lvdm/models/modules/attention_temporal.py

from typing import Optional, Any

import torch
import torch as th
from torch import nn, einsum
from einops import rearrange, repeat
try:
    import xformers
    import xformers.ops
    XFORMERS_IS_AVAILBLE = True
except:
    XFORMERS_IS_AVAILBLE = False

from lvdm.models.modules.util import (
    GEGLU,
    exists,
    default,
    Normalize,
    checkpoint,
    zero_module,
)


# ---------------------------------------------------------------------------------------------------
class FeedForward(nn.Module):
    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
        super().__init__()
        inner_dim = int(dim * mult)
        dim_out = default(dim_out, dim)
        project_in = nn.Sequential(
            nn.Linear(dim, inner_dim),
            nn.GELU()
        ) if not glu else GEGLU(dim, inner_dim)

        self.net = nn.Sequential(
            project_in,
            nn.Dropout(dropout),
            nn.Linear(inner_dim, dim_out)
        )

    def forward(self, x):
        return self.net(x)


# ---------------------------------------------------------------------------------------------------
class RelativePosition(nn.Module):
    """ https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """

    def __init__(self, num_units, max_relative_position):
        super().__init__()
        self.num_units = num_units
        self.max_relative_position = max_relative_position
        self.embeddings_table = nn.Parameter(th.Tensor(max_relative_position * 2 + 1, num_units))
        nn.init.xavier_uniform_(self.embeddings_table)

    def forward(self, length_q, length_k):
        device = self.embeddings_table.device
        range_vec_q = th.arange(length_q, device=device)
        range_vec_k = th.arange(length_k, device=device)
        distance_mat = range_vec_k[None, :] - range_vec_q[:, None]
        distance_mat_clipped = th.clamp(distance_mat, -self.max_relative_position, self.max_relative_position)
        final_mat = distance_mat_clipped + self.max_relative_position
        final_mat = final_mat.long()
        embeddings = self.embeddings_table[final_mat]
        return embeddings


# ---------------------------------------------------------------------------------------------------
class TemporalCrossAttention(nn.Module):
    def __init__(self, 
        query_dim, 
        context_dim=None, 
        heads=8, 
        dim_head=64, 
        dropout=0.,
        use_relative_position=False,    # whether use relative positional representation in temporal attention.
        temporal_length=None,           # relative positional representation
        **kwargs,
    ):
        super().__init__()
        inner_dim = dim_head * heads
        context_dim = default(context_dim, query_dim)
        self.context_dim = context_dim
        self.scale = dim_head ** -0.5
        self.heads = heads
        self.temporal_length = temporal_length
        self.use_relative_position = use_relative_position
        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
        self.to_out = nn.Sequential(
            nn.Linear(inner_dim, query_dim),
            nn.Dropout(dropout)
        )

        if use_relative_position:
            assert(temporal_length is not None)
            self.relative_position_k = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
            self.relative_position_v = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)

        nn.init.constant_(self.to_q.weight, 0)
        nn.init.constant_(self.to_k.weight, 0)
        nn.init.constant_(self.to_v.weight, 0)
        nn.init.constant_(self.to_out[0].weight, 0)
        nn.init.constant_(self.to_out[0].bias, 0)

    def forward(self, x, context=None, mask=None):
        nh = self.heads
        out = x

        # cal qkv
        q = self.to_q(out)
        context = default(context, x)
        k = self.to_k(context)
        v = self.to_v(context)

        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=nh), (q, k, v))
        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale

        # relative positional embedding
        if self.use_relative_position:
            len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
            k2 = self.relative_position_k(len_q, len_k) 
            sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale
            sim += sim2
        
        # mask attention
        if mask is not None:
            max_neg_value = -1e9
            sim = sim + (1-mask.float()) * max_neg_value # 1=masking,0=no masking
        
        # attend to values
        attn = sim.softmax(dim=-1)
        out = einsum('b i j, b j d -> b i d', attn, v)
        
        # relative positional embedding
        if self.use_relative_position:
            v2 = self.relative_position_v(len_q, len_v)
            out2 = einsum('b t s, t s d -> b t d', attn, v2)
            out += out2
        
        # merge head
        out = rearrange(out, '(b h) n d -> b n (h d)', h=nh)
        return self.to_out(out)


# ---------------------------------------------------------------------------------------------------
class CrossAttention(nn.Module):
    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.,
                 **kwargs,):
        super().__init__()
        inner_dim = dim_head * heads
        context_dim = default(context_dim, query_dim)
        
        self.scale = dim_head ** -0.5
        self.heads = heads

        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)

        self.to_out = nn.Sequential(
            nn.Linear(inner_dim, query_dim),
            nn.Dropout(dropout)
        )

    def forward(self, x, context=None, mask=None):
        h = self.heads
        b = x.shape[0]
        
        q = self.to_q(x)
        context = default(context, x)
        k = self.to_k(context)
        v = self.to_v(context)
        
        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))

        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale

        if exists(mask):
            mask = rearrange(mask, 'b ... -> b (...)')
            max_neg_value = -torch.finfo(sim.dtype).max
            mask = repeat(mask, 'b j -> (b h) () j', h=h)
            sim.masked_fill_(~mask, max_neg_value)

        attn = sim.softmax(dim=-1)

        out = einsum('b i j, b j d -> b i d', attn, v)
        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
        return self.to_out(out)
    

# ---------------------------------------------------------------------------------------------------
class MemoryEfficientCrossAttention(nn.Module):
    """https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
    """
    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0,
                 **kwargs,):
        super().__init__()
        print(f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using "
              f"{heads} heads."
        )
        inner_dim = dim_head * heads
        context_dim = default(context_dim, query_dim)

        self.heads = heads
        self.dim_head = dim_head

        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)

        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
        self.attention_op: Optional[Any] = None

    def forward(self, x, context=None, mask=None):
        q = self.to_q(x)
        context = default(context, x)
        k = self.to_k(context)
        v = self.to_v(context)

        b, _, _ = q.shape
        q, k, v = map(
            lambda t: t.unsqueeze(3)
            .reshape(b, t.shape[1], self.heads, self.dim_head)
            .permute(0, 2, 1, 3)
            .reshape(b * self.heads, t.shape[1], self.dim_head)
            .contiguous(),
            (q, k, v),
        )
        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)

        if exists(mask):
            raise NotImplementedError
        out = (
            out.unsqueeze(0)
            .reshape(b, self.heads, out.shape[1], self.dim_head)
            .permute(0, 2, 1, 3)
            .reshape(b, out.shape[1], self.heads * self.dim_head)
        )
        return self.to_out(out)


# ---------------------------------------------------------------------------------------------------
class BasicTransformerBlockST(nn.Module):
    """
    if no context is given to forward function, cross-attention defaults to self-attention
    """
    def __init__(self, 
        # Spatial
        dim, 
        n_heads, 
        d_head, 
        dropout=0., 
        context_dim=None, 
        gated_ff=True, 
        checkpoint=True,
        # Temporal
        temporal_length=None,   
        use_relative_position=True,
        **kwargs,
    ):
        super().__init__()

        # spatial self attention (if context_dim is None) and spatial cross attention
        if XFORMERS_IS_AVAILBLE:
            self.attn1 = MemoryEfficientCrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, **kwargs,)
            self.attn2 = MemoryEfficientCrossAttention(query_dim=dim, context_dim=context_dim,
                                    heads=n_heads, dim_head=d_head, dropout=dropout, **kwargs,)
        else:
            self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, **kwargs,)
            self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
                                    heads=n_heads, dim_head=d_head, dropout=dropout, **kwargs,)
        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
        
        self.norm1 = nn.LayerNorm(dim)
        self.norm2 = nn.LayerNorm(dim)
        self.norm3 = nn.LayerNorm(dim)
        self.checkpoint = checkpoint
        
        # Temporal attention
        self.attn1_tmp = TemporalCrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
                                                temporal_length=temporal_length,
                                                use_relative_position=use_relative_position,
                                                **kwargs,
        )
        self.attn2_tmp = TemporalCrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
                                                # cross attn
                                                context_dim=None,
                                                # temporal attn
                                                temporal_length=temporal_length,
                                                use_relative_position=use_relative_position,
                                                **kwargs,
        )
        self.norm4 = nn.LayerNorm(dim)
        self.norm5 = nn.LayerNorm(dim)
        
    def forward(self, x, context=None, **kwargs):
        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
        
    def _forward(self, x, context=None, mask=None,):
        assert(x.dim() == 5), f"x shape = {x.shape}"
        b, c, t, h, w = x.shape
        
        # spatial self attention
        x = rearrange(x, 'b c t h w -> (b t) (h w) c')
        x = self.attn1(self.norm1(x)) + x
        x = rearrange(x, '(b t) (h w) c -> b c t h w', b=b,h=h)
        
        # temporal self attention
        x = rearrange(x, 'b c t h w -> (b h w) t c')
        x = self.attn1_tmp(self.norm4(x), mask=mask) + x
        x = rearrange(x, '(b h w) t c -> b c t h w', b=b,h=h,w=w) # 3d -> 5d
        
        # spatial cross attention
        x = rearrange(x, 'b c t h w -> (b t) (h w) c')
        if context is not None:
            context_ = []
            for i in range(context.shape[0]):
                context_.append(context[i].unsqueeze(0).repeat(t, 1, 1))
            context_ = torch.cat(context_,dim=0)
        else:
            context_ = None
        x = self.attn2(self.norm2(x), context=context_) + x
        x = rearrange(x, '(b t) (h w) c -> b c t h w', b=b,h=h)

        # temporal cross attention
        x = rearrange(x, 'b c t h w -> (b h w) t c')
        x = self.attn2_tmp(self.norm5(x), context=None, mask=mask) + x

        # feedforward
        x = self.ff(self.norm3(x)) + x
        x = rearrange(x, '(b h w) t c -> b c t h w', b=b,h=h,w=w) # 3d -> 5d
        
        return x


# ---------------------------------------------------------------------------------------------------
class SpatialTemporalTransformer(nn.Module):
    """
    Transformer block for video-like data (5D tensor).
    First, project the input (aka embedding) with NO reshape.
    Then apply standard transformer action.
    The 5D -> 3D reshape operation will be done in the specific attention module.
    """
    def __init__(
        self,
        in_channels, n_heads, d_head,
        depth=1, dropout=0.,
        context_dim=None,
        # Temporal
        temporal_length=None,
        use_relative_position=True,
        **kwargs,
        ):
        super().__init__()

        self.in_channels = in_channels
        inner_dim = n_heads * d_head

        self.norm = Normalize(in_channels)
        self.proj_in = nn.Conv3d(in_channels,
                                 inner_dim,
                                 kernel_size=1,
                                 stride=1,
                                 padding=0)

        self.transformer_blocks = nn.ModuleList(
            [BasicTransformerBlockST(
                inner_dim, n_heads, d_head, dropout=dropout,
                # cross attn
                context_dim=context_dim,
                # temporal attn
                temporal_length=temporal_length,   
                use_relative_position=use_relative_position,
                **kwargs
                ) for d in range(depth)]
        )

        self.proj_out = zero_module(nn.Conv3d(inner_dim,
                                              in_channels,
                                              kernel_size=1,
                                              stride=1,
                                              padding=0))
        
    def forward(self, x, context=None, **kwargs):
        
        assert(x.dim() == 5), f"x shape = {x.shape}"
        x_in = x
        
        x = self.norm(x)
        x = self.proj_in(x)
        
        for block in self.transformer_blocks:
            x = block(x, context=context, **kwargs)
        
        x = self.proj_out(x)

        return x + x_in