onescotch
add huggingface implementation
2de1f98
raw
history blame
28.4 kB
# Copyright (c) OpenMMLab. All rights reserved.
from collections import OrderedDict
from copy import deepcopy
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as cp
from mmcv.cnn import build_norm_layer, constant_init, trunc_normal_init
from mmcv.cnn.bricks.transformer import FFN, build_dropout
from mmcv.cnn.utils.weight_init import trunc_normal_
from mmcv.runner import _load_checkpoint
from mmcv.utils import to_2tuple
from ...utils import get_root_logger
from ..builder import BACKBONES
from ..utils.transformer import PatchEmbed, PatchMerging
from .base_backbone import BaseBackbone
from .utils.ckpt_convert import swin_converter
class WindowMSA(nn.Module):
"""Window based multi-head self-attention (W-MSA) module with relative
position bias.
Args:
embed_dims (int): Number of input channels.
num_heads (int): Number of attention heads.
window_size (tuple[int]): The height and width of the window.
qkv_bias (bool, optional): If True, add a learnable bias to q, k, v.
Default: True.
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Default: None.
attn_drop_rate (float, optional): Dropout ratio of attention weight.
Default: 0.0
proj_drop_rate (float, optional): Dropout ratio of output. Default: 0.
"""
def __init__(self,
embed_dims,
num_heads,
window_size,
qkv_bias=True,
qk_scale=None,
attn_drop_rate=0.,
proj_drop_rate=0.):
super().__init__()
self.embed_dims = embed_dims
self.window_size = window_size # Wh, Ww
self.num_heads = num_heads
head_embed_dims = embed_dims // num_heads
self.scale = qk_scale or head_embed_dims**-0.5
# define a parameter table of relative position bias
self.relative_position_bias_table = nn.Parameter(
torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1),
num_heads)) # 2*Wh-1 * 2*Ww-1, nH
# About 2x faster than original impl
Wh, Ww = self.window_size
rel_index_coords = self.double_step_seq(2 * Ww - 1, Wh, 1, Ww)
rel_position_index = rel_index_coords + rel_index_coords.T
rel_position_index = rel_position_index.flip(1).contiguous()
self.register_buffer('relative_position_index', rel_position_index)
self.qkv = nn.Linear(embed_dims, embed_dims * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop_rate)
self.proj = nn.Linear(embed_dims, embed_dims)
self.proj_drop = nn.Dropout(proj_drop_rate)
self.softmax = nn.Softmax(dim=-1)
def init_weights(self):
trunc_normal_(self.relative_position_bias_table, std=0.02)
def forward(self, x, mask=None):
"""
Args:
x (tensor): input features with shape of (num_windows*B, N, C)
mask (tensor | None, Optional): mask with shape of (num_windows,
Wh*Ww, Wh*Ww), value should be between (-inf, 0].
"""
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)
# make torchscript happy (cannot use tensor as tuple)
q, k, v = qkv[0], qkv[1], qkv[2]
q = q * self.scale
attn = (q @ k.transpose(-2, -1))
relative_position_bias = self.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
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)
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
@staticmethod
def double_step_seq(step1, len1, step2, len2):
seq1 = torch.arange(0, step1 * len1, step1)
seq2 = torch.arange(0, step2 * len2, step2)
return (seq1[:, None] + seq2[None, :]).reshape(1, -1)
class ShiftWindowMSA(nn.Module):
"""Shifted Window Multihead Self-Attention Module.
Args:
embed_dims (int): Number of input channels.
num_heads (int): Number of attention heads.
window_size (int): The height and width of the window.
shift_size (int, optional): The shift step of each window towards
right-bottom. If zero, act as regular window-msa. Defaults to 0.
qkv_bias (bool, optional): If True, add a learnable bias to q, k, v.
Default: True
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Defaults: None.
attn_drop_rate (float, optional): Dropout ratio of attention weight.
Defaults: 0.
proj_drop_rate (float, optional): Dropout ratio of output.
Defaults: 0.
dropout_layer (dict, optional): The dropout_layer used before output.
Defaults: dict(type='DropPath', drop_prob=0.).
"""
def __init__(self,
embed_dims,
num_heads,
window_size,
shift_size=0,
qkv_bias=True,
qk_scale=None,
attn_drop_rate=0,
proj_drop_rate=0,
dropout_layer=dict(type='DropPath', drop_prob=0.)):
super().__init__()
self.window_size = window_size
self.shift_size = shift_size
assert 0 <= self.shift_size < self.window_size
self.w_msa = WindowMSA(
embed_dims=embed_dims,
num_heads=num_heads,
window_size=to_2tuple(window_size),
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop_rate=attn_drop_rate,
proj_drop_rate=proj_drop_rate)
self.drop = build_dropout(dropout_layer)
def forward(self, query, hw_shape):
B, L, C = query.shape
H, W = hw_shape
assert L == H * W, 'input feature has wrong size'
query = query.view(B, H, W, C)
# pad feature maps to multiples of window size
pad_r = (self.window_size - W % self.window_size) % self.window_size
pad_b = (self.window_size - H % self.window_size) % self.window_size
query = F.pad(query, (0, 0, 0, pad_r, 0, pad_b))
H_pad, W_pad = query.shape[1], query.shape[2]
# cyclic shift
if self.shift_size > 0:
shifted_query = torch.roll(
query,
shifts=(-self.shift_size, -self.shift_size),
dims=(1, 2))
# calculate attention mask for SW-MSA
img_mask = torch.zeros((1, H_pad, W_pad, 1), device=query.device)
h_slices = (slice(0, -self.window_size),
slice(-self.window_size,
-self.shift_size), slice(-self.shift_size, None))
w_slices = (slice(0, -self.window_size),
slice(-self.window_size,
-self.shift_size), slice(-self.shift_size, None))
cnt = 0
for h in h_slices:
for w in w_slices:
img_mask[:, h, w, :] = cnt
cnt += 1
# nW, window_size, window_size, 1
mask_windows = self.window_partition(img_mask)
mask_windows = mask_windows.view(
-1, self.window_size * self.window_size)
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:
shifted_query = query
attn_mask = None
# nW*B, window_size, window_size, C
query_windows = self.window_partition(shifted_query)
# nW*B, window_size*window_size, C
query_windows = query_windows.view(-1, self.window_size**2, C)
# W-MSA/SW-MSA (nW*B, window_size*window_size, C)
attn_windows = self.w_msa(query_windows, mask=attn_mask)
# merge windows
attn_windows = attn_windows.view(-1, self.window_size,
self.window_size, C)
# B H' W' C
shifted_x = self.window_reverse(attn_windows, H_pad, W_pad)
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(
shifted_x,
shifts=(self.shift_size, self.shift_size),
dims=(1, 2))
else:
x = shifted_x
if pad_r > 0 or pad_b:
x = x[:, :H, :W, :].contiguous()
x = x.view(B, H * W, C)
x = self.drop(x)
return x
def window_reverse(self, windows, H, W):
"""
Args:
windows: (num_windows*B, window_size, window_size, C)
H (int): Height of image
W (int): Width of image
Returns:
x: (B, H, W, C)
"""
window_size = self.window_size
B = int(windows.shape[0] / (H * W / window_size / window_size))
x = windows.view(B, H // window_size, W // window_size, window_size,
window_size, -1)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x
def window_partition(self, x):
"""
Args:
x: (B, H, W, C)
Returns:
windows: (num_windows*B, window_size, window_size, C)
"""
B, H, W, C = x.shape
window_size = self.window_size
x = x.view(B, H // window_size, window_size, W // window_size,
window_size, C)
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous()
windows = windows.view(-1, window_size, window_size, C)
return windows
class SwinBlock(nn.Module):
""""
Args:
embed_dims (int): The feature dimension.
num_heads (int): Parallel attention heads.
feedforward_channels (int): The hidden dimension for FFNs.
window_size (int, optional): The local window scale. Default: 7.
shift (bool, optional): whether to shift window or not. Default False.
qkv_bias (bool, optional): enable bias for qkv if True. Default: True.
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Default: None.
drop_rate (float, optional): Dropout rate. Default: 0.
attn_drop_rate (float, optional): Attention dropout rate. Default: 0.
drop_path_rate (float, optional): Stochastic depth rate. Default: 0.
act_cfg (dict, optional): The config dict of activation function.
Default: dict(type='GELU').
norm_cfg (dict, optional): The config dict of normalization.
Default: dict(type='LN').
with_cp (bool, optional): Use checkpoint or not. Using checkpoint
will save some memory while slowing down the training speed.
Default: False.
"""
def __init__(self,
embed_dims,
num_heads,
feedforward_channels,
window_size=7,
shift=False,
qkv_bias=True,
qk_scale=None,
drop_rate=0.,
attn_drop_rate=0.,
drop_path_rate=0.,
act_cfg=dict(type='GELU'),
norm_cfg=dict(type='LN'),
with_cp=False):
super(SwinBlock, self).__init__()
self.with_cp = with_cp
self.norm1 = build_norm_layer(norm_cfg, embed_dims)[1]
self.attn = ShiftWindowMSA(
embed_dims=embed_dims,
num_heads=num_heads,
window_size=window_size,
shift_size=window_size // 2 if shift else 0,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop_rate=attn_drop_rate,
proj_drop_rate=drop_rate,
dropout_layer=dict(type='DropPath', drop_prob=drop_path_rate))
self.norm2 = build_norm_layer(norm_cfg, embed_dims)[1]
self.ffn = FFN(
embed_dims=embed_dims,
feedforward_channels=feedforward_channels,
num_fcs=2,
ffn_drop=drop_rate,
dropout_layer=dict(type='DropPath', drop_prob=drop_path_rate),
act_cfg=act_cfg,
add_identity=True,
init_cfg=None)
def forward(self, x, hw_shape):
def _inner_forward(x):
identity = x
x = self.norm1(x)
x = self.attn(x, hw_shape)
x = x + identity
identity = x
x = self.norm2(x)
x = self.ffn(x, identity=identity)
return x
if self.with_cp and x.requires_grad:
x = cp.checkpoint(_inner_forward, x)
else:
x = _inner_forward(x)
return x
class SwinBlockSequence(nn.Module):
"""Implements one stage in Swin Transformer.
Args:
embed_dims (int): The feature dimension.
num_heads (int): Parallel attention heads.
feedforward_channels (int): The hidden dimension for FFNs.
depth (int): The number of blocks in this stage.
window_size (int, optional): The local window scale. Default: 7.
qkv_bias (bool, optional): enable bias for qkv if True. Default: True.
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Default: None.
drop_rate (float, optional): Dropout rate. Default: 0.
attn_drop_rate (float, optional): Attention dropout rate. Default: 0.
drop_path_rate (float | list[float], optional): Stochastic depth
rate. Default: 0.
downsample (nn.Module | None, optional): The downsample operation
module. Default: None.
act_cfg (dict, optional): The config dict of activation function.
Default: dict(type='GELU').
norm_cfg (dict, optional): The config dict of normalization.
Default: dict(type='LN').
with_cp (bool, optional): Use checkpoint or not. Using checkpoint
will save some memory while slowing down the training speed.
Default: False.
"""
def __init__(self,
embed_dims,
num_heads,
feedforward_channels,
depth,
window_size=7,
qkv_bias=True,
qk_scale=None,
drop_rate=0.,
attn_drop_rate=0.,
drop_path_rate=0.,
downsample=None,
act_cfg=dict(type='GELU'),
norm_cfg=dict(type='LN'),
with_cp=False):
super().__init__()
if isinstance(drop_path_rate, list):
drop_path_rates = drop_path_rate
assert len(drop_path_rates) == depth
else:
drop_path_rates = [deepcopy(drop_path_rate) for _ in range(depth)]
self.blocks = nn.ModuleList()
for i in range(depth):
block = SwinBlock(
embed_dims=embed_dims,
num_heads=num_heads,
feedforward_channels=feedforward_channels,
window_size=window_size,
shift=False if i % 2 == 0 else True,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
drop_path_rate=drop_path_rates[i],
act_cfg=act_cfg,
norm_cfg=norm_cfg,
with_cp=with_cp)
self.blocks.append(block)
self.downsample = downsample
def forward(self, x, hw_shape):
for block in self.blocks:
x = block(x, hw_shape)
if self.downsample:
x_down, down_hw_shape = self.downsample(x, hw_shape)
return x_down, down_hw_shape, x, hw_shape
else:
return x, hw_shape, x, hw_shape
@BACKBONES.register_module()
class SwinTransformer(BaseBackbone):
""" Swin Transformer
A PyTorch implement of : `Swin Transformer:
Hierarchical Vision Transformer using Shifted Windows` -
https://arxiv.org/abs/2103.14030
Inspiration from
https://github.com/microsoft/Swin-Transformer
Args:
pretrain_img_size (int | tuple[int]): The size of input image when
pretrain. Defaults: 224.
in_channels (int): The num of input channels.
Defaults: 3.
embed_dims (int): The feature dimension. Default: 96.
patch_size (int | tuple[int]): Patch size. Default: 4.
window_size (int): Window size. Default: 7.
mlp_ratio (int): Ratio of mlp hidden dim to embedding dim.
Default: 4.
depths (tuple[int]): Depths of each Swin Transformer stage.
Default: (2, 2, 6, 2).
num_heads (tuple[int]): Parallel attention heads of each Swin
Transformer stage. Default: (3, 6, 12, 24).
strides (tuple[int]): The patch merging or patch embedding stride of
each Swin Transformer stage. (In swin, we set kernel size equal to
stride.) Default: (4, 2, 2, 2).
out_indices (tuple[int]): Output from which stages.
Default: (0, 1, 2, 3).
qkv_bias (bool, optional): If True, add a learnable bias to query, key,
value. Default: True
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Default: None.
patch_norm (bool): If add a norm layer for patch embed and patch
merging. Default: True.
drop_rate (float): Dropout rate. Defaults: 0.
attn_drop_rate (float): Attention dropout rate. Default: 0.
drop_path_rate (float): Stochastic depth rate. Defaults: 0.1.
use_abs_pos_embed (bool): If True, add absolute position embedding to
the patch embedding. Defaults: False.
act_cfg (dict): Config dict for activation layer.
Default: dict(type='LN').
norm_cfg (dict): Config dict for normalization layer at
output of backone. Defaults: dict(type='LN').
with_cp (bool, optional): Use checkpoint or not. Using checkpoint
will save some memory while slowing down the training speed.
Default: False.
pretrained (str, optional): model pretrained path. Default: None.
convert_weights (bool): The flag indicates whether the
pre-trained model is from the original repo. We may need
to convert some keys to make it compatible.
Default: False.
frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
Default: -1 (-1 means not freezing any parameters).
"""
def __init__(
self,
pretrain_img_size=224,
in_channels=3,
embed_dims=96,
patch_size=4,
window_size=7,
mlp_ratio=4,
depths=(2, 2, 6, 2),
num_heads=(3, 6, 12, 24),
strides=(4, 2, 2, 2),
out_indices=(0, 1, 2, 3),
qkv_bias=True,
qk_scale=None,
patch_norm=True,
drop_rate=0.,
attn_drop_rate=0.,
drop_path_rate=0.1,
use_abs_pos_embed=False,
act_cfg=dict(type='GELU'),
norm_cfg=dict(type='LN'),
with_cp=False,
convert_weights=False,
frozen_stages=-1,
):
self.convert_weights = convert_weights
self.frozen_stages = frozen_stages
if isinstance(pretrain_img_size, int):
pretrain_img_size = to_2tuple(pretrain_img_size)
elif isinstance(pretrain_img_size, tuple):
if len(pretrain_img_size) == 1:
pretrain_img_size = to_2tuple(pretrain_img_size[0])
assert len(pretrain_img_size) == 2, \
f'The size of image should have length 1 or 2, ' \
f'but got {len(pretrain_img_size)}'
super(SwinTransformer, self).__init__()
num_layers = len(depths)
self.out_indices = out_indices
self.use_abs_pos_embed = use_abs_pos_embed
assert strides[0] == patch_size, 'Use non-overlapping patch embed.'
self.patch_embed = PatchEmbed(
in_channels=in_channels,
embed_dims=embed_dims,
conv_type='Conv2d',
kernel_size=patch_size,
stride=strides[0],
norm_cfg=norm_cfg if patch_norm else None,
init_cfg=None)
if self.use_abs_pos_embed:
patch_row = pretrain_img_size[0] // patch_size
patch_col = pretrain_img_size[1] // patch_size
num_patches = patch_row * patch_col
self.absolute_pos_embed = nn.Parameter(
torch.zeros((1, num_patches, embed_dims)))
self.drop_after_pos = nn.Dropout(p=drop_rate)
# set stochastic depth decay rule
total_depth = sum(depths)
dpr = [
x.item() for x in torch.linspace(0, drop_path_rate, total_depth)
]
self.stages = nn.ModuleList()
in_channels = embed_dims
for i in range(num_layers):
if i < num_layers - 1:
downsample = PatchMerging(
in_channels=in_channels,
out_channels=2 * in_channels,
stride=strides[i + 1],
norm_cfg=norm_cfg if patch_norm else None,
init_cfg=None)
else:
downsample = None
stage = SwinBlockSequence(
embed_dims=in_channels,
num_heads=num_heads[i],
feedforward_channels=mlp_ratio * in_channels,
depth=depths[i],
window_size=window_size,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop_rate=drop_rate,
attn_drop_rate=attn_drop_rate,
drop_path_rate=dpr[sum(depths[:i]):sum(depths[:i + 1])],
downsample=downsample,
act_cfg=act_cfg,
norm_cfg=norm_cfg,
with_cp=with_cp)
self.stages.append(stage)
if downsample:
in_channels = downsample.out_channels
self.num_features = [int(embed_dims * 2**i) for i in range(num_layers)]
# Add a norm layer for each output
for i in out_indices:
layer = build_norm_layer(norm_cfg, self.num_features[i])[1]
layer_name = f'norm{i}'
self.add_module(layer_name, layer)
def train(self, mode=True):
"""Convert the model into training mode while keep layers freezed."""
super(SwinTransformer, self).train(mode)
self._freeze_stages()
def _freeze_stages(self):
if self.frozen_stages >= 0:
self.patch_embed.eval()
for param in self.patch_embed.parameters():
param.requires_grad = False
if self.use_abs_pos_embed:
self.absolute_pos_embed.requires_grad = False
self.drop_after_pos.eval()
for i in range(1, self.frozen_stages + 1):
if (i - 1) in self.out_indices:
norm_layer = getattr(self, f'norm{i-1}')
norm_layer.eval()
for param in norm_layer.parameters():
param.requires_grad = False
m = self.stages[i - 1]
m.eval()
for param in m.parameters():
param.requires_grad = False
def init_weights(self, pretrained=None):
"""Initialize the weights in backbone.
Args:
pretrained (str, optional): Path to pre-trained weights.
Defaults to None.
"""
if isinstance(pretrained, str):
logger = get_root_logger()
ckpt = _load_checkpoint(
pretrained, logger=logger, map_location='cpu')
if 'state_dict' in ckpt:
_state_dict = ckpt['state_dict']
elif 'model' in ckpt:
_state_dict = ckpt['model']
else:
_state_dict = ckpt
if self.convert_weights:
# supported loading weight from original repo,
_state_dict = swin_converter(_state_dict)
state_dict = OrderedDict()
for k, v in _state_dict.items():
if k.startswith('backbone.'):
state_dict[k[9:]] = v
# strip prefix of state_dict
if list(state_dict.keys())[0].startswith('module.'):
state_dict = {k[7:]: v for k, v in state_dict.items()}
# reshape absolute position embedding
if state_dict.get('absolute_pos_embed') is not None:
absolute_pos_embed = state_dict['absolute_pos_embed']
N1, L, C1 = absolute_pos_embed.size()
N2, C2, H, W = self.absolute_pos_embed.size()
if N1 != N2 or C1 != C2 or L != H * W:
logger.warning('Error in loading absolute_pos_embed, pass')
else:
state_dict['absolute_pos_embed'] = absolute_pos_embed.view(
N2, H, W, C2).permute(0, 3, 1, 2).contiguous()
# interpolate position bias table if needed
relative_position_bias_table_keys = [
k for k in state_dict.keys()
if 'relative_position_bias_table' in k
]
for table_key in relative_position_bias_table_keys:
table_pretrained = state_dict[table_key]
table_current = self.state_dict()[table_key]
L1, nH1 = table_pretrained.size()
L2, nH2 = table_current.size()
if nH1 != nH2:
logger.warning(f'Error in loading {table_key}, pass')
elif L1 != L2:
S1 = int(L1**0.5)
S2 = int(L2**0.5)
table_pretrained_resized = F.interpolate(
table_pretrained.permute(1, 0).reshape(1, nH1, S1, S1),
size=(S2, S2),
mode='bicubic')
state_dict[table_key] = table_pretrained_resized.view(
nH2, L2).permute(1, 0).contiguous()
# load state_dict
self.load_state_dict(state_dict, False)
elif pretrained is None:
if self.use_abs_pos_embed:
trunc_normal_(self.absolute_pos_embed, std=0.02)
for m in self.modules():
if isinstance(m, nn.Linear):
trunc_normal_init(m, std=.02, bias=0.)
elif isinstance(m, nn.LayerNorm):
constant_init(m, 1.0)
else:
raise TypeError('pretrained must be a str or None')
def forward(self, x):
x, hw_shape = self.patch_embed(x)
if self.use_abs_pos_embed:
x = x + self.absolute_pos_embed
x = self.drop_after_pos(x)
outs = []
for i, stage in enumerate(self.stages):
x, hw_shape, out, out_hw_shape = stage(x, hw_shape)
if i in self.out_indices:
norm_layer = getattr(self, f'norm{i}')
out = norm_layer(out)
out = out.view(-1, *out_hw_shape,
self.num_features[i]).permute(0, 3, 1,
2).contiguous()
outs.append(out)
return outs