = 0.0
        norm_num_groups: int = 32
        cross_attention_dim: Optional[int] = None
        attention_bias: bool = False

    cfg: Config

    def configure(self) -> None:
        self.num_attention_heads = self.cfg.num_attention_heads
        self.attention_head_dim = self.cfg.attention_head_dim
        inner_dim = self.num_attention_heads * self.attention_head_dim

        # Define input layers
        self.norm = torch.nn.GroupNorm(
            num_groups=self.cfg.norm_num_groups,
            num_channels=self.cfg.in_channels,
            eps=1e-6,
            affine=True,
        )
        self.proj_in = nn.Linear(self.cfg.in_channels, inner_dim)

        # Define transformers blocks
        self.transformer_blocks = nn.ModuleList(
            [
                BasicBlock(
                    inner_dim,
                    kv_dim=self.cfg.cross_attention_dim,
                    num_heads=self.num_attention_heads,
                    qkv_bias=self.cfg.attention_bias,
                    proj_drop=self.cfg.dropout,
                    ff_drop=self.cfg.dropout,
                )
                for d in range(self.cfg.num_layers)
            ]
        )

        # 4. Define output layers
        self.proj_out = nn.Linear(inner_dim, self.cfg.in_channels)

    def forward(self, hidden_states, encoder_hidden_states=None, **kwargs):
        residual = hidden_states
        hidden_states = self.norm(hidden_states)
        hidden_states = hidden_states.permute(0, 2, 1)
        hidden_states = self.proj_in(hidden_states)
        for block in self.transformer_blocks:
            hidden_states = block(hidden_states, encoder_hidden_states)
        hidden_states = self.proj_out(hidden_states).permute(0, 2, 1).contiguous()
        # TODO: do we really need to add the residual?
        hidden_states = hidden_states + residual
        return hidden_states


class FuseBlock(nn.Module):
    """
    Fuse X in to Z with cross attention
    """

    def __init__(
        self,
        dim_z: int,
        dim_x: int,
        num_heads: int = 16,
        qkv_bias: bool = False,
        attn_drop: float = 0.0,
        proj_drop: float = 0.0,
        ff_drop: float = 0.0,
        norm_x_input: bool = True,
    ):
        super().__init__()
        self.norm_x_input = norm_x_input
        if self.norm_x_input:
            self.norm_x = nn.LayerNorm(dim_x)
        self.attn = CrossAttention(
            dim_z,
            kv_dim=dim_x,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=proj_drop,
        )
        self.norm_z1 = nn.LayerNorm(dim_z)
        self.norm_z2 = nn.LayerNorm(dim_z)
        self.ff = FeedForward(dim_z, dropout=ff_drop)

    def forward(self, z, x):
        # TODO: do we need to normalize x?
        z = z + self.attn(self.norm_z1(z), self.norm_x(x) if self.norm_x_input else x)
        z = z + self.ff(self.norm_z2(z))
        return z


@torch.no_grad()
def get_triplane_attention_mask(res):
    N = 3 * res * res
    attn_mask = torch.zeros(3, res, res, 3, res, res)

    i, j = torch.meshgrid(torch.arange(res), torch.arange(res))

    attn_mask[0, i, j, 1, i, :] = 1.0
    attn_mask[0, i, j, 2, j, :] = 1.0
    attn_mask[1, i, j, 0, i, :] = 1.0
    attn_mask[1, i, j, 2, :, j] = 1.0
    attn_mask[2, i, j, 0, :, i] = 1.0
    attn_mask[2, i, j, 1, :, j] = 1.0
    attn_mask = attn_mask.bool()

    attn_bias = torch.empty_like(attn_mask, dtype=torch.float)
    attn_bias.masked_fill_(attn_mask, 0.0)
    attn_bias.masked_fill_(~attn_mask, float("-inf"))

    return attn_bias.reshape(N, N)


class TriplaneAttention(nn.Module):
    def __init__(
        self,
        dim: int,
        resolution: int,
        num_heads: int = 16,
        qkv_bias: bool = False,
        attn_drop: float = 0.0,
        proj_drop: float = 0.0,
        full_attention: bool = False,
    ):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = head_dim**-0.5
        self.wq = nn.Linear(dim, dim, bias=qkv_bias)
        self.wk = nn.Linear(dim, dim, bias=qkv_bias)
        self.wv = nn.Linear(dim, dim, bias=qkv_bias)
        self.attn_drop = attn_drop
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

        self.resolution = resolution
        self.full_attention = full_attention
        self.attn_mask = (
            get_triplane_attention_mask(resolution) if not full_attention else None
        )

    def forward(self, x):
        B, N, C = x.shape
        # [B, N, C] -> [B, N, H, C/H]
        q = self.wq(x).reshape(B, N, self.num_heads, C // self.num_heads)
        k = self.wk(x).reshape(B, N, self.num_heads, C // self.num_heads)
        v = self.wv(x).reshape(B, N, self.num_heads, C // self.num_heads)

        # detokenize the planes
        assert N == self.resolution**2 * 3
        attn_bias = (
            self.attn_mask.to(q)
            .unsqueeze(0)
            .unsqueeze(0)
            .expand(B, self.num_heads, -1, -1)
            if not self.full_attention
            else None
        )

        # full attention
        x = torch.nn.functional.scaled_dot_product_attention(
            q.transpose(1, 2),
            k.transpose(1, 2),
            v.transpose(1, 2),
            attn_mask=attn_bias,
            dropout_p=self.attn_drop,
            scale=self.scale,
        ).transpose(1, 2)

        # [B, N_q, H, C/H] -> [B, N_q, C]
        x = x.reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class TwoStreamBlock(nn.Module):
    def __init__(
        self,
        dim_latent: int,
        dim_input: int,
        num_basic_blocks: int = 4,
        num_heads: int = 16,
        qkv_bias: bool = False,
        attn_drop: float = 0.0,
        proj_drop: float = 0.0,
        ff_drop: float = 0.0,
        norm_x_input: bool = True,
        dim_cross: Optional[int] = None,
    ):
        super().__init__()

        # Define the fuse block that fuse the input into the latent
        self.fuse_block_in = FuseBlock(
            dim_latent,
            dim_input,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=proj_drop,
            ff_drop=ff_drop,
            norm_x_input=norm_x_input,
        )

        # Define the transformer block that process the latent
        self.transformer_block = nn.ModuleList(
            [
                BasicBlock(
                    dim_latent,
                    kv_dim=dim_cross,
                    num_heads=num_heads,
                    qkv_bias=qkv_bias,
                    proj_drop=proj_drop,
                    ff_drop=ff_drop,
                )
                for _ in range(num_basic_blocks)
            ]
        )

        # Define the fuse block that fuse the latent into the input
        self.fuse_block_out = FuseBlock(
            dim_input,
            dim_latent,
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            attn_drop=attn_drop,
            proj_drop=proj_drop,
            ff_drop=ff_drop,
            norm_x_input=norm_x_input,
        )

    def forward(self, latent, input, cross_input):
        latent = self.fuse_block_in(latent, input)
        for block in self.transformer_block:
            latent = block(latent, cross_input)
        input = self.fuse_block_out(input, latent)
        return latent, input


class TwoStreamInterleaveTransformer(BaseModule):
    @dataclass
    class Config(BaseModule.Config):
        num_attention_heads: int = 16
        attention_head_dim: int = 64
        raw_triplane_channels: int = 1024
        triplane_channels: int = 1024
        raw_image_channels: int = 1024
        num_latents: int = 1792
        num_blocks: int = 4
        num_basic_blocks: int = 3
        dropout: float = 0.0
        latent_init_std: float = 0.02
        norm_num_groups: int = 32
        attention_bias: bool = False
        norm_x_input: bool = False
        cross_attention_dim: int = 1024
        mix_latent: bool = True

    cfg: Config

    def configure(self) -> None:
        self.mix_latent = self.cfg.mix_latent

        # Define the dimensions
        self.num_attention_heads = self.cfg.num_attention_heads
        self.attention_head_dim = self.cfg.attention_head_dim
        self.num_latents = self.cfg.num_latents
        self.latent_dim = self.num_attention_heads * self.attention_head_dim

        # Define input layers
        if self.cfg.norm_num_groups > 0:
            self.norm_triplane = torch.nn.GroupNorm(
                num_groups=self.cfg.norm_num_groups,
                num_channels=self.cfg.raw_triplane_channels,
                eps=1e-6,
                affine=True,
            )
        else:
            self.norm_triplane = nn.LayerNorm(self.cfg.raw_triplane_channels)
        self.proj_triplane = nn.Linear(
            self.cfg.raw_triplane_channels, self.cfg.triplane_channels
        )
        if self.mix_latent:
            self.norm_image = nn.LayerNorm(self.cfg.raw_image_channels)
            self.proj_image = nn.Linear(self.cfg.raw_image_channels, self.latent_dim)
        self.norm_latent = nn.LayerNorm(self.latent_dim)
        self.proj_latent = nn.Linear(self.latent_dim, self.latent_dim)

        # Define the latents
        self.latent_init = nn.Parameter(
            torch.zeros(1, self.num_latents, self.latent_dim)
        )
        nn.init.normal_(self.latent_init, std=self.cfg.latent_init_std)

        # Define the transformer blocks
        self.main_blocks = nn.ModuleList(
            [
                TwoStreamBlock(
                    self.latent_dim,
                    self.cfg.triplane_channels,
                    num_basic_blocks=self.cfg.num_basic_blocks,
                    num_heads=self.num_attention_heads,
                    qkv_bias=self.cfg.attention_bias,
                    proj_drop=self.cfg.dropout,
                    ff_drop=self.cfg.dropout,
                    norm_x_input=self.cfg.norm_x_input,
                    dim_cross=self.cfg.cross_attention_dim,
                )
                for _ in range(self.cfg.num_blocks)
            ]
        )

        # 4. Define output layers
        self.proj_out = nn.Linear(
            self.cfg.triplane_channels, self.cfg.raw_triplane_channels
        )

    def forward(self, hidden_states, encoder_hidden_states, **kwargs):
        # hidden_states: [B, triplane_dim, N_triplane] is triplane tokens
        # encoder_hidden_states: [B, N_image, image_dim] is the image tokens
        if isinstance(self.norm_triplane, nn.GroupNorm):
            triplane_tokens = self.norm_triplane(hidden_states)
            triplane_tokens = triplane_tokens.permute(
                0, 2, 1
            )  # [B, N_triplane, triplane_dim]
        elif isinstance(self.norm_triplane, nn.LayerNorm):
            triplane_tokens = self.norm_triplane(hidden_states.permute(0, 2, 1))
        else:
            raise ValueError("Unknown normalization layer")
        triplane_tokens = self.proj_triplane(triplane_tokens)
        if self.mix_latent:
            image_tokens = self.norm_image(
                encoder_hidden_states
            )  # [B, N_image, image_dim]
            image_tokens = self.proj_image(image_tokens)
        init_latents = self.latent_init.expand(
            hidden_states.shape[0], -1, -1
        )  # [B, N_latent_init, latent_dim]
        init_latents = self.norm_latent(init_latents)
        init_latents = self.proj_latent(init_latents)
        if self.mix_latent:
            latent_tokens = torch.cat(
                [image_tokens, init_latents], dim=1
            )  # [B, N_latent, latent_dim]
        else:
            latent_tokens = init_latents

        # forward the main blocks
        for block in self.main_blocks:
            latent_tokens, triplane_tokens = block(
                latent_tokens, triplane_tokens, encoder_hidden_states
            )

        # project the triplane tokens back to the original dimension
        triplane_tokens = self.proj_out(triplane_tokens).permute(0, 2, 1).contiguous()
        triplane_tokens = triplane_tokens + hidden_states
        return triplane_tokens