SuyeonJ's picture
Upload folder using huggingface_hub
8d015d4 verified
""" Twins
A PyTorch impl of : `Twins: Revisiting the Design of Spatial Attention in Vision Transformers`
- https://arxiv.org/pdf/2104.13840.pdf
Code/weights from https://github.com/Meituan-AutoML/Twins, original copyright/license info below
"""
# --------------------------------------------------------
# Twins
# Copyright (c) 2021 Meituan
# Licensed under The Apache 2.0 License [see LICENSE for details]
# Written by Xinjie Li, Xiangxiang Chu
# --------------------------------------------------------
import math
from copy import deepcopy
from typing import Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from functools import partial
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.models.layers import Mlp, DropPath, to_2tuple, trunc_normal_
from timm.models.registry import register_model
from timm.models.vision_transformer import Attention
from .attention import MultiHeadAttention, LinearPositionEmbeddingSine
from .utils import coords_grid, bilinear_sampler, upflow8
def _cfg(url='', **kwargs):
return {
'url': url,
'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
'crop_pct': .9, 'interpolation': 'bicubic', 'fixed_input_size': True,
'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
'first_conv': 'patch_embeds.0.proj', 'classifier': 'head',
**kwargs
}
default_cfgs = {
'twins_pcpvt_small': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_pcpvt_small-e70e7e7a.pth',
),
'twins_pcpvt_base': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_pcpvt_base-e5ecb09b.pth',
),
'twins_pcpvt_large': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_pcpvt_large-d273f802.pth',
),
'twins_svt_small': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_svt_small-42e5f78c.pth',
),
'twins_svt_base': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_svt_base-c2265010.pth',
),
'twins_svt_large': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vt3p-weights/twins_svt_large-90f6aaa9.pth',
),
}
Size_ = Tuple[int, int]
class GroupAttnRPEContext(nn.Module):
""" Latent cost tokens attend to different group
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., ws=1, vert_c_dim=0):
super(GroupAttnRPEContext, self).__init__()
assert ws != 1
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
assert vert_c_dim > 0, "vert_c_dim should not be 0"
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.vert_c_dim = vert_c_dim
self.context_proj = nn.Linear(256, vert_c_dim)
self.q = nn.Linear(dim + vert_c_dim, dim, bias=True)
self.k = nn.Linear(dim + vert_c_dim, dim, bias=True)
self.v = nn.Linear(dim, dim, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.ws = ws
def forward(self, x, size: Size_, context=None):
B, N, C = x.shape
C_qk = C + self.vert_c_dim
H, W = size
batch_num = B // 5
context = context.repeat(B // context.shape[0], 1, 1, 1)
context = context.view(B, -1, H * W).permute(0, 2, 1)
context = self.context_proj(context)
context = context.view(B, H, W, -1)
x = x.view(B, H, W, C)
x_qk = torch.cat([x, context], dim=-1)
pad_l = pad_t = 0
pad_r = (self.ws - W % self.ws) % self.ws
pad_b = (self.ws - H % self.ws) % self.ws
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
x_qk = F.pad(x_qk, (0, 0, pad_l, pad_r, pad_t, pad_b))
_, Hp, Wp, _ = x.shape
_h, _w = Hp // self.ws, Wp // self.ws
padded_N = Hp * Wp
coords = coords_grid(B, Hp, Wp).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1)
coords_enc = LinearPositionEmbeddingSine(coords, dim=C_qk)
coords_enc = coords_enc.reshape(B, Hp, Wp, C_qk)
q = self.q(x_qk + coords_enc).reshape(B, _h, self.ws, _w, self.ws, self.num_heads,
C // self.num_heads).transpose(2, 3)
q = q.reshape(B, _h * _w, self.ws * self.ws, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4)
v = self.v(x)
k = self.k(x_qk + coords_enc)
# concate and do shifting operation together
kv = torch.cat([k, v], dim=-1)
kv_up = torch.cat(
[kv[:batch_num, self.ws:Hp, :, :],
kv[:batch_num, Hp - self.ws:Hp, :, :]], dim=1)
kv_down = torch.cat(
[kv[batch_num:batch_num * 2, :self.ws, :, :],
kv[batch_num:batch_num * 2, :Hp - self.ws, :, :]], dim=1)
kv_left = torch.cat(
[kv[batch_num * 2:batch_num * 3, :, self.ws:Wp, :],
kv[batch_num * 2:batch_num * 3, :, Wp - self.ws:Wp, :]], dim=2)
kv_right = torch.cat(
[kv[batch_num * 3:batch_num * 4, :, :self.ws, :],
kv[batch_num * 3:batch_num * 4, :, :Wp - self.ws, :]], dim=2)
kv_center = kv[batch_num * 4:batch_num * 5, :, :, :]
kv_shifted = torch.cat([kv_up, kv_down, kv_left, kv_right, kv_center], dim=0)
k, v = torch.split(kv_shifted, [self.dim, self.dim], dim=-1)
k = k.reshape(B, _h, self.ws, _w, self.ws, self.num_heads, C // self.num_heads).transpose(2, 3)
k = k.reshape(B, _h * _w, self.ws * self.ws, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4)
v = v.reshape(B, _h, self.ws, _w, self.ws, self.num_heads, C // self.num_heads).transpose(2, 3)
v = v.reshape(B, _h * _w, self.ws * self.ws, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4)
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
attn = (attn @ v).transpose(2, 3).reshape(B, _h, _w, self.ws, self.ws, C)
x = attn.transpose(2, 3).reshape(B, _h * self.ws, _w * self.ws, C)
if pad_r > 0 or pad_b > 0:
x = x[:, :H, :W, :].contiguous()
x = x.reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class GroupAttnRPE(nn.Module):
""" Latent cost tokens attend to different group
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., ws=1):
super(GroupAttnRPE, self).__init__()
assert ws != 1
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.q = nn.Linear(dim, dim, bias=True)
self.k = nn.Linear(dim, dim, bias=True)
self.v = nn.Linear(dim, dim, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.ws = ws
def forward(self, x, size: Size_, context=None):
B, N, C = x.shape
H, W = size
batch_num = B // 5
x = x.view(B, H, W, C)
pad_l = pad_t = 0
pad_r = (self.ws - W % self.ws) % self.ws
pad_b = (self.ws - H % self.ws) % self.ws
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
_, Hp, Wp, _ = x.shape
_h, _w = Hp // self.ws, Wp // self.ws
padded_N = Hp * Wp
coords = coords_grid(B, Hp, Wp).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1)
coords_enc = LinearPositionEmbeddingSine(coords, dim=C)
coords_enc = coords_enc.reshape(B, Hp, Wp, C)
q = self.q(x + coords_enc).reshape(B, _h, self.ws, _w, self.ws, self.num_heads, C // self.num_heads).transpose(
2, 3)
q = q.reshape(B, _h * _w, self.ws * self.ws, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4)
v = self.v(x)
k = self.k(x + coords_enc)
# concate and do shifting operation together
kv = torch.cat([k, v], dim=-1)
kv_up = torch.cat([kv[:batch_num, self.ws:Hp, :, :], kv[:batch_num, Hp - self.ws:Hp, :, :]], dim=1)
kv_down = torch.cat(
[kv[batch_num:batch_num * 2, :self.ws, :, :], kv[batch_num:batch_num * 2, :Hp - self.ws, :, :]], dim=1)
kv_left = torch.cat(
[kv[batch_num * 2:batch_num * 3, :, self.ws:Wp, :], kv[batch_num * 2:batch_num * 3, :, Wp - self.ws:Wp, :]],
dim=2)
kv_right = torch.cat(
[kv[batch_num * 3:batch_num * 4, :, :self.ws, :], kv[batch_num * 3:batch_num * 4, :, :Wp - self.ws, :]],
dim=2)
kv_center = kv[batch_num * 4:batch_num * 5, :, :, :]
kv_shifted = torch.cat([kv_up, kv_down, kv_left, kv_right, kv_center], dim=0)
k, v = torch.split(kv_shifted, [self.dim, self.dim], dim=-1)
k = k.reshape(B, _h, self.ws, _w, self.ws, self.num_heads, C // self.num_heads).transpose(2, 3)
k = k.reshape(B, _h * _w, self.ws * self.ws, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4)
v = v.reshape(B, _h, self.ws, _w, self.ws, self.num_heads, C // self.num_heads).transpose(2, 3)
v = v.reshape(B, _h * _w, self.ws * self.ws, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4)
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
attn = (attn @ v).transpose(2, 3).reshape(B, _h, _w, self.ws, self.ws, C)
x = attn.transpose(2, 3).reshape(B, _h * self.ws, _w * self.ws, C)
if pad_r > 0 or pad_b > 0:
x = x[:, :H, :W, :].contiguous()
x = x.reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class LocallyGroupedAttnRPEContext(nn.Module):
""" LSA: self attention within a group
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., ws=1, vert_c_dim=0):
assert ws != 1
super(LocallyGroupedAttnRPEContext, self).__init__()
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.vert_c_dim = vert_c_dim
self.context_proj = nn.Linear(256, vert_c_dim)
# context are not added to value
self.q = nn.Linear(dim + vert_c_dim, dim, bias=True)
self.k = nn.Linear(dim + vert_c_dim, dim, bias=True)
self.v = nn.Linear(dim, dim, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.ws = ws
def forward(self, x, size: Size_, context=None):
# There are two implementations for this function, zero padding or mask. We don't observe obvious difference for
# both. You can choose any one, we recommend forward_padding because it's neat. However,
# the masking implementation is more reasonable and accurate.
B, N, C = x.shape
H, W = size
C_qk = C + self.vert_c_dim
context = context.repeat(B // context.shape[0], 1, 1, 1)
context = context.view(B, -1, H * W).permute(0, 2, 1)
context = self.context_proj(context)
context = context.view(B, H, W, -1)
x = x.view(B, H, W, C)
x_qk = torch.cat([x, context], dim=-1)
pad_l = pad_t = 0
pad_r = (self.ws - W % self.ws) % self.ws
pad_b = (self.ws - H % self.ws) % self.ws
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
x_qk = F.pad(x_qk, (0, 0, pad_l, pad_r, pad_t, pad_b))
_, Hp, Wp, _ = x.shape
_h, _w = Hp // self.ws, Wp // self.ws
x = x.reshape(B, _h, self.ws, _w, self.ws, C).transpose(2, 3)
x_qk = x_qk.reshape(B, _h, self.ws, _w, self.ws, C_qk).transpose(2, 3)
v = self.v(x).reshape(
B, _h * _w, self.ws * self.ws, 1, self.num_heads, C // self.num_heads).permute(3, 0, 1, 4, 2, 5)[0]
coords = coords_grid(B, self.ws, self.ws).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1)
coords_enc = LinearPositionEmbeddingSine(coords, dim=C_qk).view(B, self.ws, self.ws, C_qk)
# coords_enc: B, ws, ws, C
# x: B, _h, _w, self.ws, self.ws, C
x_qk = x_qk + coords_enc[:, None, None, :, :, :]
q = self.q(x_qk).reshape(
B, _h * _w, self.ws * self.ws, 1, self.num_heads, C // self.num_heads).permute(3, 0, 1, 4, 2, 5)[0]
k = self.k(x_qk).reshape(
B, _h * _w, self.ws * self.ws, 1, self.num_heads, C // self.num_heads).permute(3, 0, 1, 4, 2, 5)[0]
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
attn = (attn @ v).transpose(2, 3).reshape(B, _h, _w, self.ws, self.ws, C)
x = attn.transpose(2, 3).reshape(B, _h * self.ws, _w * self.ws, C)
if pad_r > 0 or pad_b > 0:
x = x[:, :H, :W, :].contiguous()
x = x.reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class GlobalSubSampleAttnRPEContext(nn.Module):
""" GSA: using a key to summarize the information for a group to be efficient.
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., sr_ratio=1, vert_c_dim=0):
super().__init__()
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.vert_c_dim = vert_c_dim
self.context_proj = nn.Linear(256, vert_c_dim)
self.q = nn.Linear(dim + vert_c_dim, dim, bias=True)
self.k = nn.Linear(dim, dim, bias=True)
self.v = nn.Linear(dim, dim, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.sr_ratio = sr_ratio
if sr_ratio > 1:
self.sr_key = nn.Conv2d(dim + vert_c_dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.sr_value = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.norm = nn.LayerNorm(dim)
else:
self.sr = None
self.norm = None
def forward(self, x, size: Size_, context=None):
B, N, C = x.shape
C_qk = C + self.vert_c_dim
H, W = size
context = context.repeat(B // context.shape[0], 1, 1, 1)
context = context.view(B, -1, H * W).permute(0, 2, 1)
context = self.context_proj(context)
context = context.view(B, H, W, -1)
x = x.view(B, H, W, C)
x_qk = torch.cat([x, context], dim=-1)
pad_l = pad_t = 0
pad_r = (self.sr_ratio - W % self.sr_ratio) % self.sr_ratio
pad_b = (self.sr_ratio - H % self.sr_ratio) % self.sr_ratio
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
x_qk = F.pad(x_qk, (0, 0, pad_l, pad_r, pad_t, pad_b))
_, Hp, Wp, _ = x.shape
padded_size = (Hp, Wp)
padded_N = Hp * Wp
x = x.view(B, -1, C)
x_qk = x_qk.view(B, -1, C_qk)
coords = coords_grid(B, *padded_size).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1)
coords_enc = LinearPositionEmbeddingSine(coords, dim=C_qk)
# coords_enc: B, Hp*Wp, C
# x: B, Hp*Wp, C
q = self.q(x_qk + coords_enc).reshape(B, padded_N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
if self.sr_key is not None:
x = x.permute(0, 2, 1).reshape(B, C, *padded_size)
x_qk = x_qk.permute(0, 2, 1).reshape(B, C_qk, *padded_size)
x = self.sr_value(x).reshape(B, C, -1).permute(0, 2, 1)
x_qk = self.sr_key(x_qk).reshape(B, C, -1).permute(0, 2, 1)
x = self.norm(x)
x_qk = self.norm(x_qk)
coords = coords_grid(B, padded_size[0] // self.sr_ratio, padded_size[1] // self.sr_ratio).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1) * self.sr_ratio
# align the coordinate of local and global
coords_enc = LinearPositionEmbeddingSine(coords, dim=C)
k = self.k(x_qk + coords_enc).reshape(B, (padded_size[0] // self.sr_ratio) * (padded_size[1] // self.sr_ratio),
self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
v = self.v(x).reshape(B, (padded_size[0] // self.sr_ratio) * (padded_size[1] // self.sr_ratio), self.num_heads,
C // self.num_heads).permute(0, 2, 1, 3)
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, Hp, Wp, C)
if pad_r > 0 or pad_b > 0:
x = x[:, :H, :W, :].contiguous()
x = x.reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class LocallyGroupedAttnRPE(nn.Module):
""" LSA: self attention within a group
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., ws=1):
assert ws != 1
super(LocallyGroupedAttnRPE, self).__init__()
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.q = nn.Linear(dim, dim, bias=True)
self.k = nn.Linear(dim, dim, bias=True)
self.v = nn.Linear(dim, dim, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.ws = ws
def forward(self, x, size: Size_, context=None):
# There are two implementations for this function, zero padding or mask. We don't observe obvious difference for
# both. You can choose any one, we recommend forward_padding because it's neat. However,
# the masking implementation is more reasonable and accurate.
B, N, C = x.shape
H, W = size
x = x.view(B, H, W, C)
pad_l = pad_t = 0
pad_r = (self.ws - W % self.ws) % self.ws
pad_b = (self.ws - H % self.ws) % self.ws
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
_, Hp, Wp, _ = x.shape
_h, _w = Hp // self.ws, Wp // self.ws
x = x.reshape(B, _h, self.ws, _w, self.ws, C).transpose(2, 3)
v = self.v(x).reshape(
B, _h * _w, self.ws * self.ws, 1, self.num_heads, C // self.num_heads).permute(3, 0, 1, 4, 2, 5)[0]
coords = coords_grid(B, self.ws, self.ws).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1)
coords_enc = LinearPositionEmbeddingSine(coords, dim=C).view(B, self.ws, self.ws, C)
# coords_enc: B, ws, ws, C
# x: B, _h, _w, self.ws, self.ws, C
x = x + coords_enc[:, None, None, :, :, :]
q = self.q(x).reshape(
B, _h * _w, self.ws * self.ws, 1, self.num_heads, C // self.num_heads).permute(3, 0, 1, 4, 2, 5)[0]
k = self.k(x).reshape(
B, _h * _w, self.ws * self.ws, 1, self.num_heads, C // self.num_heads).permute(3, 0, 1, 4, 2, 5)[0]
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
attn = (attn @ v).transpose(2, 3).reshape(B, _h, _w, self.ws, self.ws, C)
x = attn.transpose(2, 3).reshape(B, _h * self.ws, _w * self.ws, C)
if pad_r > 0 or pad_b > 0:
x = x[:, :H, :W, :].contiguous()
x = x.reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class GlobalSubSampleAttnRPE(nn.Module):
""" GSA: using a key to summarize the information for a group to be efficient.
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., sr_ratio=1):
super().__init__()
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.q = nn.Linear(dim, dim, bias=True)
self.k = nn.Linear(dim, dim, bias=True)
self.v = nn.Linear(dim, dim, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.sr_ratio = sr_ratio
if sr_ratio > 1:
self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.norm = nn.LayerNorm(dim)
else:
self.sr = None
self.norm = None
def forward(self, x, size: Size_, context=None):
B, N, C = x.shape
H, W = size
x = x.view(B, H, W, C)
pad_l = pad_t = 0
pad_r = (self.sr_ratio - W % self.sr_ratio) % self.sr_ratio
pad_b = (self.sr_ratio - H % self.sr_ratio) % self.sr_ratio
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
_, Hp, Wp, _ = x.shape
padded_size = (Hp, Wp)
padded_N = Hp * Wp
x = x.view(B, -1, C)
coords = coords_grid(B, *padded_size).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1)
coords_enc = LinearPositionEmbeddingSine(coords, dim=C)
# coords_enc: B, Hp*Wp, C
# x: B, Hp*Wp, C
q = self.q(x + coords_enc).reshape(B, padded_N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
if self.sr is not None:
x = x.permute(0, 2, 1).reshape(B, C, *padded_size)
x = self.sr(x).reshape(B, C, -1).permute(0, 2, 1)
x = self.norm(x)
coords = coords_grid(B, padded_size[0] // self.sr_ratio, padded_size[1] // self.sr_ratio).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1) * self.sr_ratio
# align the coordinate of local and global
coords_enc = LinearPositionEmbeddingSine(coords, dim=C)
k = self.k(x + coords_enc).reshape(B, (padded_size[0] // self.sr_ratio) * (padded_size[1] // self.sr_ratio),
self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
v = self.v(x).reshape(B, (padded_size[0] // self.sr_ratio) * (padded_size[1] // self.sr_ratio), self.num_heads,
C // self.num_heads).permute(0, 2, 1, 3)
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, Hp, Wp, C)
if pad_r > 0 or pad_b > 0:
x = x[:, :H, :W, :].contiguous()
x = x.reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class CrossGlobalSubSampleAttnRPE(nn.Module):
""" GSA: using a key to summarize the information for a group to be efficient.
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., sr_ratio=1):
super().__init__()
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.q = nn.Linear(dim, dim, bias=True)
self.k = nn.Linear(dim, dim, bias=True)
self.v = nn.Linear(dim, dim, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.sr_ratio = sr_ratio
if sr_ratio > 1:
self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.norm = nn.LayerNorm(dim)
else:
self.sr = None
self.norm = None
def forward(self, x, tgt, size: Size_):
B, N, C = x.shape
coords = coords_grid(B, *size).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1)
coords_enc = LinearPositionEmbeddingSine(coords, dim=C)
# coords_enc: B, H*W, C
# x: B, H*W, C
q = self.q(x + coords_enc).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
if self.sr is not None:
tgt = tgt.permute(0, 2, 1).reshape(B, C, *size)
tgt = self.sr(tgt).reshape(B, C, -1).permute(0, 2, 1)
tgt = self.norm(tgt)
coords = coords_grid(B, size[0] // self.sr_ratio, size[1] // self.sr_ratio).to(x.device)
coords = coords.view(B, 2, -1).permute(0, 2, 1) * self.sr_ratio
# align the coordinate of local and global
coords_enc = LinearPositionEmbeddingSine(coords, dim=C)
k = self.k(tgt + coords_enc).reshape(B, (size[0] // self.sr_ratio) * (size[1] // self.sr_ratio), self.num_heads,
C // self.num_heads).permute(0, 2, 1, 3)
v = self.v(tgt).reshape(B, (size[0] // self.sr_ratio) * (size[1] // self.sr_ratio), self.num_heads,
C // self.num_heads).permute(0, 2, 1, 3)
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class LocallyGroupedAttn(nn.Module):
""" LSA: self attention within a group
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., ws=1):
assert ws != 1
super(LocallyGroupedAttn, self).__init__()
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.qkv = nn.Linear(dim, dim * 3, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.ws = ws
def forward(self, x, size: Size_):
# There are two implementations for this function, zero padding or mask. We don't observe obvious difference for
# both. You can choose any one, we recommend forward_padding because it's neat. However,
# the masking implementation is more reasonable and accurate.
B, N, C = x.shape
H, W = size
x = x.view(B, H, W, C)
pad_l = pad_t = 0
pad_r = (self.ws - W % self.ws) % self.ws
pad_b = (self.ws - H % self.ws) % self.ws
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
_, Hp, Wp, _ = x.shape
_h, _w = Hp // self.ws, Wp // self.ws
x = x.reshape(B, _h, self.ws, _w, self.ws, C).transpose(2, 3)
qkv = self.qkv(x).reshape(
B, _h * _w, self.ws * self.ws, 3, self.num_heads, C // self.num_heads).permute(3, 0, 1, 4, 2, 5)
q, k, v = qkv[0], qkv[1], qkv[2]
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
attn = (attn @ v).transpose(2, 3).reshape(B, _h, _w, self.ws, self.ws, C)
x = attn.transpose(2, 3).reshape(B, _h * self.ws, _w * self.ws, C)
if pad_r > 0 or pad_b > 0:
x = x[:, :H, :W, :].contiguous()
x = x.reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class GlobalSubSampleAttn(nn.Module):
""" GSA: using a key to summarize the information for a group to be efficient.
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., sr_ratio=1):
super().__init__()
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.q = nn.Linear(dim, dim, bias=True)
self.kv = nn.Linear(dim, dim * 2, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.sr_ratio = sr_ratio
if sr_ratio > 1:
self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.norm = nn.LayerNorm(dim)
else:
self.sr = None
self.norm = None
def forward(self, x, size: Size_):
B, N, C = x.shape
q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
if self.sr is not None:
x = x.permute(0, 2, 1).reshape(B, C, *size)
x = self.sr(x).reshape(B, C, -1).permute(0, 2, 1)
x = self.norm(x)
kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
k, v = kv[0], kv[1]
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class CrossGlobalSubSampleAttn(nn.Module):
""" GSA: using a key to summarize the information for a group to be efficient.
"""
def __init__(self, dim, num_heads=8, attn_drop=0., proj_drop=0., sr_ratio=1):
super().__init__()
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim ** -0.5
self.q = nn.Linear(dim, dim, bias=True)
self.kv = nn.Linear(dim, dim * 2, bias=True)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.sr_ratio = sr_ratio
if sr_ratio > 1:
self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
self.norm = nn.LayerNorm(dim)
else:
self.sr = None
self.norm = None
def forward(self, x, tgt, size: Size_):
B, N, C = x.shape
q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
if self.sr is not None:
tgt = tgt.permute(0, 2, 1).reshape(B, C, *size)
tgt = self.sr(tgt).reshape(B, C, -1).permute(0, 2, 1)
tgt = self.norm(tgt)
kv = self.kv(tgt).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
k, v = kv[0], kv[1]
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class CrossBlock(nn.Module):
def __init__(self, dim, num_heads, mlp_ratio=4., drop=0., attn_drop=0., drop_path=0.,
act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1, ws=None, with_rpe=True):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = CrossGlobalSubSampleAttnRPE(dim, num_heads, attn_drop, drop, sr_ratio)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
def forward(self, src, tgt, size: Size_):
src_shortcut, tgt_shortcut = src, tgt
src, tgt = self.norm1(src), self.norm1(tgt)
src = src_shortcut + self.drop_path(self.attn(src, tgt, size))
tgt = tgt_shortcut + self.drop_path(self.attn(tgt, src, size))
src = src + self.drop_path(self.mlp(self.norm2(src)))
tgt = tgt + self.drop_path(self.mlp(self.norm2(tgt)))
return src, tgt
class Block(nn.Module):
def __init__(self, dim, num_heads, mlp_ratio=4., drop=0., attn_drop=0., drop_path=0.,
act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1, ws=None, with_rpe=False, vert_c_dim=0):
super().__init__()
self.norm1 = norm_layer(dim)
if ws == 1:
if with_rpe:
if vert_c_dim > 0:
self.attn = GlobalSubSampleAttnRPEContext(dim, num_heads, attn_drop, drop, sr_ratio, vert_c_dim)
else:
self.attn = GlobalSubSampleAttnRPE(dim, num_heads, attn_drop, drop, sr_ratio)
else:
self.attn = GlobalSubSampleAttn(dim, num_heads, attn_drop, drop, sr_ratio)
else:
if with_rpe:
if vert_c_dim > 0:
self.attn = LocallyGroupedAttnRPEContext(dim, num_heads, attn_drop, drop, ws, vert_c_dim)
else:
self.attn = LocallyGroupedAttnRPE(dim, num_heads, attn_drop, drop, ws)
else:
self.attn = LocallyGroupedAttn(dim, num_heads, attn_drop, drop, ws)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
def forward(self, x, size: Size_, context=None):
x = x + self.drop_path(self.attn(self.norm1(x), size, context))
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
class PosConv(nn.Module):
# PEG from https://arxiv.org/abs/2102.10882
def __init__(self, in_chans, embed_dim=768, stride=1):
super(PosConv, self).__init__()
self.proj = nn.Sequential(nn.Conv2d(in_chans, embed_dim, 3, stride, 1, bias=True, groups=embed_dim), )
self.stride = stride
def forward(self, x, size: Size_):
B, N, C = x.shape
cnn_feat_token = x.transpose(1, 2).view(B, C, *size)
x = self.proj(cnn_feat_token)
if self.stride == 1:
x += cnn_feat_token
x = x.flatten(2).transpose(1, 2)
return x
def no_weight_decay(self):
return ['proj.%d.weight' % i for i in range(4)]
class PatchEmbed(nn.Module):
""" Image to Patch Embedding
"""
def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
self.img_size = img_size
self.patch_size = patch_size
assert img_size[0] % patch_size[0] == 0 and img_size[1] % patch_size[1] == 0, \
f"img_size {img_size} should be divided by patch_size {patch_size}."
self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
self.num_patches = self.H * self.W
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
self.norm = nn.LayerNorm(embed_dim)
def forward(self, x) -> Tuple[torch.Tensor, Size_]:
B, C, H, W = x.shape
x = self.proj(x).flatten(2).transpose(1, 2)
x = self.norm(x)
out_size = (H // self.patch_size[0], W // self.patch_size[1])
return x, out_size
class Twins(nn.Module):
""" Twins Vision Transfomer (Revisiting Spatial Attention)
Adapted from PVT (PyramidVisionTransformer) class at https://github.com/whai362/PVT.git
"""
def __init__(
self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dims=(64, 128, 256, 512),
num_heads=(1, 2, 4, 8), mlp_ratios=(4, 4, 4, 4), drop_rate=0., attn_drop_rate=0., drop_path_rate=0.,
norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=(3, 4, 6, 3), sr_ratios=(8, 4, 2, 1), wss=None,
block_cls=Block, init_weight=True):
super().__init__()
self.num_classes = num_classes
self.depths = depths
self.embed_dims = embed_dims
self.num_features = embed_dims[-1]
img_size = to_2tuple(img_size)
prev_chs = in_chans
self.patch_embeds = nn.ModuleList()
self.pos_drops = nn.ModuleList()
for i in range(len(depths)):
self.patch_embeds.append(PatchEmbed(img_size, patch_size, prev_chs, embed_dims[i]))
self.pos_drops.append(nn.Dropout(p=drop_rate))
prev_chs = embed_dims[i]
img_size = tuple(t // patch_size for t in img_size)
patch_size = 2
self.blocks = nn.ModuleList()
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
cur = 0
for k in range(len(depths)):
_block = nn.ModuleList([block_cls(
dim=embed_dims[k], num_heads=num_heads[k], mlp_ratio=mlp_ratios[k], drop=drop_rate,
attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer, sr_ratio=sr_ratios[k],
ws=1 if wss is None or i % 2 == 1 else wss[k]) for i in range(depths[k])])
self.blocks.append(_block)
cur += depths[k]
self.pos_block = nn.ModuleList([PosConv(embed_dim, embed_dim) for embed_dim in embed_dims])
self.norm = norm_layer(self.num_features)
# classification head
self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
# init weights
if init_weight:
self.apply(self._init_weights)
@torch.jit.ignore
def no_weight_decay(self):
return set(['pos_block.' + n for n, p in self.pos_block.named_parameters()])
def get_classifier(self):
return self.head
def reset_classifier(self, num_classes, global_pool=''):
self.num_classes = num_classes
self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
elif isinstance(m, nn.Conv2d):
fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
fan_out //= m.groups
m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
if m.bias is not None:
m.bias.data.zero_()
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1.0)
m.bias.data.zero_()
def forward_features(self, x):
B = x.shape[0]
for i, (embed, drop, blocks, pos_blk) in enumerate(
zip(self.patch_embeds, self.pos_drops, self.blocks, self.pos_block)):
x, size = embed(x)
x = drop(x)
for j, blk in enumerate(blocks):
x = blk(x, size)
if j == 0:
x = pos_blk(x, size) # PEG here
if i < len(self.depths) - 1:
x = x.reshape(B, *size, -1).permute(0, 3, 1, 2).contiguous()
x = self.norm(x)
return x.mean(dim=1) # GAP here
def forward(self, x):
x = self.forward_features(x)
x = self.head(x)
return x
# def _create_twins(variant, pretrained=False, **kwargs):
# if kwargs.get('features_only', None):
# raise RuntimeError('features_only not implemented for Vision Transformer models.')
# model = build_model_with_cfg(
# Twins, variant, pretrained,
# default_cfg=default_cfgs[variant],
# **kwargs)
# return model
# @register_model
# def twins_pcpvt_small(pretrained=False, **kwargs):
# model_kwargs = dict(
# patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
# depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], **kwargs)
# return _create_twins('twins_pcpvt_small', pretrained=pretrained, **model_kwargs)
# @register_model
# def twins_pcpvt_base(pretrained=False, **kwargs):
# model_kwargs = dict(
# patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
# depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1], **kwargs)
# return _create_twins('twins_pcpvt_base', pretrained=pretrained, **model_kwargs)
# @register_model
# def twins_pcpvt_large(pretrained=False, **kwargs):
# model_kwargs = dict(
# patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[8, 8, 4, 4],
# depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1], **kwargs)
# return _create_twins('twins_pcpvt_large', pretrained=pretrained, **model_kwargs)
# @register_model
# def twins_svt_small(pretrained=False, **kwargs):
# model_kwargs = dict(
# patch_size=4, embed_dims=[64, 128, 256, 512], num_heads=[2, 4, 8, 16], mlp_ratios=[4, 4, 4, 4],
# depths=[2, 2, 10, 4], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], **kwargs)
# return _create_twins('twins_svt_small', pretrained=pretrained, **model_kwargs)
# @register_model
# def twins_svt_base(pretrained=False, **kwargs):
# model_kwargs = dict(
# patch_size=4, embed_dims=[96, 192, 384, 768], num_heads=[3, 6, 12, 24], mlp_ratios=[4, 4, 4, 4],
# depths=[2, 2, 18, 2], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], **kwargs)
# return _create_twins('twins_svt_base', pretrained=pretrained, **model_kwargs)
# @register_model
# def twins_svt_large(pretrained=False, **kwargs):
# model_kwargs = dict(
# patch_size=4, embed_dims=[128, 256, 512, 1024], num_heads=[4, 8, 16, 32], mlp_ratios=[4, 4, 4, 4],
# depths=[2, 2, 18, 2], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], **kwargs)
# return _create_twins('twins_svt_large', pretrained=pretrained, **model_kwargs)
# @register_model
# def twins_svt_large_context(pretrained=False, **kwargs):
# model_kwargs = dict(
# patch_size=4, embed_dims=[128, 256, 512, 1024], num_heads=[4, 8, 16, 32], mlp_ratios=[4, 4, 4, 4],
# depths=[2, 2, 18, 2], wss=[7, 7, 7, 7], sr_ratios=[8, 4, 2, 1], in_chans=6, init_weight=False, **kwargs)
# return _create_twins('twins_svt_large', pretrained=pretrained, **model_kwargs)
# # def twins_svt_large_context(pretrained=False, **kwargs):
# # model_kwargs = dict(
# # patch_size=4, embed_dims=[128, 256], num_heads=[4, 8], mlp_ratios=[4, 4],
# # depths=[2, 2], wss=[7, 7], sr_ratios=[8, 4], in_chans=6, init_weight=False, **kwargs)
# # return _create_twins('twins_svt_large', pretrained=pretrained, **model_kwargs)