You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
pytorch-image-models/timm/layers/halo_attn.py

234 lines
10 KiB

""" Halo Self Attention
Paper: `Scaling Local Self-Attention for Parameter Efficient Visual Backbones`
- https://arxiv.org/abs/2103.12731
@misc{2103.12731,
Author = {Ashish Vaswani and Prajit Ramachandran and Aravind Srinivas and Niki Parmar and Blake Hechtman and
Jonathon Shlens},
Title = {Scaling Local Self-Attention for Parameter Efficient Visual Backbones},
Year = {2021},
}
Status:
This impl is a WIP, there is no official ref impl and some details in paper weren't clear to me.
The attention mechanism works but it's slow as implemented.
Hacked together by / Copyright 2021 Ross Wightman
"""
3 years ago
from typing import List
import torch
from torch import nn
import torch.nn.functional as F
from .helpers import make_divisible
from .weight_init import trunc_normal_
3 years ago
from .trace_utils import _assert
def rel_logits_1d(q, rel_k, permute_mask: List[int]):
""" Compute relative logits along one dimension
As per: https://gist.github.com/aravindsrinivas/56359b79f0ce4449bcb04ab4b56a57a2
Originally from: `Attention Augmented Convolutional Networks` - https://arxiv.org/abs/1904.09925
Args:
q: (batch, height, width, dim)
rel_k: (2 * window - 1, dim)
permute_mask: permute output dim according to this
"""
B, H, W, dim = q.shape
rel_size = rel_k.shape[0]
win_size = (rel_size + 1) // 2
3 years ago
x = (q @ rel_k.transpose(-1, -2))
x = x.reshape(-1, W, rel_size)
# pad to shift from relative to absolute indexing
x_pad = F.pad(x, [0, 1]).flatten(1)
x_pad = F.pad(x_pad, [0, rel_size - W])
# reshape and slice out the padded elements
x_pad = x_pad.reshape(-1, W + 1, rel_size)
x = x_pad[:, :W, win_size - 1:]
# reshape and tile
x = x.reshape(B, H, 1, W, win_size).expand(-1, -1, win_size, -1, -1)
return x.permute(permute_mask)
class PosEmbedRel(nn.Module):
""" Relative Position Embedding
As per: https://gist.github.com/aravindsrinivas/56359b79f0ce4449bcb04ab4b56a57a2
Originally from: `Attention Augmented Convolutional Networks` - https://arxiv.org/abs/1904.09925
"""
def __init__(self, block_size, win_size, dim_head, scale):
"""
Args:
block_size (int): block size
win_size (int): neighbourhood window size
dim_head (int): attention head dim
scale (float): scale factor (for init)
"""
super().__init__()
self.block_size = block_size
self.dim_head = dim_head
self.height_rel = nn.Parameter(torch.randn(win_size * 2 - 1, dim_head) * scale)
self.width_rel = nn.Parameter(torch.randn(win_size * 2 - 1, dim_head) * scale)
def forward(self, q):
B, BB, HW, _ = q.shape
# relative logits in width dimension.
q = q.reshape(-1, self.block_size, self.block_size, self.dim_head)
rel_logits_w = rel_logits_1d(q, self.width_rel, permute_mask=(0, 1, 3, 2, 4))
# relative logits in height dimension.
q = q.transpose(1, 2)
rel_logits_h = rel_logits_1d(q, self.height_rel, permute_mask=(0, 3, 1, 4, 2))
rel_logits = rel_logits_h + rel_logits_w
rel_logits = rel_logits.reshape(B, BB, HW, -1)
return rel_logits
class HaloAttn(nn.Module):
""" Halo Attention
Paper: `Scaling Local Self-Attention for Parameter Efficient Visual Backbones`
- https://arxiv.org/abs/2103.12731
The internal dimensions of the attention module are controlled by the interaction of several arguments.
* the output dimension of the module is specified by dim_out, which falls back to input dim if not set
* the value (v) dimension is set to dim_out // num_heads, the v projection determines the output dim
* the query and key (qk) dimensions are determined by
* num_heads * dim_head if dim_head is not None
* num_heads * (dim_out * attn_ratio // num_heads) if dim_head is None
* as seen above, attn_ratio determines the ratio of q and k relative to the output if dim_head not used
Args:
dim (int): input dimension to the module
dim_out (int): output dimension of the module, same as dim if not set
feat_size (Tuple[int, int]): size of input feature_map (not used, for arg compat with bottle/lambda)
stride: output stride of the module, query downscaled if > 1 (default: 1).
num_heads: parallel attention heads (default: 8).
dim_head: dimension of query and key heads, calculated from dim_out * attn_ratio // num_heads if not set
block_size (int): size of blocks. (default: 8)
halo_size (int): size of halo overlap. (default: 3)
qk_ratio (float): ratio of q and k dimensions to output dimension when dim_head not set. (default: 1.0)
qkv_bias (bool) : add bias to q, k, and v projections
avg_down (bool): use average pool downsample instead of strided query blocks
scale_pos_embed (bool): scale the position embedding as well as Q @ K
"""
def __init__(
self, dim, dim_out=None, feat_size=None, stride=1, num_heads=8, dim_head=None, block_size=8, halo_size=3,
qk_ratio=1.0, qkv_bias=False, avg_down=False, scale_pos_embed=False):
super().__init__()
dim_out = dim_out or dim
assert dim_out % num_heads == 0
assert stride in (1, 2)
self.num_heads = num_heads
self.dim_head_qk = dim_head or make_divisible(dim_out * qk_ratio, divisor=8) // num_heads
self.dim_head_v = dim_out // self.num_heads
self.dim_out_qk = num_heads * self.dim_head_qk
self.dim_out_v = num_heads * self.dim_head_v
self.scale = self.dim_head_qk ** -0.5
self.scale_pos_embed = scale_pos_embed
self.block_size = self.block_size_ds = block_size
self.halo_size = halo_size
self.win_size = block_size + halo_size * 2 # neighbourhood window size
self.block_stride = 1
use_avg_pool = False
if stride > 1:
use_avg_pool = avg_down or block_size % stride != 0
self.block_stride = 1 if use_avg_pool else stride
self.block_size_ds = self.block_size // self.block_stride
# FIXME not clear if this stride behaviour is what the paper intended
# Also, the paper mentions using a 3D conv for dealing with the blocking/gather, and leaving
# data in unfolded block form. I haven't wrapped my head around how that'd look.
self.q = nn.Conv2d(dim, self.dim_out_qk, 1, stride=self.block_stride, bias=qkv_bias)
self.kv = nn.Conv2d(dim, self.dim_out_qk + self.dim_out_v, 1, bias=qkv_bias)
self.pos_embed = PosEmbedRel(
block_size=self.block_size_ds, win_size=self.win_size, dim_head=self.dim_head_qk, scale=self.scale)
self.pool = nn.AvgPool2d(2, 2) if use_avg_pool else nn.Identity()
self.reset_parameters()
def reset_parameters(self):
std = self.q.weight.shape[1] ** -0.5 # fan-in
trunc_normal_(self.q.weight, std=std)
trunc_normal_(self.kv.weight, std=std)
trunc_normal_(self.pos_embed.height_rel, std=self.scale)
trunc_normal_(self.pos_embed.width_rel, std=self.scale)
def forward(self, x):
B, C, H, W = x.shape
3 years ago
_assert(H % self.block_size == 0, '')
_assert(W % self.block_size == 0, '')
num_h_blocks = H // self.block_size
num_w_blocks = W // self.block_size
num_blocks = num_h_blocks * num_w_blocks
q = self.q(x)
# unfold
q = q.reshape(
-1, self.dim_head_qk,
num_h_blocks, self.block_size_ds, num_w_blocks, self.block_size_ds).permute(0, 1, 3, 5, 2, 4)
# B, num_heads * dim_head * block_size ** 2, num_blocks
q = q.reshape(B * self.num_heads, self.dim_head_qk, -1, num_blocks).transpose(1, 3)
# B * num_heads, num_blocks, block_size ** 2, dim_head
kv = self.kv(x)
# Generate overlapping windows for kv. This approach is good for GPU and CPU. However, unfold() is not
# lowered for PyTorch XLA so it will be very slow. See code at bottom of file for XLA friendly approach.
# FIXME figure out how to switch impl between this and conv2d if XLA being used.
kv = F.pad(kv, [self.halo_size, self.halo_size, self.halo_size, self.halo_size])
kv = kv.unfold(2, self.win_size, self.block_size).unfold(3, self.win_size, self.block_size).reshape(
B * self.num_heads, self.dim_head_qk + self.dim_head_v, num_blocks, -1).permute(0, 2, 3, 1)
k, v = torch.split(kv, [self.dim_head_qk, self.dim_head_v], dim=-1)
# B * num_heads, num_blocks, win_size ** 2, dim_head_qk or dim_head_v
if self.scale_pos_embed:
attn = (q @ k.transpose(-1, -2) + self.pos_embed(q)) * self.scale
else:
attn = (q @ k.transpose(-1, -2)) * self.scale + self.pos_embed(q)
# B * num_heads, num_blocks, block_size ** 2, win_size ** 2
attn = attn.softmax(dim=-1)
out = (attn @ v).transpose(1, 3) # B * num_heads, dim_head_v, block_size ** 2, num_blocks
# fold
out = out.reshape(-1, self.block_size_ds, self.block_size_ds, num_h_blocks, num_w_blocks)
out = out.permute(0, 3, 1, 4, 2).contiguous().view(
B, self.dim_out_v, H // self.block_stride, W // self.block_stride)
# B, dim_out, H // block_stride, W // block_stride
out = self.pool(out)
return out
""" Three alternatives for overlapping windows.
`.unfold().unfold()` is same speed as stride tricks with similar clarity as F.unfold()
if is_xla:
# This code achieves haloing on PyTorch XLA with reasonable runtime trade-off, it is
# EXTREMELY slow for backward on a GPU though so I need a way of selecting based on environment.
WW = self.win_size ** 2
pw = torch.eye(WW, dtype=x.dtype, device=x.device).reshape(WW, 1, self.win_size, self.win_size)
kv = F.conv2d(kv.reshape(-1, 1, H, W), pw, stride=self.block_size, padding=self.halo_size)
elif self.stride_tricks:
kv = F.pad(kv, [self.halo_size, self.halo_size, self.halo_size, self.halo_size]).contiguous()
kv = kv.as_strided((
B, self.dim_out_qk + self.dim_out_v, self.win_size, self.win_size, num_h_blocks, num_w_blocks),
stride=(kv.stride(0), kv.stride(1), kv.shape[-1], 1, self.block_size * kv.shape[-1], self.block_size))
else:
kv = F.unfold(kv, kernel_size=self.win_size, stride=self.block_size, padding=self.halo_size)
kv = kv.reshape(
B * self.num_heads, self.dim_head_qk + self.dim_head_v, -1, num_blocks).transpose(1, 3)
"""