2025-05-06 02:59:33 +08:00

602 lines
22 KiB
Python

"""SwiftFormer
SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications
Code: https://github.com/Amshaker/SwiftFormer
Paper: https://arxiv.org/pdf/2303.15446
@InProceedings{Shaker_2023_ICCV,
author = {Shaker, Abdelrahman and Maaz, Muhammad and Rasheed, Hanoona and Khan, Salman and Yang, Ming-Hsuan and Khan, Fahad Shahbaz},
title = {SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
year = {2023},
}
"""
import re
from typing import Any, Dict, List, Optional, Set, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.layers import DropPath, Linear, LayerType, to_2tuple, trunc_normal_
from ._builder import build_model_with_cfg
from ._features import feature_take_indices
from ._manipulate import checkpoint_seq
from ._registry import generate_default_cfgs, register_model
__all__ = ['SwiftFormer']
class LayerScale2d(nn.Module):
def __init__(self, dim: int, init_values: float = 1e-5, inplace: bool = False):
super().__init__()
self.inplace = inplace
self.gamma = nn.Parameter(
init_values * torch.ones(dim, 1, 1), requires_grad=True)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return x.mul_(self.gamma) if self.inplace else x * self.gamma
class Embedding(nn.Module):
"""
Patch Embedding that is implemented by a layer of conv.
Input: tensor in shape [B, C, H, W]
Output: tensor in shape [B, C, H/stride, W/stride]
"""
def __init__(
self,
in_chans: int = 3,
embed_dim: int = 768,
patch_size: int = 16,
stride: int = 16,
padding: int = 0,
norm_layer: LayerType = nn.BatchNorm2d,
):
super().__init__()
patch_size = to_2tuple(patch_size)
stride = to_2tuple(stride)
padding = to_2tuple(padding)
self.proj = nn.Conv2d(in_chans, embed_dim, patch_size, stride, padding)
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.proj(x)
x = self.norm(x)
return x
class ConvEncoder(nn.Module):
"""
Implementation of ConvEncoder with 3*3 and 1*1 convolutions.
Input: tensor with shape [B, C, H, W]
Output: tensor with shape [B, C, H, W]
"""
def __init__(
self,
dim: int,
hidden_dim: int = 64,
kernel_size: int = 3,
drop_path: float = 0.,
act_layer: LayerType = nn.GELU,
norm_layer: LayerType = nn.BatchNorm2d,
use_layer_scale: bool = True,
):
super().__init__()
self.dwconv = nn.Conv2d(dim, dim, kernel_size, padding=kernel_size // 2, groups=dim)
self.norm = norm_layer(dim)
self.pwconv1 = nn.Conv2d(dim, hidden_dim, 1)
self.act = act_layer()
self.pwconv2 = nn.Conv2d(hidden_dim, dim, 1)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.layer_scale = LayerScale2d(dim, 1) if use_layer_scale else nn.Identity()
def forward(self, x: torch.Tensor) -> torch.Tensor:
input = x
x = self.dwconv(x)
x = self.norm(x)
x = self.pwconv1(x)
x = self.act(x)
x = self.pwconv2(x)
x = self.layer_scale(x)
x = input + self.drop_path(x)
return x
class Mlp(nn.Module):
"""
Implementation of MLP layer with 1*1 convolutions.
Input: tensor with shape [B, C, H, W]
Output: tensor with shape [B, C, H, W]
"""
def __init__(
self,
in_features: int,
hidden_features: Optional[int] = None,
out_features: Optional[int] = None,
act_layer: LayerType = nn.GELU,
norm_layer: LayerType = nn.BatchNorm2d,
drop: float = 0.,
):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.norm1 = norm_layer(in_features)
self.fc1 = nn.Conv2d(in_features, hidden_features, 1)
self.act = act_layer()
self.fc2 = nn.Conv2d(hidden_features, out_features, 1)
self.drop = nn.Dropout(drop)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.norm1(x)
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
class EfficientAdditiveAttention(nn.Module):
"""
Efficient Additive Attention module for SwiftFormer.
Input: tensor in shape [B, C, H, W]
Output: tensor in shape [B, C, H, W]
"""
def __init__(self, in_dims: int = 512, token_dim: int = 256, num_heads: int = 1):
super().__init__()
self.scale_factor = token_dim ** -0.5
self.to_query = nn.Linear(in_dims, token_dim * num_heads)
self.to_key = nn.Linear(in_dims, token_dim * num_heads)
self.w_g = nn.Parameter(torch.randn(token_dim * num_heads, 1))
self.proj = nn.Linear(token_dim * num_heads, token_dim * num_heads)
self.final = nn.Linear(token_dim * num_heads, token_dim)
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, _, H, W = x.shape
x = x.flatten(2).permute(0, 2, 1)
query = F.normalize(self.to_query(x), dim=-1)
key = F.normalize(self.to_key(x), dim=-1)
attn = F.normalize(query @ self.w_g * self.scale_factor, dim=1)
attn = torch.sum(attn * query, dim=1, keepdim=True)
out = self.proj(attn * key) + query
out = self.final(out).permute(0, 2, 1).reshape(B, -1, H, W)
return out
class LocalRepresentation(nn.Module):
"""
Local Representation module for SwiftFormer that is implemented by 3*3 depth-wise and point-wise convolutions.
Input: tensor in shape [B, C, H, W]
Output: tensor in shape [B, C, H, W]
"""
def __init__(
self,
dim: int,
kernel_size: int = 3,
drop_path: float = 0.,
use_layer_scale: bool = True,
act_layer: LayerType = nn.GELU,
norm_layer: LayerType = nn.BatchNorm2d,
):
super().__init__()
self.dwconv = nn.Conv2d(dim, dim, kernel_size, padding=kernel_size // 2, groups=dim)
self.norm = norm_layer(dim)
self.pwconv1 = nn.Conv2d(dim, dim, kernel_size=1)
self.act = act_layer()
self.pwconv2 = nn.Conv2d(dim, dim, kernel_size=1)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.layer_scale = LayerScale2d(dim, 1) if use_layer_scale else nn.Identity()
def forward(self, x: torch.Tensor) -> torch.Tensor:
skip = x
x = self.dwconv(x)
x = self.norm(x)
x = self.pwconv1(x)
x = self.act(x)
x = self.pwconv2(x)
x = self.layer_scale(x)
x = skip + self.drop_path(x)
return x
class Block(nn.Module):
"""
SwiftFormer Encoder Block for SwiftFormer. It consists of :
(1) Local representation module, (2) EfficientAdditiveAttention, and (3) MLP block.
Input: tensor in shape [B, C, H, W]
Output: tensor in shape [B, C, H, W]
"""
def __init__(
self,
dim: int,
mlp_ratio: float = 4.,
drop_rate: float = 0.,
drop_path: float = 0.,
act_layer: LayerType = nn.GELU,
norm_layer: LayerType = nn.BatchNorm2d,
use_layer_scale: bool = True,
layer_scale_init_value: float = 1e-5,
):
super().__init__()
self.local_representation = LocalRepresentation(
dim=dim,
use_layer_scale=use_layer_scale,
act_layer=act_layer,
norm_layer=norm_layer,
)
self.attn = EfficientAdditiveAttention(in_dims=dim, token_dim=dim)
self.linear = Mlp(
in_features=dim,
hidden_features=int(dim * mlp_ratio),
act_layer=act_layer,
norm_layer=norm_layer,
drop=drop_rate,
)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.layer_scale_1 = LayerScale2d(dim, layer_scale_init_value) \
if use_layer_scale else nn.Identity()
self.layer_scale_2 = LayerScale2d(dim, layer_scale_init_value) \
if use_layer_scale else nn.Identity()
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.local_representation(x)
x = x + self.drop_path(self.layer_scale_1(self.attn(x)))
x = x + self.drop_path(self.layer_scale_2(self.linear(x)))
return x
class Stage(nn.Module):
"""
Implementation of each SwiftFormer stages. Here, SwiftFormerEncoder used as the last block in all stages, while ConvEncoder used in the rest of the blocks.
Input: tensor in shape [B, C, H, W]
Output: tensor in shape [B, C, H, W]
"""
def __init__(
self,
dim: int,
index: int,
layers: List[int],
mlp_ratio: float = 4.,
act_layer: LayerType = nn.GELU,
norm_layer: LayerType = nn.BatchNorm2d,
drop_rate: float = 0.,
drop_path_rate: float = 0.,
use_layer_scale: bool = True,
layer_scale_init_value: float = 1e-5,
downsample: Optional[LayerType] = None,
):
super().__init__()
self.grad_checkpointing = False
self.downsample = downsample if downsample is not None else nn.Identity()
blocks = []
for block_idx in range(layers[index]):
block_dpr = drop_path_rate * (block_idx + sum(layers[:index])) / (sum(layers) - 1)
if layers[index] - block_idx <= 1:
blocks.append(Block(
dim,
mlp_ratio=mlp_ratio,
drop_rate=drop_rate,
drop_path=block_dpr,
act_layer=act_layer,
norm_layer=norm_layer,
use_layer_scale=use_layer_scale,
layer_scale_init_value=layer_scale_init_value,
))
else:
blocks.append(ConvEncoder(
dim=dim,
hidden_dim=int(mlp_ratio * dim),
kernel_size=3,
drop_path=block_dpr,
act_layer=act_layer,
norm_layer=norm_layer,
use_layer_scale=use_layer_scale,
))
self.blocks = nn.Sequential(*blocks)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.downsample(x)
if self.grad_checkpointing and not torch.jit.is_scripting():
x = checkpoint_seq(self.blocks, x, flatten=True)
else:
x = self.blocks(x)
return x
class SwiftFormer(nn.Module):
def __init__(
self,
layers: List[int] = [3, 3, 6, 4],
embed_dims: List[int] = [48, 56, 112, 220],
mlp_ratios: int = 4,
downsamples: List[bool] = [False, True, True, True],
act_layer: LayerType = nn.GELU,
down_patch_size: int = 3,
down_stride: int = 2,
down_pad: int = 1,
num_classes: int = 1000,
drop_rate: float = 0.,
drop_path_rate: float = 0.,
use_layer_scale: bool = True,
layer_scale_init_value: float = 1e-5,
global_pool: str = 'avg',
output_stride: int = 32,
in_chans: int = 3,
**kwargs,
):
super().__init__()
assert output_stride == 32
self.num_classes = num_classes
self.global_pool = global_pool
self.feature_info = []
self.stem = nn.Sequential(
nn.Conv2d(in_chans, embed_dims[0] // 2, 3, 2, 1),
nn.BatchNorm2d(embed_dims[0] // 2),
nn.ReLU(),
nn.Conv2d(embed_dims[0] // 2, embed_dims[0], 3, 2, 1),
nn.BatchNorm2d(embed_dims[0]),
nn.ReLU(),
)
prev_dim = embed_dims[0]
stages = []
for i in range(len(layers)):
downsample = Embedding(
in_chans=prev_dim,
embed_dim=embed_dims[i],
patch_size=down_patch_size,
stride=down_stride,
padding=down_pad,
) if downsamples[i] else nn.Identity()
stage = Stage(
dim=embed_dims[i],
index=i,
layers=layers,
mlp_ratio=mlp_ratios,
act_layer=act_layer,
drop_rate=drop_rate,
drop_path_rate=drop_path_rate,
use_layer_scale=use_layer_scale,
layer_scale_init_value=layer_scale_init_value,
downsample=downsample,
)
prev_dim = embed_dims[i]
stages.append(stage)
self.feature_info += [dict(num_chs=embed_dims[i], reduction=2**(i+2), module=f'stages.{i}')]
self.stages = nn.Sequential(*stages)
# Classifier head
self.num_features = self.head_hidden_size = out_chs = embed_dims[-1]
self.norm = nn.BatchNorm2d(out_chs)
self.head_drop = nn.Dropout(drop_rate)
self.head = Linear(out_chs, num_classes) if num_classes > 0 else nn.Identity()
# assuming model is always distilled (valid for current checkpoints, will split def if that changes)
self.head_dist = Linear(out_chs, num_classes) if num_classes > 0 else nn.Identity()
self.distilled_training = False # must set this True to train w/ distillation token
self._initialize_weights()
def _initialize_weights(self):
for name, m in self.named_modules():
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Conv2d):
trunc_normal_(m.weight, std=.02)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
@torch.jit.ignore
def no_weight_decay(self) -> Set:
return set()
@torch.jit.ignore
def group_matcher(self, coarse: bool = False) -> Dict[str, Any]:
matcher = dict(
stem=r'^stem', # stem and embed
blocks=[(r'^stages\.(\d+)', None), (r'^norm', (99999,))]
)
return matcher
@torch.jit.ignore
def set_grad_checkpointing(self, enable: bool = True):
for s in self.stages:
s.grad_checkpointing = enable
@torch.jit.ignore
def get_classifier(self) -> Tuple[nn.Module, nn.Module]:
return self.head, self.head_dist
def reset_classifier(self, num_classes: int, global_pool: Optional[str] = None):
self.num_classes = num_classes
if global_pool is not None:
self.global_pool = global_pool
self.head = Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
self.head_dist = Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
@torch.jit.ignore
def set_distilled_training(self, enable: bool = True):
self.distilled_training = enable
def forward_intermediates(
self,
x: torch.Tensor,
indices: Optional[Union[int, List[int]]] = None,
norm: bool = False,
stop_early: bool = False,
output_fmt: str = 'NCHW',
intermediates_only: bool = False,
) -> Union[List[torch.Tensor], Tuple[torch.Tensor, List[torch.Tensor]]]:
""" Forward features that returns intermediates.
Args:
x: Input image tensor
indices: Take last n blocks if int, all if None, select matching indices if sequence
norm: Apply norm layer to compatible intermediates
stop_early: Stop iterating over blocks when last desired intermediate hit
output_fmt: Shape of intermediate feature outputs
intermediates_only: Only return intermediate features
Returns:
"""
assert output_fmt in ('NCHW',), 'Output shape must be NCHW.'
intermediates = []
take_indices, max_index = feature_take_indices(len(self.stages), indices)
last_idx = len(self.stages) - 1
# forward pass
x = self.stem(x)
if torch.jit.is_scripting() or not stop_early: # can't slice blocks in torchscript
stages = self.stages
else:
stages = self.stages[:max_index + 1]
for feat_idx, stage in enumerate(stages):
x = stage(x)
if feat_idx in take_indices:
if norm and feat_idx == last_idx:
x_inter = self.norm(x) # applying final norm last intermediate
else:
x_inter = x
intermediates.append(x_inter)
if intermediates_only:
return intermediates
if feat_idx == last_idx:
x = self.norm(x)
return x, intermediates
def prune_intermediate_layers(
self,
indices: Union[int, List[int]] = 1,
prune_norm: bool = False,
prune_head: bool = True,
):
""" Prune layers not required for specified intermediates.
"""
take_indices, max_index = feature_take_indices(len(self.stages), indices)
self.stages = self.stages[:max_index + 1] # truncate blocks w/ stem as idx 0
if prune_norm:
self.norm = nn.Identity()
if prune_head:
self.reset_classifier(0, '')
return take_indices
def forward_features(self, x: torch.Tensor) -> torch.Tensor:
x = self.stem(x)
x = self.stages(x)
x = self.norm(x)
return x
def forward_head(self, x: torch.Tensor, pre_logits: bool = False):
if self.global_pool == 'avg':
x = x.mean(dim=(2, 3))
x = self.head_drop(x)
if pre_logits:
return x
x, x_dist = self.head(x), self.head_dist(x)
if self.distilled_training and self.training and not torch.jit.is_scripting():
# only return separate classification predictions when training in distilled mode
return x, x_dist
else:
# during standard train/finetune, inference average the classifier predictions
return (x + x_dist) / 2
def forward(self, x: torch.Tensor):
x = self.forward_features(x)
x = self.forward_head(x)
return x
def checkpoint_filter_fn(state_dict: Dict[str, torch.Tensor], model: nn.Module) -> Dict[str, torch.Tensor]:
if 'model' in state_dict:
state_dict = state_dict['model']
out_dict = {}
for k, v in state_dict.items():
k = k.replace('patch_embed.', 'stem.')
k = k.replace('dist_head.', 'head_dist.')
k = k.replace('attn.Proj.', 'attn.proj.')
k = k.replace('.layer_scale_1', '.layer_scale_1.gamma')
k = k.replace('.layer_scale_2', '.layer_scale_2.gamma')
k = re.sub(r'\.layer_scale(?=$|\.)', '.layer_scale.gamma', k)
m = re.match(r'^network\.(\d+)\.(.*)', k)
if m:
n_idx, rest = int(m.group(1)), m.group(2)
stage_idx = n_idx // 2
if n_idx % 2 == 0:
k = f'stages.{stage_idx}.blocks.{rest}'
else:
k = f'stages.{stage_idx+1}.downsample.{rest}'
out_dict[k] = v
return out_dict
def _cfg(url: str = '', **kwargs: Any) -> Dict[str, Any]:
return {
'url': url,
'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None, 'fixed_input_size': True,
'crop_pct': .95, 'interpolation': 'bicubic',
'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
'first_conv': 'stem.0', 'classifier': ('head', 'head_dist'),
'paper_ids': 'arXiv:2303.15446',
'paper_name': 'SwiftFormer: Efficient Additive Attention for Transformer-based Real-time Mobile Vision Applications',
'origin_url': 'https://github.com/Amshaker/SwiftFormer',
**kwargs
}
default_cfgs = generate_default_cfgs({
# 'swiftformer_xs.dist_in1k': _cfg(hf_hub_id='timm/'),
# 'swiftformer_s.dist_in1k': _cfg(hf_hub_id='timm/'),
# 'swiftformer_l1.dist_in1k': _cfg(hf_hub_id='timm/'),
# 'swiftformer_l3.dist_in1k': _cfg(hf_hub_id='timm/'),
'swiftformer_xs.untrained': _cfg(),
'swiftformer_s.untrained': _cfg(),
'swiftformer_l1.untrained': _cfg(),
'swiftformer_l3.untrained': _cfg(),
})
def _create_swiftformer(variant: str, pretrained: bool = False, **kwargs: Any) -> SwiftFormer:
model = build_model_with_cfg(
SwiftFormer, variant, pretrained,
pretrained_filter_fn=checkpoint_filter_fn,
feature_cfg=dict(out_indices=(0, 1, 2, 3), flatten_sequential=True),
**kwargs,
)
return model
@register_model
def swiftformer_xs(pretrained: bool = False, **kwargs: Any) -> SwiftFormer:
model_args = dict(layers=[3, 3, 6, 4], embed_dims=[48, 56, 112, 220])
return _create_swiftformer('swiftformer_xs', pretrained=pretrained, **dict(model_args, **kwargs))
@register_model
def swiftformer_s(pretrained: bool = False, **kwargs: Any) -> SwiftFormer:
model_args = dict(layers=[3, 3, 9, 6], embed_dims=[48, 64, 168, 224])
return _create_swiftformer('swiftformer_s', pretrained=pretrained, **dict(model_args, **kwargs))
@register_model
def swiftformer_l1(pretrained: bool = False, **kwargs: Any) -> SwiftFormer:
model_args = dict(layers=[4, 3, 10, 5], embed_dims=[48, 96, 192, 384])
return _create_swiftformer('swiftformer_l1', pretrained=pretrained, **dict(model_args, **kwargs))
@register_model
def swiftformer_l3(pretrained: bool = False, **kwargs: Any) -> SwiftFormer:
model_args = dict(layers=[4, 4, 12, 6], embed_dims=[64, 128, 320, 512])
return _create_swiftformer('swiftformer_l3', pretrained=pretrained, **dict(model_args, **kwargs))