Ross Wightman
2 years ago
commit
dff33730b3
@ -0,0 +1,417 @@
|
||||
""" Sequencer
|
||||
|
||||
Paper: `Sequencer: Deep LSTM for Image Classification` - https://arxiv.org/pdf/2205.01972.pdf
|
||||
|
||||
"""
|
||||
# Copyright (c) 2022. Yuki Tatsunami
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
|
||||
|
||||
import math
|
||||
from functools import partial
|
||||
from typing import Tuple
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD, DEFAULT_CROP_PCT
|
||||
from .helpers import build_model_with_cfg, named_apply
|
||||
from .layers import lecun_normal_, DropPath, Mlp, PatchEmbed as TimmPatchEmbed
|
||||
from .registry import register_model
|
||||
|
||||
|
||||
def _cfg(url='', **kwargs):
|
||||
return {
|
||||
'url': url,
|
||||
'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
|
||||
'crop_pct': DEFAULT_CROP_PCT, 'interpolation': 'bicubic', 'fixed_input_size': True,
|
||||
'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
|
||||
'first_conv': 'stem.proj', 'classifier': 'head',
|
||||
**kwargs
|
||||
}
|
||||
|
||||
|
||||
default_cfgs = dict(
|
||||
sequencer2d_s=_cfg(url="https://github.com/okojoalg/sequencer/releases/download/weights/sequencer2d_s.pth"),
|
||||
sequencer2d_m=_cfg(url="https://github.com/okojoalg/sequencer/releases/download/weights/sequencer2d_m.pth"),
|
||||
sequencer2d_l=_cfg(url="https://github.com/okojoalg/sequencer/releases/download/weights/sequencer2d_l.pth"),
|
||||
)
|
||||
|
||||
|
||||
def _init_weights(module: nn.Module, name: str, head_bias: float = 0., flax=False):
|
||||
if isinstance(module, nn.Linear):
|
||||
if name.startswith('head'):
|
||||
nn.init.zeros_(module.weight)
|
||||
nn.init.constant_(module.bias, head_bias)
|
||||
else:
|
||||
if flax:
|
||||
# Flax defaults
|
||||
lecun_normal_(module.weight)
|
||||
if module.bias is not None:
|
||||
nn.init.zeros_(module.bias)
|
||||
else:
|
||||
nn.init.xavier_uniform_(module.weight)
|
||||
if module.bias is not None:
|
||||
if 'mlp' in name:
|
||||
nn.init.normal_(module.bias, std=1e-6)
|
||||
else:
|
||||
nn.init.zeros_(module.bias)
|
||||
elif isinstance(module, nn.Conv2d):
|
||||
lecun_normal_(module.weight)
|
||||
if module.bias is not None:
|
||||
nn.init.zeros_(module.bias)
|
||||
elif isinstance(module, (nn.LayerNorm, nn.BatchNorm2d, nn.GroupNorm)):
|
||||
nn.init.ones_(module.weight)
|
||||
nn.init.zeros_(module.bias)
|
||||
elif isinstance(module, (nn.RNN, nn.GRU, nn.LSTM)):
|
||||
stdv = 1.0 / math.sqrt(module.hidden_size)
|
||||
for weight in module.parameters():
|
||||
nn.init.uniform_(weight, -stdv, stdv)
|
||||
elif hasattr(module, 'init_weights'):
|
||||
module.init_weights()
|
||||
|
||||
|
||||
def get_stage(
|
||||
index, layers, patch_sizes, embed_dims, hidden_sizes, mlp_ratios, block_layer, rnn_layer, mlp_layer,
|
||||
norm_layer, act_layer, num_layers, bidirectional, union,
|
||||
with_fc, drop=0., drop_path_rate=0., **kwargs):
|
||||
assert len(layers) == len(patch_sizes) == len(embed_dims) == len(hidden_sizes) == len(mlp_ratios)
|
||||
blocks = []
|
||||
for block_idx in range(layers[index]):
|
||||
drop_path = drop_path_rate * (block_idx + sum(layers[:index])) / (sum(layers) - 1)
|
||||
blocks.append(block_layer(
|
||||
embed_dims[index], hidden_sizes[index], mlp_ratio=mlp_ratios[index],
|
||||
rnn_layer=rnn_layer, mlp_layer=mlp_layer, norm_layer=norm_layer, act_layer=act_layer,
|
||||
num_layers=num_layers, bidirectional=bidirectional, union=union, with_fc=with_fc,
|
||||
drop=drop, drop_path=drop_path))
|
||||
|
||||
if index < len(embed_dims) - 1:
|
||||
blocks.append(Downsample2D(embed_dims[index], embed_dims[index + 1], patch_sizes[index + 1]))
|
||||
|
||||
blocks = nn.Sequential(*blocks)
|
||||
return blocks
|
||||
|
||||
|
||||
class RNNIdentity(nn.Module):
|
||||
def __init__(self, *args, **kwargs):
|
||||
super(RNNIdentity, self).__init__()
|
||||
|
||||
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, None]:
|
||||
return x, None
|
||||
|
||||
|
||||
class RNN2DBase(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self, input_size: int, hidden_size: int,
|
||||
num_layers: int = 1, bias: bool = True, bidirectional: bool = True,
|
||||
union="cat", with_fc=True):
|
||||
super().__init__()
|
||||
|
||||
self.input_size = input_size
|
||||
self.hidden_size = hidden_size
|
||||
self.output_size = 2 * hidden_size if bidirectional else hidden_size
|
||||
self.union = union
|
||||
|
||||
self.with_vertical = True
|
||||
self.with_horizontal = True
|
||||
self.with_fc = with_fc
|
||||
|
||||
self.fc = None
|
||||
if with_fc:
|
||||
if union == "cat":
|
||||
self.fc = nn.Linear(2 * self.output_size, input_size)
|
||||
elif union == "add":
|
||||
self.fc = nn.Linear(self.output_size, input_size)
|
||||
elif union == "vertical":
|
||||
self.fc = nn.Linear(self.output_size, input_size)
|
||||
self.with_horizontal = False
|
||||
elif union == "horizontal":
|
||||
self.fc = nn.Linear(self.output_size, input_size)
|
||||
self.with_vertical = False
|
||||
else:
|
||||
raise ValueError("Unrecognized union: " + union)
|
||||
elif union == "cat":
|
||||
pass
|
||||
if 2 * self.output_size != input_size:
|
||||
raise ValueError(f"The output channel {2 * self.output_size} is different from the input channel {input_size}.")
|
||||
elif union == "add":
|
||||
pass
|
||||
if self.output_size != input_size:
|
||||
raise ValueError(f"The output channel {self.output_size} is different from the input channel {input_size}.")
|
||||
elif union == "vertical":
|
||||
if self.output_size != input_size:
|
||||
raise ValueError(f"The output channel {self.output_size} is different from the input channel {input_size}.")
|
||||
self.with_horizontal = False
|
||||
elif union == "horizontal":
|
||||
if self.output_size != input_size:
|
||||
raise ValueError(f"The output channel {self.output_size} is different from the input channel {input_size}.")
|
||||
self.with_vertical = False
|
||||
else:
|
||||
raise ValueError("Unrecognized union: " + union)
|
||||
|
||||
self.rnn_v = RNNIdentity()
|
||||
self.rnn_h = RNNIdentity()
|
||||
|
||||
def forward(self, x):
|
||||
B, H, W, C = x.shape
|
||||
|
||||
if self.with_vertical:
|
||||
v = x.permute(0, 2, 1, 3)
|
||||
v = v.reshape(-1, H, C)
|
||||
v, _ = self.rnn_v(v)
|
||||
v = v.reshape(B, W, H, -1)
|
||||
v = v.permute(0, 2, 1, 3)
|
||||
else:
|
||||
v = None
|
||||
|
||||
if self.with_horizontal:
|
||||
h = x.reshape(-1, W, C)
|
||||
h, _ = self.rnn_h(h)
|
||||
h = h.reshape(B, H, W, -1)
|
||||
else:
|
||||
h = None
|
||||
|
||||
if v is not None and h is not None:
|
||||
if self.union == "cat":
|
||||
x = torch.cat([v, h], dim=-1)
|
||||
else:
|
||||
x = v + h
|
||||
elif v is not None:
|
||||
x = v
|
||||
elif h is not None:
|
||||
x = h
|
||||
|
||||
if self.fc is not None:
|
||||
x = self.fc(x)
|
||||
|
||||
return x
|
||||
|
||||
|
||||
class LSTM2D(RNN2DBase):
|
||||
|
||||
def __init__(
|
||||
self, input_size: int, hidden_size: int,
|
||||
num_layers: int = 1, bias: bool = True, bidirectional: bool = True,
|
||||
union="cat", with_fc=True):
|
||||
super().__init__(input_size, hidden_size, num_layers, bias, bidirectional, union, with_fc)
|
||||
if self.with_vertical:
|
||||
self.rnn_v = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True, bias=bias, bidirectional=bidirectional)
|
||||
if self.with_horizontal:
|
||||
self.rnn_h = nn.LSTM(input_size, hidden_size, num_layers, batch_first=True, bias=bias, bidirectional=bidirectional)
|
||||
|
||||
|
||||
class Sequencer2DBlock(nn.Module):
|
||||
def __init__(
|
||||
self, dim, hidden_size, mlp_ratio=3.0, rnn_layer=LSTM2D, mlp_layer=Mlp,
|
||||
norm_layer=partial(nn.LayerNorm, eps=1e-6), act_layer=nn.GELU,
|
||||
num_layers=1, bidirectional=True, union="cat", with_fc=True, drop=0., drop_path=0.):
|
||||
super().__init__()
|
||||
channels_dim = int(mlp_ratio * dim)
|
||||
self.norm1 = norm_layer(dim)
|
||||
self.rnn_tokens = rnn_layer(dim, hidden_size, num_layers=num_layers, bidirectional=bidirectional,
|
||||
union=union, with_fc=with_fc)
|
||||
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
self.norm2 = norm_layer(dim)
|
||||
self.mlp_channels = mlp_layer(dim, channels_dim, act_layer=act_layer, drop=drop)
|
||||
|
||||
def forward(self, x):
|
||||
x = x + self.drop_path(self.rnn_tokens(self.norm1(x)))
|
||||
x = x + self.drop_path(self.mlp_channels(self.norm2(x)))
|
||||
return x
|
||||
|
||||
|
||||
class PatchEmbed(TimmPatchEmbed):
|
||||
def forward(self, x):
|
||||
x = self.proj(x)
|
||||
if self.flatten:
|
||||
x = x.flatten(2).transpose(1, 2) # BCHW -> BNC
|
||||
else:
|
||||
x = x.permute(0, 2, 3, 1) # BCHW -> BHWC
|
||||
x = self.norm(x)
|
||||
return x
|
||||
|
||||
|
||||
class Shuffle(nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
|
||||
def forward(self, x):
|
||||
if self.training:
|
||||
B, H, W, C = x.shape
|
||||
r = torch.randperm(H * W)
|
||||
x = x.reshape(B, -1, C)
|
||||
x = x[:, r, :].reshape(B, H, W, -1)
|
||||
return x
|
||||
|
||||
|
||||
class Downsample2D(nn.Module):
|
||||
def __init__(self, input_dim, output_dim, patch_size):
|
||||
super().__init__()
|
||||
self.down = nn.Conv2d(input_dim, output_dim, kernel_size=patch_size, stride=patch_size)
|
||||
|
||||
def forward(self, x):
|
||||
x = x.permute(0, 3, 1, 2)
|
||||
x = self.down(x)
|
||||
x = x.permute(0, 2, 3, 1)
|
||||
return x
|
||||
|
||||
|
||||
class Sequencer2D(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
num_classes=1000,
|
||||
img_size=224,
|
||||
in_chans=3,
|
||||
global_pool='avg',
|
||||
layers=[4, 3, 8, 3],
|
||||
patch_sizes=[7, 2, 1, 1],
|
||||
embed_dims=[192, 384, 384, 384],
|
||||
hidden_sizes=[48, 96, 96, 96],
|
||||
mlp_ratios=[3.0, 3.0, 3.0, 3.0],
|
||||
block_layer=Sequencer2DBlock,
|
||||
rnn_layer=LSTM2D,
|
||||
mlp_layer=Mlp,
|
||||
norm_layer=partial(nn.LayerNorm, eps=1e-6),
|
||||
act_layer=nn.GELU,
|
||||
num_rnn_layers=1,
|
||||
bidirectional=True,
|
||||
union="cat",
|
||||
with_fc=True,
|
||||
drop_rate=0.,
|
||||
drop_path_rate=0.,
|
||||
nlhb=False,
|
||||
stem_norm=False,
|
||||
):
|
||||
super().__init__()
|
||||
assert global_pool in ('', 'avg')
|
||||
self.num_classes = num_classes
|
||||
self.global_pool = global_pool
|
||||
self.num_features = embed_dims[-1] # num_features for consistency with other models
|
||||
self.feature_dim = -1 # channel dim index for feature outputs (rank 4, NHWC)
|
||||
self.embed_dims = embed_dims
|
||||
self.stem = PatchEmbed(
|
||||
img_size=img_size, patch_size=patch_sizes[0], in_chans=in_chans,
|
||||
embed_dim=embed_dims[0], norm_layer=norm_layer if stem_norm else None,
|
||||
flatten=False)
|
||||
|
||||
self.blocks = nn.Sequential(*[
|
||||
get_stage(
|
||||
i, layers, patch_sizes, embed_dims, hidden_sizes, mlp_ratios, block_layer=block_layer,
|
||||
rnn_layer=rnn_layer, mlp_layer=mlp_layer, norm_layer=norm_layer, act_layer=act_layer,
|
||||
num_layers=num_rnn_layers, bidirectional=bidirectional,
|
||||
union=union, with_fc=with_fc, drop=drop_rate, drop_path_rate=drop_path_rate,
|
||||
)
|
||||
for i, _ in enumerate(embed_dims)])
|
||||
|
||||
self.norm = norm_layer(embed_dims[-1])
|
||||
self.head = nn.Linear(embed_dims[-1], self.num_classes) if num_classes > 0 else nn.Identity()
|
||||
|
||||
self.init_weights(nlhb=nlhb)
|
||||
|
||||
def init_weights(self, nlhb=False):
|
||||
head_bias = -math.log(self.num_classes) if nlhb else 0.
|
||||
named_apply(partial(_init_weights, head_bias=head_bias), module=self) # depth-first
|
||||
|
||||
@torch.jit.ignore
|
||||
def group_matcher(self, coarse=False):
|
||||
return dict(
|
||||
stem=r'^stem',
|
||||
blocks=[
|
||||
(r'^blocks\.(\d+)\..*\.down', (99999,)),
|
||||
(r'^blocks\.(\d+)', None) if coarse else (r'^blocks\.(\d+)\.(\d+)', None),
|
||||
(r'^norm', (99999,))
|
||||
]
|
||||
)
|
||||
|
||||
@torch.jit.ignore
|
||||
def set_grad_checkpointing(self, enable=True):
|
||||
assert not enable, 'gradient checkpointing not supported'
|
||||
|
||||
@torch.jit.ignore
|
||||
def get_classifier(self):
|
||||
return self.head
|
||||
|
||||
def reset_classifier(self, num_classes, global_pool=None):
|
||||
self.num_classes = num_classes
|
||||
if global_pool is not None:
|
||||
assert global_pool in ('', 'avg')
|
||||
self.global_pool = global_pool
|
||||
self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
||||
|
||||
def forward_features(self, x):
|
||||
x = self.stem(x)
|
||||
x = self.blocks(x)
|
||||
x = self.norm(x)
|
||||
return x
|
||||
|
||||
def forward_head(self, x, pre_logits: bool = False):
|
||||
if self.global_pool == 'avg':
|
||||
x = x.mean(dim=(1, 2))
|
||||
return x if pre_logits else self.head(x)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.forward_features(x)
|
||||
x = self.forward_head(x)
|
||||
return x
|
||||
|
||||
|
||||
def _create_sequencer2d(variant, pretrained=False, **kwargs):
|
||||
if kwargs.get('features_only', None):
|
||||
raise RuntimeError('features_only not implemented for Sequencer2D models.')
|
||||
|
||||
model = build_model_with_cfg(Sequencer2D, variant, pretrained, **kwargs)
|
||||
return model
|
||||
|
||||
|
||||
# main
|
||||
|
||||
@register_model
|
||||
def sequencer2d_s(pretrained=False, **kwargs):
|
||||
model_args = dict(
|
||||
layers=[4, 3, 8, 3],
|
||||
patch_sizes=[7, 2, 1, 1],
|
||||
embed_dims=[192, 384, 384, 384],
|
||||
hidden_sizes=[48, 96, 96, 96],
|
||||
mlp_ratios=[3.0, 3.0, 3.0, 3.0],
|
||||
rnn_layer=LSTM2D,
|
||||
bidirectional=True,
|
||||
union="cat",
|
||||
with_fc=True,
|
||||
**kwargs)
|
||||
model = _create_sequencer2d('sequencer2d_s', pretrained=pretrained, **model_args)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def sequencer2d_m(pretrained=False, **kwargs):
|
||||
model_args = dict(
|
||||
layers=[4, 3, 14, 3],
|
||||
patch_sizes=[7, 2, 1, 1],
|
||||
embed_dims=[192, 384, 384, 384],
|
||||
hidden_sizes=[48, 96, 96, 96],
|
||||
mlp_ratios=[3.0, 3.0, 3.0, 3.0],
|
||||
rnn_layer=LSTM2D,
|
||||
bidirectional=True,
|
||||
union="cat",
|
||||
with_fc=True,
|
||||
**kwargs)
|
||||
model = _create_sequencer2d('sequencer2d_m', pretrained=pretrained, **model_args)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def sequencer2d_l(pretrained=False, **kwargs):
|
||||
model_args = dict(
|
||||
layers=[8, 8, 16, 4],
|
||||
patch_sizes=[7, 2, 1, 1],
|
||||
embed_dims=[192, 384, 384, 384],
|
||||
hidden_sizes=[48, 96, 96, 96],
|
||||
mlp_ratios=[3.0, 3.0, 3.0, 3.0],
|
||||
rnn_layer=LSTM2D,
|
||||
bidirectional=True,
|
||||
union="cat",
|
||||
with_fc=True,
|
||||
**kwargs)
|
||||
model = _create_sequencer2d('sequencer2d_l', pretrained=pretrained, **model_args)
|
||||
return model
|
||||
@ -0,0 +1,753 @@
|
||||
""" Swin Transformer V2
|
||||
A PyTorch impl of : `Swin Transformer V2: Scaling Up Capacity and Resolution`
|
||||
- https://arxiv.org/abs/2111.09883
|
||||
|
||||
Code/weights from https://github.com/microsoft/Swin-Transformer, original copyright/license info below
|
||||
|
||||
Modifications and additions for timm hacked together by / Copyright 2022, Ross Wightman
|
||||
"""
|
||||
# --------------------------------------------------------
|
||||
# Swin Transformer V2
|
||||
# Copyright (c) 2022 Microsoft
|
||||
# Licensed under The MIT License [see LICENSE for details]
|
||||
# Written by Ze Liu
|
||||
# --------------------------------------------------------
|
||||
import math
|
||||
from typing import Tuple, Optional
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
import torch.utils.checkpoint as checkpoint
|
||||
|
||||
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
|
||||
from .fx_features import register_notrace_function
|
||||
from .helpers import build_model_with_cfg, named_apply
|
||||
from .layers import PatchEmbed, Mlp, DropPath, to_2tuple, to_ntuple, trunc_normal_, _assert
|
||||
from .registry import register_model
|
||||
|
||||
|
||||
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 = {
|
||||
'swinv2_tiny_window8_256': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_tiny_patch4_window8_256.pth',
|
||||
input_size=(3, 256, 256)
|
||||
),
|
||||
'swinv2_tiny_window16_256': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_tiny_patch4_window16_256.pth',
|
||||
input_size=(3, 256, 256)
|
||||
),
|
||||
'swinv2_small_window8_256': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_small_patch4_window8_256.pth',
|
||||
input_size=(3, 256, 256)
|
||||
),
|
||||
'swinv2_small_window16_256': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_small_patch4_window16_256.pth',
|
||||
input_size=(3, 256, 256)
|
||||
),
|
||||
'swinv2_base_window8_256': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window8_256.pth',
|
||||
input_size=(3, 256, 256)
|
||||
),
|
||||
'swinv2_base_window16_256': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window16_256.pth',
|
||||
input_size=(3, 256, 256)
|
||||
),
|
||||
|
||||
'swinv2_base_window12_192_22k': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window12_192_22k.pth',
|
||||
num_classes=21841, input_size=(3, 192, 192)
|
||||
),
|
||||
'swinv2_base_window12to16_192to256_22kft1k': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window12to16_192to256_22kto1k_ft.pth',
|
||||
input_size=(3, 256, 256)
|
||||
),
|
||||
'swinv2_base_window12to24_192to384_22kft1k': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window12to24_192to384_22kto1k_ft.pth',
|
||||
input_size=(3, 384, 384), crop_pct=1.0,
|
||||
),
|
||||
'swinv2_large_window12_192_22k': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_large_patch4_window12_192_22k.pth',
|
||||
num_classes=21841, input_size=(3, 192, 192)
|
||||
),
|
||||
'swinv2_large_window12to16_192to256_22kft1k': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_large_patch4_window12to16_192to256_22kto1k_ft.pth',
|
||||
input_size=(3, 256, 256)
|
||||
),
|
||||
'swinv2_large_window12to24_192to384_22kft1k': _cfg(
|
||||
url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_large_patch4_window12to24_192to384_22kto1k_ft.pth',
|
||||
input_size=(3, 384, 384), crop_pct=1.0,
|
||||
),
|
||||
}
|
||||
|
||||
|
||||
def window_partition(x, window_size: Tuple[int, int]):
|
||||
"""
|
||||
Args:
|
||||
x: (B, H, W, C)
|
||||
window_size (int): window size
|
||||
|
||||
Returns:
|
||||
windows: (num_windows*B, window_size, window_size, C)
|
||||
"""
|
||||
B, H, W, C = x.shape
|
||||
x = x.view(B, H // window_size[0], window_size[0], W // window_size[1], window_size[1], C)
|
||||
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size[0], window_size[1], C)
|
||||
return windows
|
||||
|
||||
|
||||
@register_notrace_function # reason: int argument is a Proxy
|
||||
def window_reverse(windows, window_size: Tuple[int, int], img_size: Tuple[int, int]):
|
||||
"""
|
||||
Args:
|
||||
windows: (num_windows * B, window_size[0], window_size[1], C)
|
||||
window_size (Tuple[int, int]): Window size
|
||||
img_size (Tuple[int, int]): Image size
|
||||
|
||||
Returns:
|
||||
x: (B, H, W, C)
|
||||
"""
|
||||
H, W = img_size
|
||||
B = int(windows.shape[0] / (H * W / window_size[0] / window_size[1]))
|
||||
x = windows.view(B, H // window_size[0], W // window_size[1], window_size[0], window_size[1], -1)
|
||||
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
|
||||
return x
|
||||
|
||||
|
||||
class WindowAttention(nn.Module):
|
||||
r""" Window based multi-head self attention (W-MSA) module with relative position bias.
|
||||
It supports both of shifted and non-shifted window.
|
||||
|
||||
Args:
|
||||
dim (int): Number of input channels.
|
||||
window_size (tuple[int]): The height and width of the window.
|
||||
num_heads (int): Number of attention heads.
|
||||
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
|
||||
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
|
||||
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
|
||||
pretrained_window_size (tuple[int]): The height and width of the window in pre-training.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self, dim, window_size, num_heads, qkv_bias=True, attn_drop=0., proj_drop=0.,
|
||||
pretrained_window_size=[0, 0]):
|
||||
|
||||
super().__init__()
|
||||
self.dim = dim
|
||||
self.window_size = window_size # Wh, Ww
|
||||
self.pretrained_window_size = pretrained_window_size
|
||||
self.num_heads = num_heads
|
||||
|
||||
self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))))
|
||||
|
||||
# mlp to generate continuous relative position bias
|
||||
self.cpb_mlp = nn.Sequential(
|
||||
nn.Linear(2, 512, bias=True),
|
||||
nn.ReLU(inplace=True),
|
||||
nn.Linear(512, num_heads, bias=False)
|
||||
)
|
||||
|
||||
# get relative_coords_table
|
||||
relative_coords_h = torch.arange(-(self.window_size[0] - 1), self.window_size[0], dtype=torch.float32)
|
||||
relative_coords_w = torch.arange(-(self.window_size[1] - 1), self.window_size[1], dtype=torch.float32)
|
||||
relative_coords_table = torch.stack(torch.meshgrid([
|
||||
relative_coords_h,
|
||||
relative_coords_w])).permute(1, 2, 0).contiguous().unsqueeze(0) # 1, 2*Wh-1, 2*Ww-1, 2
|
||||
if pretrained_window_size[0] > 0:
|
||||
relative_coords_table[:, :, :, 0] /= (pretrained_window_size[0] - 1)
|
||||
relative_coords_table[:, :, :, 1] /= (pretrained_window_size[1] - 1)
|
||||
else:
|
||||
relative_coords_table[:, :, :, 0] /= (self.window_size[0] - 1)
|
||||
relative_coords_table[:, :, :, 1] /= (self.window_size[1] - 1)
|
||||
relative_coords_table *= 8 # normalize to -8, 8
|
||||
relative_coords_table = torch.sign(relative_coords_table) * torch.log2(
|
||||
torch.abs(relative_coords_table) + 1.0) / math.log2(8)
|
||||
|
||||
self.register_buffer("relative_coords_table", relative_coords_table, persistent=False)
|
||||
|
||||
# get pair-wise relative position index for each token inside the window
|
||||
coords_h = torch.arange(self.window_size[0])
|
||||
coords_w = torch.arange(self.window_size[1])
|
||||
coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
|
||||
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
|
||||
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
|
||||
relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
|
||||
relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
|
||||
relative_coords[:, :, 1] += self.window_size[1] - 1
|
||||
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
|
||||
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
|
||||
self.register_buffer("relative_position_index", relative_position_index, persistent=False)
|
||||
|
||||
self.qkv = nn.Linear(dim, dim * 3, bias=False)
|
||||
if qkv_bias:
|
||||
self.q_bias = nn.Parameter(torch.zeros(dim))
|
||||
self.register_buffer('k_bias', torch.zeros(dim), persistent=False)
|
||||
self.v_bias = nn.Parameter(torch.zeros(dim))
|
||||
else:
|
||||
self.q_bias = None
|
||||
self.k_bias = None
|
||||
self.v_bias = None
|
||||
self.attn_drop = nn.Dropout(attn_drop)
|
||||
self.proj = nn.Linear(dim, dim)
|
||||
self.proj_drop = nn.Dropout(proj_drop)
|
||||
self.softmax = nn.Softmax(dim=-1)
|
||||
|
||||
def forward(self, x, mask: Optional[torch.Tensor] = None):
|
||||
"""
|
||||
Args:
|
||||
x: input features with shape of (num_windows*B, N, C)
|
||||
mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
|
||||
"""
|
||||
B_, N, C = x.shape
|
||||
qkv_bias = None
|
||||
if self.q_bias is not None:
|
||||
qkv_bias = torch.cat((self.q_bias, self.k_bias, self.v_bias))
|
||||
qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
|
||||
qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
|
||||
q, k, v = qkv.unbind(0)
|
||||
|
||||
# cosine attention
|
||||
attn = (F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1))
|
||||
logit_scale = torch.clamp(self.logit_scale, max=math.log(1. / 0.01)).exp()
|
||||
attn = attn * logit_scale
|
||||
|
||||
relative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view(-1, self.num_heads)
|
||||
relative_position_bias = relative_position_bias_table[self.relative_position_index.view(-1)].view(
|
||||
self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH
|
||||
relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
|
||||
relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
|
||||
attn = attn + relative_position_bias.unsqueeze(0)
|
||||
|
||||
if mask is not None:
|
||||
nW = mask.shape[0]
|
||||
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
|
||||
attn = attn.view(-1, self.num_heads, N, N)
|
||||
attn = self.softmax(attn)
|
||||
else:
|
||||
attn = self.softmax(attn)
|
||||
|
||||
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 SwinTransformerBlock(nn.Module):
|
||||
r""" Swin Transformer Block.
|
||||
|
||||
Args:
|
||||
dim (int): Number of input channels.
|
||||
input_resolution (tuple[int]): Input resolution.
|
||||
num_heads (int): Number of attention heads.
|
||||
window_size (int): Window size.
|
||||
shift_size (int): Shift size for SW-MSA.
|
||||
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
|
||||
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
|
||||
drop (float, optional): Dropout rate. Default: 0.0
|
||||
attn_drop (float, optional): Attention dropout rate. Default: 0.0
|
||||
drop_path (float, optional): Stochastic depth rate. Default: 0.0
|
||||
act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
|
||||
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
|
||||
pretrained_window_size (int): Window size in pretraining.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
|
||||
mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,
|
||||
act_layer=nn.GELU, norm_layer=nn.LayerNorm, pretrained_window_size=0):
|
||||
super().__init__()
|
||||
self.dim = dim
|
||||
self.input_resolution = to_2tuple(input_resolution)
|
||||
self.num_heads = num_heads
|
||||
ws, ss = self._calc_window_shift(window_size, shift_size)
|
||||
self.window_size: Tuple[int, int] = ws
|
||||
self.shift_size: Tuple[int, int] = ss
|
||||
self.window_area = self.window_size[0] * self.window_size[1]
|
||||
self.mlp_ratio = mlp_ratio
|
||||
|
||||
self.attn = WindowAttention(
|
||||
dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
|
||||
qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,
|
||||
pretrained_window_size=to_2tuple(pretrained_window_size))
|
||||
self.norm1 = norm_layer(dim)
|
||||
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
|
||||
self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)
|
||||
self.norm2 = norm_layer(dim)
|
||||
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
|
||||
if any(self.shift_size):
|
||||
# calculate attention mask for SW-MSA
|
||||
H, W = self.input_resolution
|
||||
img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1
|
||||
cnt = 0
|
||||
for h in (
|
||||
slice(0, -self.window_size[0]),
|
||||
slice(-self.window_size[0], -self.shift_size[0]),
|
||||
slice(-self.shift_size[0], None)):
|
||||
for w in (
|
||||
slice(0, -self.window_size[1]),
|
||||
slice(-self.window_size[1], -self.shift_size[1]),
|
||||
slice(-self.shift_size[1], None)):
|
||||
img_mask[:, h, w, :] = cnt
|
||||
cnt += 1
|
||||
mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
|
||||
mask_windows = mask_windows.view(-1, self.window_area)
|
||||
attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
|
||||
attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
|
||||
else:
|
||||
attn_mask = None
|
||||
|
||||
self.register_buffer("attn_mask", attn_mask)
|
||||
|
||||
def _calc_window_shift(self, target_window_size, target_shift_size) -> Tuple[Tuple[int, int], Tuple[int, int]]:
|
||||
target_window_size = to_2tuple(target_window_size)
|
||||
target_shift_size = to_2tuple(target_shift_size)
|
||||
window_size = [r if r <= w else w for r, w in zip(self.input_resolution, target_window_size)]
|
||||
shift_size = [0 if r <= w else s for r, w, s in zip(self.input_resolution, window_size, target_shift_size)]
|
||||
return tuple(window_size), tuple(shift_size)
|
||||
|
||||
def _attn(self, x):
|
||||
H, W = self.input_resolution
|
||||
B, L, C = x.shape
|
||||
_assert(L == H * W, "input feature has wrong size")
|
||||
x = x.view(B, H, W, C)
|
||||
|
||||
# cyclic shift
|
||||
has_shift = any(self.shift_size)
|
||||
if has_shift:
|
||||
shifted_x = torch.roll(x, shifts=(-self.shift_size[0], -self.shift_size[1]), dims=(1, 2))
|
||||
else:
|
||||
shifted_x = x
|
||||
|
||||
# partition windows
|
||||
x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
|
||||
x_windows = x_windows.view(-1, self.window_area, C) # nW*B, window_size*window_size, C
|
||||
|
||||
# W-MSA/SW-MSA
|
||||
attn_windows = self.attn(x_windows, mask=self.attn_mask) # nW*B, window_size*window_size, C
|
||||
|
||||
# merge windows
|
||||
attn_windows = attn_windows.view(-1, self.window_size[0], self.window_size[1], C)
|
||||
shifted_x = window_reverse(attn_windows, self.window_size, self.input_resolution) # B H' W' C
|
||||
|
||||
# reverse cyclic shift
|
||||
if has_shift:
|
||||
x = torch.roll(shifted_x, shifts=self.shift_size, dims=(1, 2))
|
||||
else:
|
||||
x = shifted_x
|
||||
x = x.view(B, H * W, C)
|
||||
return x
|
||||
|
||||
def forward(self, x):
|
||||
x = x + self.drop_path1(self.norm1(self._attn(x)))
|
||||
x = x + self.drop_path2(self.norm2(self.mlp(x)))
|
||||
return x
|
||||
|
||||
|
||||
class PatchMerging(nn.Module):
|
||||
r""" Patch Merging Layer.
|
||||
|
||||
Args:
|
||||
input_resolution (tuple[int]): Resolution of input feature.
|
||||
dim (int): Number of input channels.
|
||||
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
|
||||
"""
|
||||
|
||||
def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
|
||||
super().__init__()
|
||||
self.input_resolution = input_resolution
|
||||
self.dim = dim
|
||||
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
|
||||
self.norm = norm_layer(2 * dim)
|
||||
|
||||
def forward(self, x):
|
||||
"""
|
||||
x: B, H*W, C
|
||||
"""
|
||||
H, W = self.input_resolution
|
||||
B, L, C = x.shape
|
||||
_assert(L == H * W, "input feature has wrong size")
|
||||
_assert(H % 2 == 0, f"x size ({H}*{W}) are not even.")
|
||||
_assert(W % 2 == 0, f"x size ({H}*{W}) are not even.")
|
||||
|
||||
x = x.view(B, H, W, C)
|
||||
|
||||
x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
|
||||
x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
|
||||
x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
|
||||
x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
|
||||
x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
|
||||
x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
|
||||
|
||||
x = self.reduction(x)
|
||||
x = self.norm(x)
|
||||
|
||||
return x
|
||||
|
||||
|
||||
class BasicLayer(nn.Module):
|
||||
""" A basic Swin Transformer layer for one stage.
|
||||
|
||||
Args:
|
||||
dim (int): Number of input channels.
|
||||
input_resolution (tuple[int]): Input resolution.
|
||||
depth (int): Number of blocks.
|
||||
num_heads (int): Number of attention heads.
|
||||
window_size (int): Local window size.
|
||||
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
|
||||
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
|
||||
drop (float, optional): Dropout rate. Default: 0.0
|
||||
attn_drop (float, optional): Attention dropout rate. Default: 0.0
|
||||
drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
|
||||
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
|
||||
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
|
||||
pretrained_window_size (int): Local window size in pre-training.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self, dim, input_resolution, depth, num_heads, window_size,
|
||||
mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,
|
||||
norm_layer=nn.LayerNorm, downsample=None, pretrained_window_size=0):
|
||||
|
||||
super().__init__()
|
||||
self.dim = dim
|
||||
self.input_resolution = input_resolution
|
||||
self.depth = depth
|
||||
self.grad_checkpointing = False
|
||||
|
||||
# build blocks
|
||||
self.blocks = nn.ModuleList([
|
||||
SwinTransformerBlock(
|
||||
dim=dim, input_resolution=input_resolution,
|
||||
num_heads=num_heads, window_size=window_size,
|
||||
shift_size=0 if (i % 2 == 0) else window_size // 2,
|
||||
mlp_ratio=mlp_ratio,
|
||||
qkv_bias=qkv_bias,
|
||||
drop=drop, attn_drop=attn_drop,
|
||||
drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
|
||||
norm_layer=norm_layer,
|
||||
pretrained_window_size=pretrained_window_size)
|
||||
for i in range(depth)])
|
||||
|
||||
# patch merging layer
|
||||
if downsample is not None:
|
||||
self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
|
||||
else:
|
||||
self.downsample = nn.Identity()
|
||||
|
||||
def forward(self, x):
|
||||
for blk in self.blocks:
|
||||
if self.grad_checkpointing and not torch.jit.is_scripting():
|
||||
x = checkpoint.checkpoint(blk, x)
|
||||
else:
|
||||
x = blk(x)
|
||||
x = self.downsample(x)
|
||||
return x
|
||||
|
||||
def _init_respostnorm(self):
|
||||
for blk in self.blocks:
|
||||
nn.init.constant_(blk.norm1.bias, 0)
|
||||
nn.init.constant_(blk.norm1.weight, 0)
|
||||
nn.init.constant_(blk.norm2.bias, 0)
|
||||
nn.init.constant_(blk.norm2.weight, 0)
|
||||
|
||||
|
||||
class SwinTransformerV2(nn.Module):
|
||||
r""" Swin Transformer V2
|
||||
A PyTorch impl of : `Swin Transformer V2: Scaling Up Capacity and Resolution`
|
||||
- https://arxiv.org/abs/2111.09883
|
||||
Args:
|
||||
img_size (int | tuple(int)): Input image size. Default 224
|
||||
patch_size (int | tuple(int)): Patch size. Default: 4
|
||||
in_chans (int): Number of input image channels. Default: 3
|
||||
num_classes (int): Number of classes for classification head. Default: 1000
|
||||
embed_dim (int): Patch embedding dimension. Default: 96
|
||||
depths (tuple(int)): Depth of each Swin Transformer layer.
|
||||
num_heads (tuple(int)): Number of attention heads in different layers.
|
||||
window_size (int): Window size. Default: 7
|
||||
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
|
||||
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
|
||||
drop_rate (float): Dropout rate. Default: 0
|
||||
attn_drop_rate (float): Attention dropout rate. Default: 0
|
||||
drop_path_rate (float): Stochastic depth rate. Default: 0.1
|
||||
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
|
||||
ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
|
||||
patch_norm (bool): If True, add normalization after patch embedding. Default: True
|
||||
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
|
||||
pretrained_window_sizes (tuple(int)): Pretrained window sizes of each layer.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, global_pool='avg',
|
||||
embed_dim=96, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24),
|
||||
window_size=7, mlp_ratio=4., qkv_bias=True,
|
||||
drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
|
||||
norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
|
||||
pretrained_window_sizes=(0, 0, 0, 0), **kwargs):
|
||||
super().__init__()
|
||||
|
||||
self.num_classes = num_classes
|
||||
assert global_pool in ('', 'avg')
|
||||
self.global_pool = global_pool
|
||||
self.num_layers = len(depths)
|
||||
self.embed_dim = embed_dim
|
||||
self.patch_norm = patch_norm
|
||||
self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
|
||||
|
||||
# split image into non-overlapping patches
|
||||
self.patch_embed = PatchEmbed(
|
||||
img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
|
||||
norm_layer=norm_layer if self.patch_norm else None)
|
||||
num_patches = self.patch_embed.num_patches
|
||||
|
||||
# absolute position embedding
|
||||
if ape:
|
||||
self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
|
||||
trunc_normal_(self.absolute_pos_embed, std=.02)
|
||||
else:
|
||||
self.absolute_pos_embed = None
|
||||
|
||||
self.pos_drop = nn.Dropout(p=drop_rate)
|
||||
|
||||
# stochastic depth
|
||||
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
|
||||
|
||||
# build layers
|
||||
self.layers = nn.ModuleList()
|
||||
for i_layer in range(self.num_layers):
|
||||
layer = BasicLayer(
|
||||
dim=int(embed_dim * 2 ** i_layer),
|
||||
input_resolution=(
|
||||
self.patch_embed.grid_size[0] // (2 ** i_layer),
|
||||
self.patch_embed.grid_size[1] // (2 ** i_layer)),
|
||||
depth=depths[i_layer],
|
||||
num_heads=num_heads[i_layer],
|
||||
window_size=window_size,
|
||||
mlp_ratio=mlp_ratio,
|
||||
qkv_bias=qkv_bias,
|
||||
drop=drop_rate, attn_drop=attn_drop_rate,
|
||||
drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
|
||||
norm_layer=norm_layer,
|
||||
downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
|
||||
pretrained_window_size=pretrained_window_sizes[i_layer]
|
||||
)
|
||||
self.layers.append(layer)
|
||||
|
||||
self.norm = norm_layer(self.num_features)
|
||||
self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
|
||||
|
||||
self.apply(self._init_weights)
|
||||
for bly in self.layers:
|
||||
bly._init_respostnorm()
|
||||
|
||||
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)
|
||||
|
||||
@torch.jit.ignore
|
||||
def no_weight_decay(self):
|
||||
nod = {'absolute_pos_embed'}
|
||||
for n, m in self.named_modules():
|
||||
if any([kw in n for kw in ("cpb_mlp", "logit_scale", 'relative_position_bias_table')]):
|
||||
nod.add(n)
|
||||
return nod
|
||||
|
||||
@torch.jit.ignore
|
||||
def group_matcher(self, coarse=False):
|
||||
return dict(
|
||||
stem=r'^absolute_pos_embed|patch_embed', # stem and embed
|
||||
blocks=r'^layers\.(\d+)' if coarse else [
|
||||
(r'^layers\.(\d+).downsample', (0,)),
|
||||
(r'^layers\.(\d+)\.\w+\.(\d+)', None),
|
||||
(r'^norm', (99999,)),
|
||||
]
|
||||
)
|
||||
|
||||
@torch.jit.ignore
|
||||
def set_grad_checkpointing(self, enable=True):
|
||||
for l in self.layers:
|
||||
l.grad_checkpointing = enable
|
||||
|
||||
@torch.jit.ignore
|
||||
def get_classifier(self):
|
||||
return self.head
|
||||
|
||||
def reset_classifier(self, num_classes, global_pool=None):
|
||||
self.num_classes = num_classes
|
||||
if global_pool is not None:
|
||||
assert global_pool in ('', 'avg')
|
||||
self.global_pool = global_pool
|
||||
self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
|
||||
|
||||
def forward_features(self, x):
|
||||
x = self.patch_embed(x)
|
||||
if self.absolute_pos_embed is not None:
|
||||
x = x + self.absolute_pos_embed
|
||||
x = self.pos_drop(x)
|
||||
|
||||
for layer in self.layers:
|
||||
x = layer(x)
|
||||
|
||||
x = self.norm(x) # B L C
|
||||
return x
|
||||
|
||||
def forward_head(self, x, pre_logits: bool = False):
|
||||
if self.global_pool == 'avg':
|
||||
x = x.mean(dim=1)
|
||||
return x if pre_logits else self.head(x)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.forward_features(x)
|
||||
x = self.forward_head(x)
|
||||
return x
|
||||
|
||||
|
||||
def checkpoint_filter_fn(state_dict, model):
|
||||
out_dict = {}
|
||||
if 'model' in state_dict:
|
||||
# For deit models
|
||||
state_dict = state_dict['model']
|
||||
for k, v in state_dict.items():
|
||||
if any([n in k for n in ('relative_position_index', 'relative_coords_table')]):
|
||||
continue # skip buffers that should not be persistent
|
||||
out_dict[k] = v
|
||||
return out_dict
|
||||
|
||||
|
||||
def _create_swin_transformer_v2(variant, pretrained=False, **kwargs):
|
||||
model = build_model_with_cfg(
|
||||
SwinTransformerV2, variant, pretrained,
|
||||
pretrained_filter_fn=checkpoint_filter_fn,
|
||||
**kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_tiny_window16_256(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=16, embed_dim=96, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24), **kwargs)
|
||||
return _create_swin_transformer_v2('swinv2_tiny_window16_256', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_tiny_window8_256(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=8, embed_dim=96, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24), **kwargs)
|
||||
return _create_swin_transformer_v2('swinv2_tiny_window8_256', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_small_window16_256(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=16, embed_dim=96, depths=(2, 2, 18, 2), num_heads=(3, 6, 12, 24), **kwargs)
|
||||
return _create_swin_transformer_v2('swinv2_small_window16_256', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_small_window8_256(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=8, embed_dim=96, depths=(2, 2, 18, 2), num_heads=(3, 6, 12, 24), **kwargs)
|
||||
return _create_swin_transformer_v2('swinv2_small_window8_256', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_base_window16_256(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=16, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32), **kwargs)
|
||||
return _create_swin_transformer_v2('swinv2_base_window16_256', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_base_window8_256(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=8, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32), **kwargs)
|
||||
return _create_swin_transformer_v2('swinv2_base_window8_256', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_base_window12_192_22k(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=12, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32), **kwargs)
|
||||
return _create_swin_transformer_v2('swinv2_base_window12_192_22k', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_base_window12to16_192to256_22kft1k(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=16, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32),
|
||||
pretrained_window_sizes=(12, 12, 12, 6), **kwargs)
|
||||
return _create_swin_transformer_v2(
|
||||
'swinv2_base_window12to16_192to256_22kft1k', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_base_window12to24_192to384_22kft1k(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=24, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32),
|
||||
pretrained_window_sizes=(12, 12, 12, 6), **kwargs)
|
||||
return _create_swin_transformer_v2(
|
||||
'swinv2_base_window12to24_192to384_22kft1k', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_large_window12_192_22k(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=12, embed_dim=192, depths=(2, 2, 18, 2), num_heads=(6, 12, 24, 48), **kwargs)
|
||||
return _create_swin_transformer_v2('swinv2_large_window12_192_22k', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_large_window12to16_192to256_22kft1k(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=16, embed_dim=192, depths=(2, 2, 18, 2), num_heads=(6, 12, 24, 48),
|
||||
pretrained_window_sizes=(12, 12, 12, 6), **kwargs)
|
||||
return _create_swin_transformer_v2(
|
||||
'swinv2_large_window12to16_192to256_22kft1k', pretrained=pretrained, **model_kwargs)
|
||||
|
||||
|
||||
@register_model
|
||||
def swinv2_large_window12to24_192to384_22kft1k(pretrained=False, **kwargs):
|
||||
"""
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
window_size=24, embed_dim=192, depths=(2, 2, 18, 2), num_heads=(6, 12, 24, 48),
|
||||
pretrained_window_sizes=(12, 12, 12, 6), **kwargs)
|
||||
return _create_swin_transformer_v2(
|
||||
'swinv2_large_window12to24_192to384_22kft1k', pretrained=pretrained, **model_kwargs)
|
||||
@ -0,0 +1,558 @@
|
||||
""" Relative Position Vision Transformer (ViT) in PyTorch
|
||||
|
||||
NOTE: these models are experimental / WIP, expect changes
|
||||
|
||||
Hacked together by / Copyright 2022, Ross Wightman
|
||||
"""
|
||||
import math
|
||||
import logging
|
||||
from functools import partial
|
||||
from collections import OrderedDict
|
||||
from typing import Optional, Tuple
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from torch.utils.checkpoint import checkpoint
|
||||
|
||||
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD, IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD
|
||||
from .helpers import build_model_with_cfg, resolve_pretrained_cfg, named_apply
|
||||
from .layers import PatchEmbed, Mlp, DropPath, trunc_normal_, lecun_normal_, to_2tuple
|
||||
from .registry import register_model
|
||||
|
||||
_logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
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_INCEPTION_MEAN, 'std': IMAGENET_INCEPTION_STD,
|
||||
'first_conv': 'patch_embed.proj', 'classifier': 'head',
|
||||
**kwargs
|
||||
}
|
||||
|
||||
|
||||
default_cfgs = {
|
||||
'vit_relpos_base_patch32_plus_rpn_256': _cfg(
|
||||
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_replos_base_patch32_plus_rpn_256-sw-dd486f51.pth',
|
||||
input_size=(3, 256, 256)),
|
||||
'vit_relpos_base_patch16_plus_240': _cfg(url='', input_size=(3, 240, 240)),
|
||||
|
||||
'vit_relpos_small_patch16_224': _cfg(
|
||||
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_small_patch16_224-sw-ec2778b4.pth'),
|
||||
'vit_relpos_medium_patch16_224': _cfg(
|
||||
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_medium_patch16_224-sw-11c174af.pth'),
|
||||
'vit_relpos_base_patch16_224': _cfg(
|
||||
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_base_patch16_224-sw-49049aed.pth'),
|
||||
|
||||
'vit_relpos_base_patch16_cls_224': _cfg(
|
||||
url=''),
|
||||
'vit_relpos_base_patch16_gapcls_224': _cfg(
|
||||
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_base_patch16_gapcls_224-sw-1a341d6c.pth'),
|
||||
|
||||
'vit_relpos_small_patch16_rpn_224': _cfg(url=''),
|
||||
'vit_relpos_medium_patch16_rpn_224': _cfg(
|
||||
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_medium_patch16_rpn_224-sw-5d2befd8.pth'),
|
||||
'vit_relpos_base_patch16_rpn_224': _cfg(url=''),
|
||||
}
|
||||
|
||||
|
||||
def gen_relative_position_index(win_size: Tuple[int, int], class_token: int = 0) -> torch.Tensor:
|
||||
# cut and paste w/ modifications from swin / beit codebase
|
||||
# cls to token & token 2 cls & cls to cls
|
||||
# get pair-wise relative position index for each token inside the window
|
||||
window_area = win_size[0] * win_size[1]
|
||||
coords = torch.stack(torch.meshgrid([torch.arange(win_size[0]), torch.arange(win_size[1])])).flatten(1) # 2, Wh, Ww
|
||||
relative_coords = coords[:, :, None] - coords[:, None, :] # 2, Wh*Ww, Wh*Ww
|
||||
relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
|
||||
relative_coords[:, :, 0] += win_size[0] - 1 # shift to start from 0
|
||||
relative_coords[:, :, 1] += win_size[1] - 1
|
||||
relative_coords[:, :, 0] *= 2 * win_size[1] - 1
|
||||
if class_token:
|
||||
num_relative_distance = (2 * win_size[0] - 1) * (2 * win_size[1] - 1) + 3
|
||||
relative_position_index = torch.zeros(size=(window_area + 1,) * 2, dtype=relative_coords.dtype)
|
||||
relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
|
||||
relative_position_index[0, 0:] = num_relative_distance - 3
|
||||
relative_position_index[0:, 0] = num_relative_distance - 2
|
||||
relative_position_index[0, 0] = num_relative_distance - 1
|
||||
else:
|
||||
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
|
||||
return relative_position_index
|
||||
|
||||
|
||||
def gen_relative_log_coords(
|
||||
win_size: Tuple[int, int],
|
||||
pretrained_win_size: Tuple[int, int] = (0, 0),
|
||||
mode='swin'
|
||||
):
|
||||
# as per official swin-v2 impl, supporting timm swin-v2-cr coords as well
|
||||
relative_coords_h = torch.arange(-(win_size[0] - 1), win_size[0], dtype=torch.float32)
|
||||
relative_coords_w = torch.arange(-(win_size[1] - 1), win_size[1], dtype=torch.float32)
|
||||
relative_coords_table = torch.stack(torch.meshgrid([relative_coords_h, relative_coords_w]))
|
||||
relative_coords_table = relative_coords_table.permute(1, 2, 0).contiguous() # 2*Wh-1, 2*Ww-1, 2
|
||||
if mode == 'swin':
|
||||
if pretrained_win_size[0] > 0:
|
||||
relative_coords_table[:, :, 0] /= (pretrained_win_size[0] - 1)
|
||||
relative_coords_table[:, :, 1] /= (pretrained_win_size[1] - 1)
|
||||
else:
|
||||
relative_coords_table[:, :, 0] /= (win_size[0] - 1)
|
||||
relative_coords_table[:, :, 1] /= (win_size[1] - 1)
|
||||
relative_coords_table *= 8 # normalize to -8, 8
|
||||
scale = math.log2(8)
|
||||
else:
|
||||
# FIXME we should support a form of normalization (to -1/1) for this mode?
|
||||
scale = math.log2(math.e)
|
||||
relative_coords_table = torch.sign(relative_coords_table) * torch.log2(
|
||||
1.0 + relative_coords_table.abs()) / scale
|
||||
return relative_coords_table
|
||||
|
||||
|
||||
class RelPosMlp(nn.Module):
|
||||
def __init__(
|
||||
self,
|
||||
window_size,
|
||||
num_heads=8,
|
||||
hidden_dim=128,
|
||||
class_token=False,
|
||||
mode='cr',
|
||||
pretrained_window_size=(0, 0)
|
||||
):
|
||||
super().__init__()
|
||||
self.window_size = window_size
|
||||
self.window_area = self.window_size[0] * self.window_size[1]
|
||||
self.class_token = 1 if class_token else 0
|
||||
self.num_heads = num_heads
|
||||
self.bias_shape = (self.window_area,) * 2 + (num_heads,)
|
||||
self.apply_sigmoid = mode == 'swin'
|
||||
|
||||
mlp_bias = (True, False) if mode == 'swin' else True
|
||||
self.mlp = Mlp(
|
||||
2, # x, y
|
||||
hidden_features=hidden_dim,
|
||||
out_features=num_heads,
|
||||
act_layer=nn.ReLU,
|
||||
bias=mlp_bias,
|
||||
drop=(0.125, 0.)
|
||||
)
|
||||
|
||||
self.register_buffer(
|
||||
"relative_position_index",
|
||||
gen_relative_position_index(window_size),
|
||||
persistent=False)
|
||||
|
||||
# get relative_coords_table
|
||||
self.register_buffer(
|
||||
"rel_coords_log",
|
||||
gen_relative_log_coords(window_size, pretrained_window_size, mode=mode),
|
||||
persistent=False)
|
||||
|
||||
def get_bias(self) -> torch.Tensor:
|
||||
relative_position_bias = self.mlp(self.rel_coords_log)
|
||||
if self.relative_position_index is not None:
|
||||
relative_position_bias = relative_position_bias.view(-1, self.num_heads)[
|
||||
self.relative_position_index.view(-1)] # Wh*Ww,Wh*Ww,nH
|
||||
relative_position_bias = relative_position_bias.view(self.bias_shape)
|
||||
relative_position_bias = relative_position_bias.permute(2, 0, 1)
|
||||
if self.apply_sigmoid:
|
||||
relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
|
||||
if self.class_token:
|
||||
relative_position_bias = F.pad(relative_position_bias, [self.class_token, 0, self.class_token, 0])
|
||||
return relative_position_bias.unsqueeze(0).contiguous()
|
||||
|
||||
def forward(self, attn, shared_rel_pos: Optional[torch.Tensor] = None):
|
||||
return attn + self.get_bias()
|
||||
|
||||
|
||||
class RelPosBias(nn.Module):
|
||||
|
||||
def __init__(self, window_size, num_heads, class_token=False):
|
||||
super().__init__()
|
||||
self.window_size = window_size
|
||||
self.window_area = window_size[0] * window_size[1]
|
||||
self.class_token = 1 if class_token else 0
|
||||
self.bias_shape = (self.window_area + self.class_token,) * 2 + (num_heads,)
|
||||
|
||||
num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3 * self.class_token
|
||||
self.relative_position_bias_table = nn.Parameter(torch.zeros(num_relative_distance, num_heads))
|
||||
self.register_buffer(
|
||||
"relative_position_index",
|
||||
gen_relative_position_index(self.window_size, class_token=self.class_token),
|
||||
persistent=False,
|
||||
)
|
||||
|
||||
self.init_weights()
|
||||
|
||||
def init_weights(self):
|
||||
trunc_normal_(self.relative_position_bias_table, std=.02)
|
||||
|
||||
def get_bias(self) -> torch.Tensor:
|
||||
relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)]
|
||||
# win_h * win_w, win_h * win_w, num_heads
|
||||
relative_position_bias = relative_position_bias.view(self.bias_shape).permute(2, 0, 1)
|
||||
return relative_position_bias.unsqueeze(0).contiguous()
|
||||
|
||||
def forward(self, attn, shared_rel_pos: Optional[torch.Tensor] = None):
|
||||
return attn + self.get_bias()
|
||||
|
||||
|
||||
class RelPosAttention(nn.Module):
|
||||
def __init__(self, dim, num_heads=8, qkv_bias=False, rel_pos_cls=None, attn_drop=0., proj_drop=0.):
|
||||
super().__init__()
|
||||
assert dim % num_heads == 0, 'dim should be divisible by num_heads'
|
||||
self.num_heads = num_heads
|
||||
head_dim = dim // num_heads
|
||||
self.scale = head_dim ** -0.5
|
||||
|
||||
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
||||
self.rel_pos = rel_pos_cls(num_heads=num_heads) if rel_pos_cls else None
|
||||
self.attn_drop = nn.Dropout(attn_drop)
|
||||
self.proj = nn.Linear(dim, dim)
|
||||
self.proj_drop = nn.Dropout(proj_drop)
|
||||
|
||||
def forward(self, x, shared_rel_pos: Optional[torch.Tensor] = None):
|
||||
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.unbind(0) # make torchscript happy (cannot use tensor as tuple)
|
||||
|
||||
attn = (q @ k.transpose(-2, -1)) * self.scale
|
||||
if self.rel_pos is not None:
|
||||
attn = self.rel_pos(attn, shared_rel_pos=shared_rel_pos)
|
||||
elif shared_rel_pos is not None:
|
||||
attn = attn + shared_rel_pos
|
||||
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 LayerScale(nn.Module):
|
||||
def __init__(self, dim, init_values=1e-5, inplace=False):
|
||||
super().__init__()
|
||||
self.inplace = inplace
|
||||
self.gamma = nn.Parameter(init_values * torch.ones(dim))
|
||||
|
||||
def forward(self, x):
|
||||
return x.mul_(self.gamma) if self.inplace else x * self.gamma
|
||||
|
||||
|
||||
class RelPosBlock(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self, dim, num_heads, mlp_ratio=4., qkv_bias=False, rel_pos_cls=None, init_values=None,
|
||||
drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
|
||||
super().__init__()
|
||||
self.norm1 = norm_layer(dim)
|
||||
self.attn = RelPosAttention(
|
||||
dim, num_heads, qkv_bias=qkv_bias, rel_pos_cls=rel_pos_cls, attn_drop=attn_drop, proj_drop=drop)
|
||||
self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
|
||||
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
|
||||
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
|
||||
self.norm2 = norm_layer(dim)
|
||||
self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)
|
||||
self.ls2 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
|
||||
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
|
||||
def forward(self, x, shared_rel_pos: Optional[torch.Tensor] = None):
|
||||
x = x + self.drop_path1(self.ls1(self.attn(self.norm1(x), shared_rel_pos=shared_rel_pos)))
|
||||
x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
|
||||
return x
|
||||
|
||||
|
||||
class ResPostRelPosBlock(nn.Module):
|
||||
|
||||
def __init__(
|
||||
self, dim, num_heads, mlp_ratio=4., qkv_bias=False, rel_pos_cls=None, init_values=None,
|
||||
drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
|
||||
super().__init__()
|
||||
self.init_values = init_values
|
||||
|
||||
self.attn = RelPosAttention(
|
||||
dim, num_heads, qkv_bias=qkv_bias, rel_pos_cls=rel_pos_cls, attn_drop=attn_drop, proj_drop=drop)
|
||||
self.norm1 = norm_layer(dim)
|
||||
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
|
||||
self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)
|
||||
self.norm2 = norm_layer(dim)
|
||||
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
||||
|
||||
self.init_weights()
|
||||
|
||||
def init_weights(self):
|
||||
# NOTE this init overrides that base model init with specific changes for the block type
|
||||
if self.init_values is not None:
|
||||
nn.init.constant_(self.norm1.weight, self.init_values)
|
||||
nn.init.constant_(self.norm2.weight, self.init_values)
|
||||
|
||||
def forward(self, x, shared_rel_pos: Optional[torch.Tensor] = None):
|
||||
x = x + self.drop_path1(self.norm1(self.attn(x, shared_rel_pos=shared_rel_pos)))
|
||||
x = x + self.drop_path2(self.norm2(self.mlp(x)))
|
||||
return x
|
||||
|
||||
|
||||
class VisionTransformerRelPos(nn.Module):
|
||||
""" Vision Transformer w/ Relative Position Bias
|
||||
|
||||
Differing from classic vit, this impl
|
||||
* uses relative position index (swin v1 / beit) or relative log coord + mlp (swin v2) pos embed
|
||||
* defaults to no class token (can be enabled)
|
||||
* defaults to global avg pool for head (can be changed)
|
||||
* layer-scale (residual branch gain) enabled
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, global_pool='avg',
|
||||
embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, init_values=1e-6,
|
||||
class_token=False, fc_norm=False, rel_pos_type='mlp', shared_rel_pos=False, rel_pos_dim=None,
|
||||
drop_rate=0., attn_drop_rate=0., drop_path_rate=0., weight_init='skip',
|
||||
embed_layer=PatchEmbed, norm_layer=None, act_layer=None, block_fn=RelPosBlock):
|
||||
"""
|
||||
Args:
|
||||
img_size (int, tuple): input image size
|
||||
patch_size (int, tuple): patch size
|
||||
in_chans (int): number of input channels
|
||||
num_classes (int): number of classes for classification head
|
||||
global_pool (str): type of global pooling for final sequence (default: 'avg')
|
||||
embed_dim (int): embedding dimension
|
||||
depth (int): depth of transformer
|
||||
num_heads (int): number of attention heads
|
||||
mlp_ratio (int): ratio of mlp hidden dim to embedding dim
|
||||
qkv_bias (bool): enable bias for qkv if True
|
||||
init_values: (float): layer-scale init values
|
||||
class_token (bool): use class token (default: False)
|
||||
fc_norm (bool): use pre classifier norm instead of pre-pool
|
||||
rel_pos_ty pe (str): type of relative position
|
||||
shared_rel_pos (bool): share relative pos across all blocks
|
||||
drop_rate (float): dropout rate
|
||||
attn_drop_rate (float): attention dropout rate
|
||||
drop_path_rate (float): stochastic depth rate
|
||||
weight_init (str): weight init scheme
|
||||
embed_layer (nn.Module): patch embedding layer
|
||||
norm_layer: (nn.Module): normalization layer
|
||||
act_layer: (nn.Module): MLP activation layer
|
||||
"""
|
||||
super().__init__()
|
||||
assert global_pool in ('', 'avg', 'token')
|
||||
assert class_token or global_pool != 'token'
|
||||
norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
|
||||
act_layer = act_layer or nn.GELU
|
||||
|
||||
self.num_classes = num_classes
|
||||
self.global_pool = global_pool
|
||||
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
|
||||
self.num_tokens = 1 if class_token else 0
|
||||
self.grad_checkpointing = False
|
||||
|
||||
self.patch_embed = embed_layer(
|
||||
img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
|
||||
feat_size = self.patch_embed.grid_size
|
||||
|
||||
rel_pos_args = dict(window_size=feat_size, class_token=class_token)
|
||||
if rel_pos_type.startswith('mlp'):
|
||||
if rel_pos_dim:
|
||||
rel_pos_args['hidden_dim'] = rel_pos_dim
|
||||
if 'swin' in rel_pos_type:
|
||||
rel_pos_args['mode'] = 'swin'
|
||||
rel_pos_cls = partial(RelPosMlp, **rel_pos_args)
|
||||
else:
|
||||
rel_pos_cls = partial(RelPosBias, **rel_pos_args)
|
||||
self.shared_rel_pos = None
|
||||
if shared_rel_pos:
|
||||
self.shared_rel_pos = rel_pos_cls(num_heads=num_heads)
|
||||
# NOTE shared rel pos currently mutually exclusive w/ per-block, but could support both...
|
||||
rel_pos_cls = None
|
||||
|
||||
self.cls_token = nn.Parameter(torch.zeros(1, self.num_tokens, embed_dim)) if self.num_tokens else None
|
||||
|
||||
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule
|
||||
self.blocks = nn.ModuleList([
|
||||
block_fn(
|
||||
dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, rel_pos_cls=rel_pos_cls,
|
||||
init_values=init_values, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i],
|
||||
norm_layer=norm_layer, act_layer=act_layer)
|
||||
for i in range(depth)])
|
||||
self.norm = norm_layer(embed_dim) if not fc_norm else nn.Identity()
|
||||
|
||||
# Classifier Head
|
||||
self.fc_norm = norm_layer(embed_dim) if fc_norm else nn.Identity()
|
||||
self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
||||
|
||||
if weight_init != 'skip':
|
||||
self.init_weights(weight_init)
|
||||
|
||||
def init_weights(self, mode=''):
|
||||
assert mode in ('jax', 'moco', '')
|
||||
if self.cls_token is not None:
|
||||
nn.init.normal_(self.cls_token, std=1e-6)
|
||||
# FIXME weight init scheme using PyTorch defaults curently
|
||||
#named_apply(get_init_weights_vit(mode, head_bias), self)
|
||||
|
||||
@torch.jit.ignore
|
||||
def no_weight_decay(self):
|
||||
return {'cls_token'}
|
||||
|
||||
@torch.jit.ignore
|
||||
def group_matcher(self, coarse=False):
|
||||
return dict(
|
||||
stem=r'^cls_token|patch_embed', # stem and embed
|
||||
blocks=[(r'^blocks\.(\d+)', None), (r'^norm', (99999,))]
|
||||
)
|
||||
|
||||
@torch.jit.ignore
|
||||
def set_grad_checkpointing(self, enable=True):
|
||||
self.grad_checkpointing = enable
|
||||
|
||||
@torch.jit.ignore
|
||||
def get_classifier(self):
|
||||
return self.head
|
||||
|
||||
def reset_classifier(self, num_classes: int, global_pool=None):
|
||||
self.num_classes = num_classes
|
||||
if global_pool is not None:
|
||||
assert global_pool in ('', 'avg', 'token')
|
||||
self.global_pool = global_pool
|
||||
self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
||||
|
||||
def forward_features(self, x):
|
||||
x = self.patch_embed(x)
|
||||
if self.cls_token is not None:
|
||||
x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
|
||||
|
||||
shared_rel_pos = self.shared_rel_pos.get_bias() if self.shared_rel_pos is not None else None
|
||||
for blk in self.blocks:
|
||||
if self.grad_checkpointing and not torch.jit.is_scripting():
|
||||
x = checkpoint(blk, x, shared_rel_pos=shared_rel_pos)
|
||||
else:
|
||||
x = blk(x, shared_rel_pos=shared_rel_pos)
|
||||
x = self.norm(x)
|
||||
return x
|
||||
|
||||
def forward_head(self, x, pre_logits: bool = False):
|
||||
if self.global_pool:
|
||||
x = x[:, self.num_tokens:].mean(dim=1) if self.global_pool == 'avg' else x[:, 0]
|
||||
x = self.fc_norm(x)
|
||||
return x if pre_logits else self.head(x)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.forward_features(x)
|
||||
x = self.forward_head(x)
|
||||
return x
|
||||
|
||||
|
||||
def _create_vision_transformer_relpos(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(VisionTransformerRelPos, variant, pretrained, **kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_base_patch32_plus_rpn_256(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/32+) w/ relative log-coord position and residual post-norm, no class token
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
patch_size=32, embed_dim=896, depth=12, num_heads=14, block_fn=ResPostRelPosBlock, **kwargs)
|
||||
model = _create_vision_transformer_relpos(
|
||||
'vit_relpos_base_patch32_plus_rpn_256', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_base_patch16_plus_240(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/16+) w/ relative log-coord position, no class token
|
||||
"""
|
||||
model_kwargs = dict(patch_size=16, embed_dim=896, depth=12, num_heads=14, **kwargs)
|
||||
model = _create_vision_transformer_relpos('vit_relpos_base_patch16_plus_240', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_small_patch16_224(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/16) w/ relative log-coord position, no class token
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
patch_size=16, embed_dim=384, depth=12, num_heads=6, qkv_bias=False, fc_norm=True, **kwargs)
|
||||
model = _create_vision_transformer_relpos('vit_relpos_small_patch16_224', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_medium_patch16_224(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/16) w/ relative log-coord position, no class token
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
patch_size=16, embed_dim=512, depth=12, num_heads=8, qkv_bias=False, fc_norm=True, **kwargs)
|
||||
model = _create_vision_transformer_relpos('vit_relpos_medium_patch16_224', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_base_patch16_224(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/16) w/ relative log-coord position, no class token
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
patch_size=16, embed_dim=768, depth=12, num_heads=12, qkv_bias=False, fc_norm=True, **kwargs)
|
||||
model = _create_vision_transformer_relpos('vit_relpos_base_patch16_224', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_base_patch16_cls_224(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/16) w/ relative log-coord position, class token present
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
patch_size=16, embed_dim=768, depth=12, num_heads=12, qkv_bias=False,
|
||||
class_token=True, global_pool='token', **kwargs)
|
||||
model = _create_vision_transformer_relpos('vit_relpos_base_patch16_cls_224', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_base_patch16_gapcls_224(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/16) w/ relative log-coord position, class token present
|
||||
NOTE this config is a bit of a mistake, class token was enabled but global avg-pool w/ fc-norm was not disabled
|
||||
Leaving here for comparisons w/ a future re-train as it performs quite well.
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
patch_size=16, embed_dim=768, depth=12, num_heads=12, qkv_bias=False, fc_norm=True, class_token=True, **kwargs)
|
||||
model = _create_vision_transformer_relpos('vit_relpos_base_patch16_gapcls_224', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_small_patch16_rpn_224(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/16) w/ relative log-coord position and residual post-norm, no class token
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
patch_size=16, embed_dim=384, depth=12, num_heads=6, qkv_bias=False, block_fn=ResPostRelPosBlock, **kwargs)
|
||||
model = _create_vision_transformer_relpos(
|
||||
'vit_relpos_small_patch16_rpn_224', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_medium_patch16_rpn_224(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/16) w/ relative log-coord position and residual post-norm, no class token
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
patch_size=16, embed_dim=512, depth=12, num_heads=8, qkv_bias=False, block_fn=ResPostRelPosBlock, **kwargs)
|
||||
model = _create_vision_transformer_relpos(
|
||||
'vit_relpos_medium_patch16_rpn_224', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def vit_relpos_base_patch16_rpn_224(pretrained=False, **kwargs):
|
||||
""" ViT-Base (ViT-B/16) w/ relative log-coord position and residual post-norm, no class token
|
||||
"""
|
||||
model_kwargs = dict(
|
||||
patch_size=16, embed_dim=768, depth=12, num_heads=12, qkv_bias=False, block_fn=ResPostRelPosBlock, **kwargs)
|
||||
model = _create_vision_transformer_relpos(
|
||||
'vit_relpos_base_patch16_rpn_224', pretrained=pretrained, **model_kwargs)
|
||||
return model
|
||||
@ -1 +1 @@
|
||||
__version__ = '0.6.1'
|
||||
__version__ = '0.7.0.dev0'
|
||||
|
||||
Loading…
Reference in new issue