Ross Wightman
3 years ago
commit
f45de37690
@ -0,0 +1,350 @@
|
||||
""" ConViT Model
|
||||
|
||||
@article{d2021convit,
|
||||
title={ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases},
|
||||
author={d'Ascoli, St{\'e}phane and Touvron, Hugo and Leavitt, Matthew and Morcos, Ari and Biroli, Giulio and Sagun, Levent},
|
||||
journal={arXiv preprint arXiv:2103.10697},
|
||||
year={2021}
|
||||
}
|
||||
|
||||
Paper link: https://arxiv.org/abs/2103.10697
|
||||
Original code: https://github.com/facebookresearch/convit, original copyright below
|
||||
"""
|
||||
# Copyright (c) 2015-present, Facebook, Inc.
|
||||
# All rights reserved.
|
||||
#
|
||||
# This source code is licensed under the CC-by-NC license found in the
|
||||
# LICENSE file in the root directory of this source tree.
|
||||
#
|
||||
'''These modules are adapted from those of timm, see
|
||||
https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
|
||||
'''
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
from functools import partial
|
||||
import torch.nn.functional as F
|
||||
|
||||
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
|
||||
from .helpers import build_model_with_cfg
|
||||
from .layers import DropPath, to_2tuple, trunc_normal_, PatchEmbed, Mlp
|
||||
from .registry import register_model
|
||||
from .vision_transformer_hybrid import HybridEmbed
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
|
||||
def _cfg(url='', **kwargs):
|
||||
return {
|
||||
'url': url,
|
||||
'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
|
||||
'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
|
||||
'first_conv': 'patch_embed.proj', 'classifier': 'head',
|
||||
**kwargs
|
||||
}
|
||||
|
||||
|
||||
default_cfgs = {
|
||||
# ConViT
|
||||
'convit_tiny': _cfg(
|
||||
url="https://dl.fbaipublicfiles.com/convit/convit_tiny.pth"),
|
||||
'convit_small': _cfg(
|
||||
url="https://dl.fbaipublicfiles.com/convit/convit_small.pth"),
|
||||
'convit_base': _cfg(
|
||||
url="https://dl.fbaipublicfiles.com/convit/convit_base.pth")
|
||||
}
|
||||
|
||||
|
||||
class GPSA(nn.Module):
|
||||
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.,
|
||||
locality_strength=1.):
|
||||
super().__init__()
|
||||
self.num_heads = num_heads
|
||||
self.dim = dim
|
||||
head_dim = dim // num_heads
|
||||
self.scale = qk_scale or head_dim ** -0.5
|
||||
self.locality_strength = locality_strength
|
||||
|
||||
self.qk = nn.Linear(dim, dim * 2, bias=qkv_bias)
|
||||
self.v = nn.Linear(dim, dim, bias=qkv_bias)
|
||||
|
||||
self.attn_drop = nn.Dropout(attn_drop)
|
||||
self.proj = nn.Linear(dim, dim)
|
||||
self.pos_proj = nn.Linear(3, num_heads)
|
||||
self.proj_drop = nn.Dropout(proj_drop)
|
||||
self.locality_strength = locality_strength
|
||||
self.gating_param = nn.Parameter(torch.ones(self.num_heads))
|
||||
self.rel_indices: torch.Tensor = torch.zeros(1, 1, 1, 3) # silly torchscript hack, won't work with None
|
||||
|
||||
def forward(self, x):
|
||||
B, N, C = x.shape
|
||||
if self.rel_indices is None or self.rel_indices.shape[1] != N:
|
||||
self.rel_indices = self.get_rel_indices(N)
|
||||
attn = self.get_attention(x)
|
||||
v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
||||
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
|
||||
def get_attention(self, x):
|
||||
B, N, C = x.shape
|
||||
qk = self.qk(x).reshape(B, N, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
||||
q, k = qk[0], qk[1]
|
||||
pos_score = self.rel_indices.expand(B, -1, -1, -1)
|
||||
pos_score = self.pos_proj(pos_score).permute(0, 3, 1, 2)
|
||||
patch_score = (q @ k.transpose(-2, -1)) * self.scale
|
||||
patch_score = patch_score.softmax(dim=-1)
|
||||
pos_score = pos_score.softmax(dim=-1)
|
||||
|
||||
gating = self.gating_param.view(1, -1, 1, 1)
|
||||
attn = (1. - torch.sigmoid(gating)) * patch_score + torch.sigmoid(gating) * pos_score
|
||||
attn /= attn.sum(dim=-1).unsqueeze(-1)
|
||||
attn = self.attn_drop(attn)
|
||||
return attn
|
||||
|
||||
def get_attention_map(self, x, return_map=False):
|
||||
attn_map = self.get_attention(x).mean(0) # average over batch
|
||||
distances = self.rel_indices.squeeze()[:, :, -1] ** .5
|
||||
dist = torch.einsum('nm,hnm->h', (distances, attn_map)) / distances.size(0)
|
||||
if return_map:
|
||||
return dist, attn_map
|
||||
else:
|
||||
return dist
|
||||
|
||||
def local_init(self):
|
||||
self.v.weight.data.copy_(torch.eye(self.dim))
|
||||
locality_distance = 1 # max(1,1/locality_strength**.5)
|
||||
|
||||
kernel_size = int(self.num_heads ** .5)
|
||||
center = (kernel_size - 1) / 2 if kernel_size % 2 == 0 else kernel_size // 2
|
||||
for h1 in range(kernel_size):
|
||||
for h2 in range(kernel_size):
|
||||
position = h1 + kernel_size * h2
|
||||
self.pos_proj.weight.data[position, 2] = -1
|
||||
self.pos_proj.weight.data[position, 1] = 2 * (h1 - center) * locality_distance
|
||||
self.pos_proj.weight.data[position, 0] = 2 * (h2 - center) * locality_distance
|
||||
self.pos_proj.weight.data *= self.locality_strength
|
||||
|
||||
def get_rel_indices(self, num_patches: int) -> torch.Tensor:
|
||||
img_size = int(num_patches ** .5)
|
||||
rel_indices = torch.zeros(1, num_patches, num_patches, 3)
|
||||
ind = torch.arange(img_size).view(1, -1) - torch.arange(img_size).view(-1, 1)
|
||||
indx = ind.repeat(img_size, img_size)
|
||||
indy = ind.repeat_interleave(img_size, dim=0).repeat_interleave(img_size, dim=1)
|
||||
indd = indx ** 2 + indy ** 2
|
||||
rel_indices[:, :, :, 2] = indd.unsqueeze(0)
|
||||
rel_indices[:, :, :, 1] = indy.unsqueeze(0)
|
||||
rel_indices[:, :, :, 0] = indx.unsqueeze(0)
|
||||
device = self.qk.weight.device
|
||||
return rel_indices.to(device)
|
||||
|
||||
|
||||
class MHSA(nn.Module):
|
||||
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
|
||||
super().__init__()
|
||||
self.num_heads = num_heads
|
||||
head_dim = dim // num_heads
|
||||
self.scale = qk_scale or head_dim ** -0.5
|
||||
|
||||
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
||||
self.attn_drop = nn.Dropout(attn_drop)
|
||||
self.proj = nn.Linear(dim, dim)
|
||||
self.proj_drop = nn.Dropout(proj_drop)
|
||||
|
||||
def get_attention_map(self, x, return_map=False):
|
||||
B, N, C = x.shape
|
||||
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
||||
q, k, v = qkv[0], qkv[1], qkv[2]
|
||||
attn_map = (q @ k.transpose(-2, -1)) * self.scale
|
||||
attn_map = attn_map.softmax(dim=-1).mean(0)
|
||||
|
||||
img_size = int(N ** .5)
|
||||
ind = torch.arange(img_size).view(1, -1) - torch.arange(img_size).view(-1, 1)
|
||||
indx = ind.repeat(img_size, img_size)
|
||||
indy = ind.repeat_interleave(img_size, dim=0).repeat_interleave(img_size, dim=1)
|
||||
indd = indx ** 2 + indy ** 2
|
||||
distances = indd ** .5
|
||||
distances = distances.to('cuda')
|
||||
|
||||
dist = torch.einsum('nm,hnm->h', (distances, attn_map)) / N
|
||||
if return_map:
|
||||
return dist, attn_map
|
||||
else:
|
||||
return dist
|
||||
|
||||
def forward(self, x):
|
||||
B, N, C = x.shape
|
||||
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
||||
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)
|
||||
|
||||
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
|
||||
x = self.proj(x)
|
||||
x = self.proj_drop(x)
|
||||
return x
|
||||
|
||||
|
||||
class Block(nn.Module):
|
||||
|
||||
def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
|
||||
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_gpsa=True, **kwargs):
|
||||
super().__init__()
|
||||
self.norm1 = norm_layer(dim)
|
||||
self.use_gpsa = use_gpsa
|
||||
if self.use_gpsa:
|
||||
self.attn = GPSA(
|
||||
dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop,
|
||||
proj_drop=drop, **kwargs)
|
||||
else:
|
||||
self.attn = MHSA(
|
||||
dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop,
|
||||
proj_drop=drop, **kwargs)
|
||||
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):
|
||||
x = x + self.drop_path(self.attn(self.norm1(x)))
|
||||
x = x + self.drop_path(self.mlp(self.norm2(x)))
|
||||
return x
|
||||
|
||||
|
||||
class ConViT(nn.Module):
|
||||
""" Vision Transformer with support for patch or hybrid CNN input stage
|
||||
"""
|
||||
|
||||
def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
|
||||
num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
|
||||
drop_path_rate=0., hybrid_backbone=None, norm_layer=nn.LayerNorm, global_pool=None,
|
||||
local_up_to_layer=3, locality_strength=1., use_pos_embed=True):
|
||||
super().__init__()
|
||||
embed_dim *= num_heads
|
||||
self.num_classes = num_classes
|
||||
self.local_up_to_layer = local_up_to_layer
|
||||
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
|
||||
self.locality_strength = locality_strength
|
||||
self.use_pos_embed = use_pos_embed
|
||||
|
||||
if hybrid_backbone is not None:
|
||||
self.patch_embed = HybridEmbed(
|
||||
hybrid_backbone, img_size=img_size, in_chans=in_chans, embed_dim=embed_dim)
|
||||
else:
|
||||
self.patch_embed = PatchEmbed(
|
||||
img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
|
||||
num_patches = self.patch_embed.num_patches
|
||||
self.num_patches = num_patches
|
||||
|
||||
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
|
||||
self.pos_drop = nn.Dropout(p=drop_rate)
|
||||
|
||||
if self.use_pos_embed:
|
||||
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
|
||||
trunc_normal_(self.pos_embed, std=.02)
|
||||
|
||||
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule
|
||||
self.blocks = nn.ModuleList([
|
||||
Block(
|
||||
dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
|
||||
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
|
||||
use_gpsa=True,
|
||||
locality_strength=locality_strength)
|
||||
if i < local_up_to_layer else
|
||||
Block(
|
||||
dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
|
||||
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
|
||||
use_gpsa=False)
|
||||
for i in range(depth)])
|
||||
self.norm = norm_layer(embed_dim)
|
||||
|
||||
# Classifier head
|
||||
self.feature_info = [dict(num_chs=embed_dim, reduction=0, module='head')]
|
||||
self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
||||
|
||||
trunc_normal_(self.cls_token, std=.02)
|
||||
self.apply(self._init_weights)
|
||||
for n, m in self.named_modules():
|
||||
if hasattr(m, 'local_init'):
|
||||
m.local_init()
|
||||
|
||||
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)
|
||||
|
||||
@torch.jit.ignore
|
||||
def no_weight_decay(self):
|
||||
return {'pos_embed', 'cls_token'}
|
||||
|
||||
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.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
||||
|
||||
def forward_features(self, x):
|
||||
B = x.shape[0]
|
||||
x = self.patch_embed(x)
|
||||
|
||||
cls_tokens = self.cls_token.expand(B, -1, -1)
|
||||
|
||||
if self.use_pos_embed:
|
||||
x = x + self.pos_embed
|
||||
x = self.pos_drop(x)
|
||||
|
||||
for u, blk in enumerate(self.blocks):
|
||||
if u == self.local_up_to_layer:
|
||||
x = torch.cat((cls_tokens, x), dim=1)
|
||||
x = blk(x)
|
||||
|
||||
x = self.norm(x)
|
||||
return x[:, 0]
|
||||
|
||||
def forward(self, x):
|
||||
x = self.forward_features(x)
|
||||
x = self.head(x)
|
||||
return x
|
||||
|
||||
|
||||
def _create_convit(variant, pretrained=False, **kwargs):
|
||||
return build_model_with_cfg(
|
||||
ConViT, variant, pretrained,
|
||||
default_cfg=default_cfgs[variant],
|
||||
**kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def convit_tiny(pretrained=False, **kwargs):
|
||||
model_args = dict(
|
||||
local_up_to_layer=10, locality_strength=1.0, embed_dim=48,
|
||||
num_heads=4, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
|
||||
model = _create_convit(variant='convit_tiny', pretrained=pretrained, **model_args)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def convit_small(pretrained=False, **kwargs):
|
||||
model_args = dict(
|
||||
local_up_to_layer=10, locality_strength=1.0, embed_dim=48,
|
||||
num_heads=9, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
|
||||
model = _create_convit(variant='convit_small', pretrained=pretrained, **model_args)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def convit_base(pretrained=False, **kwargs):
|
||||
model_args = dict(
|
||||
local_up_to_layer=10, locality_strength=1.0, embed_dim=48,
|
||||
num_heads=16, norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
|
||||
model = _create_convit(variant='convit_base', pretrained=pretrained, **model_args)
|
||||
return model
|
||||
@ -0,0 +1,431 @@
|
||||
""" 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 .layers import Mlp, DropPath, to_2tuple, trunc_normal_
|
||||
from .registry import register_model
|
||||
from .vision_transformer import Attention
|
||||
from .helpers import build_model_with_cfg, overlay_external_default_cfg
|
||||
|
||||
|
||||
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_embed.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 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
|
||||
|
||||
# def forward_mask(self, x, size: Size_):
|
||||
# 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
|
||||
# mask = torch.zeros((1, Hp, Wp), device=x.device)
|
||||
# mask[:, -pad_b:, :].fill_(1)
|
||||
# mask[:, :, -pad_r:].fill_(1)
|
||||
#
|
||||
# x = x.reshape(B, _h, self.ws, _w, self.ws, C).transpose(2, 3) # B, _h, _w, ws, ws, C
|
||||
# mask = mask.reshape(1, _h, self.ws, _w, self.ws).transpose(2, 3).reshape(1, _h * _w, self.ws * self.ws)
|
||||
# attn_mask = mask.unsqueeze(2) - mask.unsqueeze(3) # 1, _h*_w, ws*ws, ws*ws
|
||||
# attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-1000.0)).masked_fill(attn_mask == 0, float(0.0))
|
||||
# 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)
|
||||
# # n_h, B, _w*_h, nhead, ws*ws, dim
|
||||
# q, k, v = qkv[0], qkv[1], qkv[2] # B, _h*_w, n_head, ws*ws, dim_head
|
||||
# attn = (q @ k.transpose(-2, -1)) * self.scale # B, _h*_w, n_head, ws*ws, ws*ws
|
||||
# attn = attn + attn_mask.unsqueeze(2)
|
||||
# attn = attn.softmax(dim=-1)
|
||||
# attn = self.attn_drop(attn) # attn @v -> B, _h*_w, n_head, ws*ws, dim_head
|
||||
# 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 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):
|
||||
super().__init__()
|
||||
self.norm1 = norm_layer(dim)
|
||||
if ws is None:
|
||||
self.attn = Attention(dim, num_heads, False, None, attn_drop, drop)
|
||||
elif ws == 1:
|
||||
self.attn = GlobalSubSampleAttn(dim, num_heads, attn_drop, drop, sr_ratio)
|
||||
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_):
|
||||
x = x + self.drop_path(self.attn(self.norm1(x), size))
|
||||
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):
|
||||
super().__init__()
|
||||
self.num_classes = num_classes
|
||||
self.depths = depths
|
||||
|
||||
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(embed_dims[-1])
|
||||
|
||||
# classification head
|
||||
self.head = nn.Linear(embed_dims[-1], num_classes) if num_classes > 0 else nn.Identity()
|
||||
|
||||
# init weights
|
||||
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.embed_dim, 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, default_cfg=None, **kwargs):
|
||||
if default_cfg is None:
|
||||
default_cfg = deepcopy(default_cfgs[variant])
|
||||
overlay_external_default_cfg(default_cfg, kwargs)
|
||||
default_num_classes = default_cfg['num_classes']
|
||||
default_img_size = default_cfg['input_size'][-2:]
|
||||
|
||||
num_classes = kwargs.pop('num_classes', default_num_classes)
|
||||
img_size = kwargs.pop('img_size', default_img_size)
|
||||
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_cfg,
|
||||
img_size=img_size,
|
||||
num_classes=num_classes,
|
||||
**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)
|
||||
Loading…
Reference in new issue