""" CrossViT Model
@inproceedings{
chen2021crossvit,
title={{CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification}},
author={Chun-Fu (Richard) Chen and Quanfu Fan and Rameswar Panda},
booktitle={International Conference on Computer Vision (ICCV)},
year={2021}
}
Paper link: https://arxiv.org/abs/2103.14899
Original code: https://github.com/IBM/CrossViT/blob/main/models/crossvit.py
NOTE: model names have been renamed from originals to represent actual input res all *_224 -> *_240 and *_384 -> *_408
Modifications and additions for timm hacked together by / Copyright 2021, Ross Wightman
"""
# Copyright IBM All Rights Reserved.
# SPDX-License-Identifier: Apache-2.0
"""
Modifed from Timm. https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
"""
from functools import partial
from typing import List
from typing import Tuple
import torch
import torch.hub
import torch.nn as nn
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.layers import DropPath, to_2tuple, trunc_normal_, _assert
from ._builder import build_model_with_cfg
from ._features_fx import register_notrace_function
from ._registry import register_model
from .vision_transformer import Block
__all__ = ['CrossViT'] # model_registry will add each entrypoint fn to this
def _cfg(url='', **kwargs):
return {
'url': url,
'num_classes': 1000, 'input_size': (3, 240, 240), 'pool_size': None, 'crop_pct': 0.875,
'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD, 'fixed_input_size': True,
'first_conv': ('patch_embed.0.proj', 'patch_embed.1.proj'),
'classifier': ('head.0', 'head.1'),
**kwargs
}
default_cfgs = {
'crossvit_15_240': _cfg(url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_15_224.pth'),
'crossvit_15_dagger_240': _cfg(
url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_15_dagger_224.pth',
first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'),
),
'crossvit_15_dagger_408': _cfg(
url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_15_dagger_384.pth',
input_size=(3, 408, 408), first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'), crop_pct=1.0,
),
'crossvit_18_240': _cfg(url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_18_224.pth'),
'crossvit_18_dagger_240': _cfg(
url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_18_dagger_224.pth',
first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'),
),
'crossvit_18_dagger_408': _cfg(
url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_18_dagger_384.pth',
input_size=(3, 408, 408), first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'), crop_pct=1.0,
),
'crossvit_9_240': _cfg(url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_9_224.pth'),
'crossvit_9_dagger_240': _cfg(
url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_9_dagger_224.pth',
first_conv=('patch_embed.0.proj.0', 'patch_embed.1.proj.0'),
),
'crossvit_base_240': _cfg(
url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_base_224.pth'),
'crossvit_small_240': _cfg(
url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_small_224.pth'),
'crossvit_tiny_240': _cfg(
url='https://github.com/IBM/CrossViT/releases/download/weights-0.1/crossvit_tiny_224.pth'),
}
class PatchEmbed(nn.Module):
""" Image to Patch Embedding
"""
def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, multi_conv=False):
super().__init__()
img_size = to_2tuple(img_size)
patch_size = to_2tuple(patch_size)
num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
self.img_size = img_size
self.patch_size = patch_size
self.num_patches = num_patches
if multi_conv:
if patch_size[0] == 12:
self.proj = nn.Sequential(
nn.Conv2d(in_chans, embed_dim // 4, kernel_size=7, stride=4, padding=3),
nn.ReLU(inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim // 2, kernel_size=3, stride=3, padding=0),
nn.ReLU(inplace=True),
nn.Conv2d(embed_dim // 2, embed_dim, kernel_size=3, stride=1, padding=1),
)
elif patch_size[0] == 16:
self.proj = nn.Sequential(
nn.Conv2d(in_chans, embed_dim // 4, kernel_size=7, stride=4, padding=3),
nn.ReLU(inplace=True),
nn.Conv2d(embed_dim // 4, embed_dim // 2, kernel_size=3, stride=2, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(embed_dim // 2, embed_dim, kernel_size=3, stride=2, padding=1),
)
else:
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
def forward(self, x):
B, C, H, W = x.shape
# FIXME look at relaxing size constraints
_assert(H == self.img_size[0],
f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]}).")
_assert(W == self.img_size[1],
f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]}).")
x = self.proj(x).flatten(2).transpose(1, 2)
return x
class CrossAttention(nn.Module):
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
# NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
self.scale = qk_scale or head_dim ** -0.5
self.wq = nn.Linear(dim, dim, bias=qkv_bias)
self.wk = nn.Linear(dim, dim, bias=qkv_bias)
self.wv = nn.Linear(dim, dim, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x):
B, N, C = x.shape
# B1C -> B1H(C/H) -> BH1(C/H)
q = self.wq(x[:, 0:1, ...]).reshape(B, 1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
# BNC -> BNH(C/H) -> BHN(C/H)
k = self.wk(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
# BNC -> BNH(C/H) -> BHN(C/H)
v = self.wv(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
attn = (q @ k.transpose(-2, -1)) * self.scale # BH1(C/H) @ BH(C/H)N -> BH1N
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, 1, C) # (BH1N @ BHN(C/H)) -> BH1(C/H) -> B1H(C/H) -> B1C
x = self.proj(x)
x = self.proj_drop(x)
return x
class CrossAttentionBlock(nn.Module):
def __init__(
self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = CrossAttention(
dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def forward(self, x):
x = x[:, 0:1, ...] + self.drop_path(self.attn(self.norm1(x)))
return x
class MultiScaleBlock(nn.Module):
def __init__(self, dim, patches, depth, num_heads, mlp_ratio, qkv_bias=False, drop=0., attn_drop=0.,
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
num_branches = len(dim)
self.num_branches = num_branches
# different branch could have different embedding size, the first one is the base
self.blocks = nn.ModuleList()
for d in range(num_branches):
tmp = []
for i in range(depth[d]):
tmp.append(Block(
dim=dim[d], num_heads=num_heads[d], mlp_ratio=mlp_ratio[d], qkv_bias=qkv_bias,
drop=drop, attn_drop=attn_drop, drop_path=drop_path[i], norm_layer=norm_layer))
if len(tmp) != 0:
self.blocks.append(nn.Sequential(*tmp))
if len(self.blocks) == 0:
self.blocks = None
self.projs = nn.ModuleList()
for d in range(num_branches):
if dim[d] == dim[(d + 1) % num_branches] and False:
tmp = [nn.Identity()]
else:
tmp = [norm_layer(dim[d]), act_layer(), nn.Linear(dim[d], dim[(d + 1) % num_branches])]
self.projs.append(nn.Sequential(*tmp))
self.fusion = nn.ModuleList()
for d in range(num_branches):
d_ = (d + 1) % num_branches
nh = num_heads[d_]
if depth[-1] == 0: # backward capability:
self.fusion.append(
CrossAttentionBlock(
dim=dim[d_], num_heads=nh, mlp_ratio=mlp_ratio[d], qkv_bias=qkv_bias,
drop=drop, attn_drop=attn_drop, drop_path=drop_path[-1], norm_layer=norm_layer))
else:
tmp = []
for _ in range(depth[-1]):
tmp.append(CrossAttentionBlock(
dim=dim[d_], num_heads=nh, mlp_ratio=mlp_ratio[d], qkv_bias=qkv_bias,
drop=drop, attn_drop=attn_drop, drop_path=drop_path[-1], norm_layer=norm_layer))
self.fusion.append(nn.Sequential(*tmp))
self.revert_projs = nn.ModuleList()
for d in range(num_branches):
if dim[(d + 1) % num_branches] == dim[d] and False:
tmp = [nn.Identity()]
else:
tmp = [norm_layer(dim[(d + 1) % num_branches]), act_layer(),
nn.Linear(dim[(d + 1) % num_branches], dim[d])]
self.revert_projs.append(nn.Sequential(*tmp))
def forward(self, x: List[torch.Tensor]) -> List[torch.Tensor]:
outs_b = []
for i, block in enumerate(self.blocks):
outs_b.append(block(x[i]))
# only take the cls token out
proj_cls_token = torch.jit.annotate(List[torch.Tensor], [])
for i, proj in enumerate(self.projs):
proj_cls_token.append(proj(outs_b[i][:, 0:1, ...]))
# cross attention
outs = []
for i, (fusion, revert_proj) in enumerate(zip(self.fusion, self.revert_projs)):
tmp = torch.cat((proj_cls_token[i], outs_b[(i + 1) % self.num_branches][:, 1:, ...]), dim=1)
tmp = fusion(tmp)
reverted_proj_cls_token = revert_proj(tmp[:, 0:1, ...])
tmp = torch.cat((reverted_proj_cls_token, outs_b[i][:, 1:, ...]), dim=1)
outs.append(tmp)
return outs
def _compute_num_patches(img_size, patches):
return [i[0] // p * i[1] // p for i, p in zip(img_size, patches)]
@register_notrace_function
def scale_image(x, ss: Tuple[int, int], crop_scale: bool = False): # annotations for torchscript
"""
Pulled out of CrossViT.forward_features to bury conditional logic in a leaf node for FX tracing.
Args:
x (Tensor): input image
ss (tuple[int, int]): height and width to scale to
crop_scale (bool): whether to crop instead of interpolate to achieve the desired scale. Defaults to False
Returns:
Tensor: the "scaled" image batch tensor
"""
H, W = x.shape[-2:]
if H != ss[0] or W != ss[1]:
if crop_scale and ss[0] <= H and ss[1] <= W:
cu, cl = int(round((H - ss[0]) / 2.)), int(round((W - ss[1]) / 2.))
x = x[:, :, cu:cu + ss[0], cl:cl + ss[1]]
else:
x = torch.nn.functional.interpolate(x, size=ss, mode='bicubic', align_corners=False)
return x
class CrossViT(nn.Module):
""" Vision Transformer with support for patch or hybrid CNN input stage
"""
def __init__(
self, img_size=224, img_scale=(1.0, 1.0), patch_size=(8, 16), in_chans=3, num_classes=1000,
embed_dim=(192, 384), depth=((1, 3, 1), (1, 3, 1), (1, 3, 1)), num_heads=(6, 12), mlp_ratio=(2., 2., 4.),
multi_conv=False, crop_scale=False, qkv_bias=True, drop_rate=0., attn_drop_rate=0., drop_path_rate=0.,
norm_layer=partial(nn.LayerNorm, eps=1e-6), global_pool='token',
):
super().__init__()
assert global_pool in ('token', 'avg')
self.num_classes = num_classes
self.global_pool = global_pool
self.img_size = to_2tuple(img_size)
img_scale = to_2tuple(img_scale)
self.img_size_scaled = [tuple([int(sj * si) for sj in self.img_size]) for si in img_scale]
self.crop_scale = crop_scale # crop instead of interpolate for scale
num_patches = _compute_num_patches(self.img_size_scaled, patch_size)
self.num_branches = len(patch_size)
self.embed_dim = embed_dim
self.num_features = sum(embed_dim)
self.patch_embed = nn.ModuleList()
# hard-coded for torch jit script
for i in range(self.num_branches):
setattr(self, f'pos_embed_{i}', nn.Parameter(torch.zeros(1, 1 + num_patches[i], embed_dim[i])))
setattr(self, f'cls_token_{i}', nn.Parameter(torch.zeros(1, 1, embed_dim[i])))
for im_s, p, d in zip(self.img_size_scaled, patch_size, embed_dim):
self.patch_embed.append(
PatchEmbed(img_size=im_s, patch_size=p, in_chans=in_chans, embed_dim=d, multi_conv=multi_conv))
self.pos_drop = nn.Dropout(p=drop_rate)
total_depth = sum([sum(x[-2:]) for x in depth])
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, total_depth)] # stochastic depth decay rule
dpr_ptr = 0
self.blocks = nn.ModuleList()
for idx, block_cfg in enumerate(depth):
curr_depth = max(block_cfg[:-1]) + block_cfg[-1]
dpr_ = dpr[dpr_ptr:dpr_ptr + curr_depth]
blk = MultiScaleBlock(
embed_dim, num_patches, block_cfg, num_heads=num_heads, mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr_, norm_layer=norm_layer)
dpr_ptr += curr_depth
self.blocks.append(blk)
self.norm = nn.ModuleList([norm_layer(embed_dim[i]) for i in range(self.num_branches)])
self.head = nn.ModuleList([
nn.Linear(embed_dim[i], num_classes) if num_classes > 0 else nn.Identity()
for i in range(self.num_branches)])
for i in range(self.num_branches):
trunc_normal_(getattr(self, f'pos_embed_{i}'), std=.02)
trunc_normal_(getattr(self, f'cls_token_{i}'), std=.02)
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
@torch.jit.ignore
def no_weight_decay(self):
out = set()
for i in range(self.num_branches):
out.add(f'cls_token_{i}')
pe = getattr(self, f'pos_embed_{i}', None)
if pe is not None and pe.requires_grad:
out.add(f'pos_embed_{i}')
return out
@torch.jit.ignore
def group_matcher(self, coarse=False):
return dict(
stem=r'^cls_token|pos_embed|patch_embed', # stem and embed
blocks=[(r'^blocks\.(\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 ('token', 'avg')
self.global_pool = global_pool
self.head = nn.ModuleList(
[nn.Linear(self.embed_dim[i], num_classes) if num_classes > 0 else nn.Identity() for i in
range(self.num_branches)])
def forward_features(self, x) -> List[torch.Tensor]:
B = x.shape[0]
xs = []
for i, patch_embed in enumerate(self.patch_embed):
x_ = x
ss = self.img_size_scaled[i]
x_ = scale_image(x_, ss, self.crop_scale)
x_ = patch_embed(x_)
cls_tokens = self.cls_token_0 if i == 0 else self.cls_token_1 # hard-coded for torch jit script
cls_tokens = cls_tokens.expand(B, -1, -1)
x_ = torch.cat((cls_tokens, x_), dim=1)
pos_embed = self.pos_embed_0 if i == 0 else self.pos_embed_1 # hard-coded for torch jit script
x_ = x_ + pos_embed
x_ = self.pos_drop(x_)
xs.append(x_)
for i, blk in enumerate(self.blocks):
xs = blk(xs)
# NOTE: was before branch token section, move to here to assure all branch token are before layer norm
xs = [norm(xs[i]) for i, norm in enumerate(self.norm)]
return xs
def forward_head(self, xs: List[torch.Tensor], pre_logits: bool = False) -> torch.Tensor:
xs = [x[:, 1:].mean(dim=1) for x in xs] if self.global_pool == 'avg' else [x[:, 0] for x in xs]
if pre_logits or isinstance(self.head[0], nn.Identity):
return torch.cat([x for x in xs], dim=1)
return torch.mean(torch.stack([head(xs[i]) for i, head in enumerate(self.head)], dim=0), dim=0)
def forward(self, x):
xs = self.forward_features(x)
x = self.forward_head(xs)
return x
def _create_crossvit(variant, pretrained=False, **kwargs):
if kwargs.get('features_only', None):
raise RuntimeError('features_only not implemented for Vision Transformer models.')
def pretrained_filter_fn(state_dict):
new_state_dict = {}
for key in state_dict.keys():
if 'pos_embed' in key or 'cls_token' in key:
new_key = key.replace(".", "_")
else:
new_key = key
new_state_dict[new_key] = state_dict[key]
return new_state_dict
return build_model_with_cfg(
CrossViT, variant, pretrained,
pretrained_filter_fn=pretrained_filter_fn,
**kwargs)
@register_model
def crossvit_tiny_240(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[96, 192], depth=[[1, 4, 0], [1, 4, 0], [1, 4, 0]],
num_heads=[3, 3], mlp_ratio=[4, 4, 1], **kwargs)
model = _create_crossvit(variant='crossvit_tiny_240', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_small_240(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[192, 384], depth=[[1, 4, 0], [1, 4, 0], [1, 4, 0]],
num_heads=[6, 6], mlp_ratio=[4, 4, 1], **kwargs)
model = _create_crossvit(variant='crossvit_small_240', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_base_240(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[384, 768], depth=[[1, 4, 0], [1, 4, 0], [1, 4, 0]],
num_heads=[12, 12], mlp_ratio=[4, 4, 1], **kwargs)
model = _create_crossvit(variant='crossvit_base_240', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_9_240(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[128, 256], depth=[[1, 3, 0], [1, 3, 0], [1, 3, 0]],
num_heads=[4, 4], mlp_ratio=[3, 3, 1], **kwargs)
model = _create_crossvit(variant='crossvit_9_240', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_15_240(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[192, 384], depth=[[1, 5, 0], [1, 5, 0], [1, 5, 0]],
num_heads=[6, 6], mlp_ratio=[3, 3, 1], **kwargs)
model = _create_crossvit(variant='crossvit_15_240', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_18_240(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 224 / 240), patch_size=[12, 16], embed_dim=[224, 448], depth=[[1, 6, 0], [1, 6, 0], [1, 6, 0]],
num_heads=[7, 7], mlp_ratio=[3, 3, 1], **kwargs)
model = _create_crossvit(variant='crossvit_18_240', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_9_dagger_240(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 224 / 240), patch_size=[12, 16], embed_dim=[128, 256], depth=[[1, 3, 0], [1, 3, 0], [1, 3, 0]],
num_heads=[4, 4], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs)
model = _create_crossvit(variant='crossvit_9_dagger_240', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_15_dagger_240(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[192, 384], depth=[[1, 5, 0], [1, 5, 0], [1, 5, 0]],
num_heads=[6, 6], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs)
model = _create_crossvit(variant='crossvit_15_dagger_240', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_15_dagger_408(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 384/408), patch_size=[12, 16], embed_dim=[192, 384], depth=[[1, 5, 0], [1, 5, 0], [1, 5, 0]],
num_heads=[6, 6], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs)
model = _create_crossvit(variant='crossvit_15_dagger_408', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_18_dagger_240(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 224/240), patch_size=[12, 16], embed_dim=[224, 448], depth=[[1, 6, 0], [1, 6, 0], [1, 6, 0]],
num_heads=[7, 7], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs)
model = _create_crossvit(variant='crossvit_18_dagger_240', pretrained=pretrained, **model_args)
return model
@register_model
def crossvit_18_dagger_408(pretrained=False, **kwargs):
model_args = dict(
img_scale=(1.0, 384/408), patch_size=[12, 16], embed_dim=[224, 448], depth=[[1, 6, 0], [1, 6, 0], [1, 6, 0]],
num_heads=[7, 7], mlp_ratio=[3, 3, 1], multi_conv=True, **kwargs)
model = _create_crossvit(variant='crossvit_18_dagger_408', pretrained=pretrained, **model_args)
return model