attention.py

import math

import torch
import torch.nn.functional as F
from torch import nn


class AttentionBlock(nn.Module):
    """
    An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted
    to the N-d case.
    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
    Uses three q, k, v linear layers to compute attention
    """

    def __init__(
        self,
        channels,
        num_head_channels=None,
        num_groups=32,
        rescale_output_factor=1.0,
        eps=1e-5,
    ):
        super().__init__()
        self.channels = channels

        self.num_heads = channels // num_head_channels if num_head_channels is not None else 1
        self.num_head_size = num_head_channels
        self.group_norm = nn.GroupNorm(num_channels=channels, num_groups=num_groups, eps=eps, affine=True)

        # define q,k,v as linear layers
        self.query = nn.Linear(channels, channels)
        self.key = nn.Linear(channels, channels)
        self.value = nn.Linear(channels, channels)

        self.rescale_output_factor = rescale_output_factor
        self.proj_attn = nn.Linear(channels, channels, 1)

    def transpose_for_scores(self, projection: torch.Tensor) -> torch.Tensor:
        new_projection_shape = projection.size()[:-1] + (self.num_heads, -1)
        # move heads to 2nd position (B, T, H * D) -> (B, T, H, D) -> (B, H, T, D)
        new_projection = projection.view(new_projection_shape).permute(0, 2, 1, 3)
        return new_projection

    def forward(self, hidden_states):
        residual = hidden_states
        batch, channel, height, width = hidden_states.shape

        # norm
        hidden_states = self.group_norm(hidden_states)

        hidden_states = hidden_states.view(batch, channel, height * width).transpose(1, 2)

        # proj to q, k, v
        query_proj = self.query(hidden_states)
        key_proj = self.key(hidden_states)
        value_proj = self.value(hidden_states)

        # transpose
        query_states = self.transpose_for_scores(query_proj)
        key_states = self.transpose_for_scores(key_proj)
        value_states = self.transpose_for_scores(value_proj)

        # get scores
        scale = 1 / math.sqrt(math.sqrt(self.channels / self.num_heads))

        attention_scores = torch.matmul(query_states * scale, key_states.transpose(-1, -2) * scale)
        attention_probs = torch.softmax(attention_scores.float(), dim=-1).type(attention_scores.dtype)

        # compute attention output
        hidden_states = torch.matmul(attention_probs, value_states)

        hidden_states = hidden_states.permute(0, 2, 1, 3).contiguous()
        new_hidden_states_shape = hidden_states.size()[:-2] + (self.channels,)
        hidden_states = hidden_states.view(new_hidden_states_shape)

        # compute next hidden_states
        hidden_states = self.proj_attn(hidden_states)
        hidden_states = hidden_states.transpose(-1, -2).reshape(batch, channel, height, width)

        # res connect and rescale
        hidden_states = (hidden_states + residual) / self.rescale_output_factor
        return hidden_states


class SpatialTransformer(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, in_channels, n_heads, d_head, depth=1, dropout=0.0, context_dim=None):
        super().__init__()
        self.n_heads = n_heads
        self.d_head = d_head
        self.in_channels = in_channels
        inner_dim = n_heads * d_head
        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)

        self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)

        self.transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
                for d in range(depth)
            ]
        )

        self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)

    def _set_attention_slice(self, slice_size):
        for block in self.transformer_blocks:
            block._set_attention_slice(slice_size)

    def forward(self, x, context=None):
        # note: if no context is given, cross-attention defaults to self-attention
        b, c, h, w = x.shape
        x_in = x
        x = self.norm(x)
        x = self.proj_in(x)
        x = x.permute(0, 2, 3, 1).reshape(b, h * w, c)
        for block in self.transformer_blocks:
            x = block(x, context=context)
        x = x.reshape(b, h, w, c).permute(0, 3, 1, 2)
        x = self.proj_out(x)
        return x + x_in


class BasicTransformerBlock(nn.Module):
    def __init__(self, dim, n_heads, d_head, dropout=0.0, context_dim=None, gated_ff=True, checkpoint=True):
        super().__init__()
        self.attn1 = CrossAttention(
            query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout
        )  # is a self-attention
        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 = nn.LayerNorm(dim)
        self.norm2 = nn.LayerNorm(dim)
        self.norm3 = nn.LayerNorm(dim)
        self.checkpoint = checkpoint

    def _set_attention_slice(self, slice_size):
        self.attn1._slice_size = slice_size
        self.attn2._slice_size = slice_size

    def forward(self, x, context=None):
        x = self.attn1(self.norm1(x)) + x
        x = self.attn2(self.norm2(x), context=context) + x
        x = self.ff(self.norm3(x)) + x
        return x


class CrossAttention(nn.Module):
    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
        super().__init__()
        inner_dim = dim_head * heads
        context_dim = context_dim if context_dim is not None else query_dim

        self.scale = dim_head**-0.5
        self.heads = heads
        # for slice_size > 0 the attention score computation
        # is split across the batch axis to save memory
        # You can set slice_size with `set_attention_slice`
        self._slice_size = None

        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 reshape_heads_to_batch_dim(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.heads
        tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size * head_size, seq_len, dim // head_size)
        return tensor

    def reshape_batch_dim_to_heads(self, tensor):
        batch_size, seq_len, dim = tensor.shape
        head_size = self.heads
        tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
        tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
        return tensor

    def forward(self, x, context=None, mask=None):
        batch_size, sequence_length, dim = x.shape

        q = self.to_q(x)
        context = context if context is not None else x
        k = self.to_k(context)
        v = self.to_v(context)

        q = self.reshape_heads_to_batch_dim(q)
        k = self.reshape_heads_to_batch_dim(k)
        v = self.reshape_heads_to_batch_dim(v)

        # TODO(PVP) - mask is currently never used. Remember to re-implement when used

        # attention, what we cannot get enough of
        hidden_states = self._attention(q, k, v, sequence_length, dim)

        return self.to_out(hidden_states)

    def _attention(self, query, key, value, sequence_length, dim):
        batch_size_attention = query.shape[0]
        hidden_states = torch.zeros(
            (batch_size_attention, sequence_length, dim // self.heads), device=query.device, dtype=query.dtype
        )
        slice_size = self._slice_size if self._slice_size is not None else hidden_states.shape[0]
        for i in range(hidden_states.shape[0] // slice_size):
            start_idx = i * slice_size
            end_idx = (i + 1) * slice_size
            attn_slice = (
                torch.einsum("b i d, b j d -> b i j", query[start_idx:end_idx], key[start_idx:end_idx]) * self.scale
            )
            attn_slice = attn_slice.softmax(dim=-1)
            attn_slice = torch.einsum("b i j, b j d -> b i d", attn_slice, value[start_idx:end_idx])

            hidden_states[start_idx:end_idx] = attn_slice

        # reshape hidden_states
        hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
        return hidden_states


class FeedForward(nn.Module):
    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
        super().__init__()
        inner_dim = int(dim * mult)
        dim_out = dim_out if dim_out is not None else dim
        project_in = 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)


# 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)