svrm/ldm/modules/attention.py

from inspect import isfunction
import math
import torch
import torch.nn.functional as F
from torch import nn, einsum
from einops import rearrange, repeat
import numpy as np

FLASH_IS_AVAILABLE = XFORMERS_IS_AVAILBLE = False
try:
    from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
    FLASH_IS_AVAILABLE = True
except:
    try:
        import xformers
        import xformers.ops
        XFORMERS_IS_AVAILBLE = True
    except:
        pass

def exists(val):
    return val is not None


def uniq(arr):
    return{el: True for el in arr}.keys()


def default(val, d):
    if exists(val):
        return val
    return d() if isfunction(d) else d


def max_neg_value(t):
    return -torch.finfo(t.dtype).max


def init_(tensor):
    dim = tensor.shape[-1]
    std = 1 / math.sqrt(dim)
    tensor.uniform_(-std, std)
    return tensor

def checkpoint(func, inputs, params, flag):
    """
    Evaluate a function without caching intermediate activations, allowing for
    reduced memory at the expense of extra compute in the backward pass.
    :param func: the function to evaluate.
    :param inputs: the argument sequence to pass to `func`.
    :param params: a sequence of parameters `func` depends on but does not
                   explicitly take as arguments.
    :param flag: if False, disable gradient checkpointing.
    """
    if flag:
        args = tuple(inputs) + tuple(params)
        return CheckpointFunction.apply(func, len(inputs), *args)
    else:
        return func(*inputs)


class CheckpointFunction(torch.autograd.Function):
    @staticmethod
    def forward(ctx, run_function, length, *args):
        ctx.run_function = run_function
        ctx.input_tensors = list(args[:length])
        ctx.input_params = list(args[length:])

        with torch.no_grad():
            output_tensors = ctx.run_function(*ctx.input_tensors)
        return output_tensors

    @staticmethod
    def backward(ctx, *output_grads):
        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
        with torch.enable_grad():
            # Fixes a bug where the first op in run_function modifies the
            # Tensor storage in place, which is not allowed for detach()'d
            # Tensors.
            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
            output_tensors = ctx.run_function(*shallow_copies)
        input_grads = torch.autograd.grad(
            output_tensors,
            ctx.input_tensors + ctx.input_params,
            output_grads,
            allow_unused=True,
        )
        del ctx.input_tensors
        del ctx.input_params
        del output_tensors
        return (None, None) + input_grads


# feedforward
class GEGLU(nn.Module):
    def __init__(self, dim_in, dim_out):
        super().__init__()
        self.proj = nn.Linear(dim_in, dim_out * 2)

    def forward(self, x):
        x, gate = self.proj(x).chunk(2, dim=-1)
        return x * F.gelu(gate)


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)


def zero_module(module):
    """
    Zero out the parameters of a module and return it.
    """
    for p in module.parameters():
        p.detach().zero_()
    return module


def Normalize(in_channels):
    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)


class LinearAttention(nn.Module):
    def __init__(self, dim, heads=4, dim_head=32):
        super().__init__()
        self.heads = heads
        hidden_dim = dim_head * heads
        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
        self.to_out = nn.Conv2d(hidden_dim, dim, 1)

    def forward(self, x):
        b, c, h, w = x.shape
        qkv = self.to_qkv(x)
        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
        k = k.softmax(dim=-1)  
        context = torch.einsum('bhdn,bhen->bhde', k, v)
        out = torch.einsum('bhde,bhdn->bhen', context, q)
        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
        return self.to_out(out)


class SpatialSelfAttention(nn.Module):
    def __init__(self, in_channels):
        super().__init__()
        self.in_channels = in_channels

        self.norm = Normalize(in_channels)
        self.q = torch.nn.Conv2d(in_channels,
                                 in_channels,
                                 kernel_size=1,
                                 stride=1,
                                 padding=0)
        self.k = torch.nn.Conv2d(in_channels,
                                 in_channels,
                                 kernel_size=1,
                                 stride=1,
                                 padding=0)
        self.v = torch.nn.Conv2d(in_channels,
                                 in_channels,
                                 kernel_size=1,
                                 stride=1,
                                 padding=0)
        self.proj_out = torch.nn.Conv2d(in_channels,
                                        in_channels,
                                        kernel_size=1,
                                        stride=1,
                                        padding=0)

    def forward(self, x):
        h_ = x
        h_ = self.norm(h_)
        q = self.q(h_)
        k = self.k(h_)
        v = self.v(h_)

        # compute attention
        b,c,h,w = q.shape
        q = rearrange(q, 'b c h w -> b (h w) c')
        k = rearrange(k, 'b c h w -> b c (h w)')
        w_ = torch.einsum('bij,bjk->bik', q, k)

        w_ = w_ * (int(c)**(-0.5))
        w_ = torch.nn.functional.softmax(w_, dim=2)

        # attend to values
        v = rearrange(v, 'b c h w -> b c (h w)')
        w_ = rearrange(w_, 'b i j -> b j i')
        h_ = torch.einsum('bij,bjk->bik', v, w_)
        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
        h_ = self.proj_out(h_)

        return x+h_


class CrossAttention(nn.Module):
    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
        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
        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)
        # attention, what we cannot get enough of
        attn = sim.softmax(dim=-1)
        out = einsum('b i j, b j d -> b i d', attn, v) # [b*h, n, d]
        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
        return self.to_out(out)


class FlashAttention(nn.Module):
    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):
        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.scale = dim_head ** -0.5
        self.heads = heads
        self.dropout = dropout
        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):
        context = default(context, x)
        h = self.heads
        dtype = torch.bfloat16 # torch.half
        q = self.to_q(x).to(dtype)
        k = self.to_k(context).to(dtype)
        v = self.to_v(context).to(dtype)
        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b n h d', h=h), (q, k, v)) # q is [b, 3079, 16, 64]
        out = flash_attn_func(q, k, v, 
                              dropout_p=self.dropout, 
                              softmax_scale=None, 
                              causal=False, 
                              window_size=(-1, -1)
                             ) # out is same shape to q
        out = rearrange(out, 'b n h d -> b n (h d)', h=h)
        return self.to_out(out.float())

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):
        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),
        )

        # actually compute the attention, what we cannot get enough of
        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 BasicTransformerBlock(nn.Module):
    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
                 disable_self_attn=False):
        super().__init__()
        self.disable_self_attn = disable_self_attn
        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
                                    context_dim=context_dim if self.disable_self_attn else None)
        # is a self-attention if not self.disable_self_attn
        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
                                    heads=n_heads, dim_head=d_head, dropout=dropout)  
        # is self-attn if context is none
        self.norm1 = Fp32LayerNorm(dim)
        self.norm2 = Fp32LayerNorm(dim)
        self.norm3 = Fp32LayerNorm(dim)
        self.checkpoint = checkpoint

    def forward(self, x, context=None):
        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)

    def _forward(self, x, context=None):
        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x
        x = self.attn2(self.norm2(x), context=context) + x
        x = self.ff(self.norm3(x)) + x
        return x

ATTENTION_MODES = {
    "softmax": CrossAttention,  # vanilla attention
    "softmax-xformers": MemoryEfficientCrossAttention, 
    "softmax-flash": FlashAttention
}

def modulate(x, shift, scale):
    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)


class Fp32LayerNorm(nn.LayerNorm):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
    def forward(self, x):
        return super().forward(x.float()).type(x.dtype)


class AdaNorm(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(),
            nn.Linear(dim, 2 * dim, bias=True)
        )
        self.norm = Fp32LayerNorm(dim, elementwise_affine=False, eps=1e-6)

    def forward(self, x, c):  # x is fp32, c is fp16
        shift, scale = self.adaLN_modulation(c.float()).chunk(2, dim=1) # bf16
        x = modulate(self.norm(x), shift, scale) # fp32
        return x


class BasicTransformerBlockLRM(nn.Module):
    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, \
                 checkpoint=True):
        super().__init__()

        attn_mode = "softmax-xformers" if XFORMERS_IS_AVAILBLE else "softmax"
        attn_mode = "softmax-flash" if FLASH_IS_AVAILABLE else attn_mode
        assert attn_mode in ATTENTION_MODES
        attn_cls = ATTENTION_MODES[attn_mode]

        self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, \
                              context_dim=context_dim) # cross-attn
        self.attn2 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, \
                              context_dim=None) # self-attn

        self.norm1 = Fp32LayerNorm(dim)
        self.norm2 = Fp32LayerNorm(dim)
        self.norm3 = Fp32LayerNorm(dim)
        
        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
        self.checkpoint = checkpoint

    def forward(self, x, context=None, cam_emb=None): # (torch.float32, torch.float32, torch.bfloat16)
        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)


    def _forward(self, x, context=None, cam_emb=None):

        x = self.attn1(self.norm1(x), context=context) + x  # cross-attn
        x = self.attn2(self.norm2(x), context=None) + x # self-attn
        x = self.ff(self.norm3(x)) + x

        return x

class ImgToTriplaneTransformer(nn.Module):
    """
    Transformer block for image-like data.
    First, project the input (aka embedding)
    and reshape to b, t, d.
    Then apply standard transformer action.
    Finally, reshape to image
    """
    def __init__(self, query_dim, n_heads, d_head, depth=1, dropout=0., context_dim=None, triplane_size=64):
        super().__init__()

        self.transformer_blocks = nn.ModuleList([
            BasicTransformerBlockLRM(query_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim) 
            for d in range(depth)])
        
        self.norm = Fp32LayerNorm(query_dim, eps=1e-6)
        
        self.initialize_weights()

    def initialize_weights(self):
        # Initialize transformer layers:
        def _basic_init(module):
            if isinstance(module, nn.Linear):
                torch.nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)
            elif isinstance(module, nn.LayerNorm):
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)
                if module.weight is not None:
                    nn.init.constant_(module.weight, 1.0)
        self.apply(_basic_init)

    def forward(self, x, context=None, cam_emb=None):
        # note: if no context is given, cross-attention defaults to self-attention
        for block in self.transformer_blocks:
            x = block(x, context=context)  
        x = self.norm(x)
        return x