mirrors/deit
1
0
Fork 0
mirror of https://github.com/facebookresearch/deit.git synced 2025-06-03 14:52:20 +08:00
Code Issues Packages Projects Releases Wiki Activity

add cait models and README

Browse Source
This commit is contained in:
Hugo Touvron 2021-04-09 16:45:17 +02:00 committed by GitHub
parent 7e0acbca7b
commit 4fcfe17ec3
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 459 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

64
README_cait.md Normal file
View File

@ -0,0 +1,64 @@
# CaiT: Going deeper with Image Transformers
This repository contains PyTorch evaluation code, training code and pretrained models for DeiT (Data-Efficient Image Transformers) and CaiT (Going deeper with Image Transformers). All models are trained during 400 epochs.
CaiT obtain competitive tradeoffs in terms of flops / precision:
<p align="center">
<img width="600" height="600" src=".github/cait.png">
</p>
For details see [Going deeper with Image Transformers](https://arxiv.org/abs/2103.17239) by Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve and Hervé Jégou
If you use this code for a paper please cite:
```
@article{touvron2021cait,
title={Going deeper with Image Transformers},
author={Hugo Touvron and Matthieu Cord and Alexandre Sablayrolles and Gabriel Synnaeve and Herv\'e J\'egou},
journal={arXiv preprint arXiv:2103.17239},
year={2021}
}
```
# Model Zoo
We provide baseline CaiT models pretrained on ImageNet1k 2012 only, using the distilled version of our method.
| name | res | acc@1 | FLOPs| #params | url |
| --- | --- | --- | --- | --- | --- |
| S24 | 83.5 | 224 |9.4B| 47M| [model](https://dl.fbaipublicfiles.com/deit/S24_224.pth) |
| XS24| 84.1 | 384 | 19.3B |27M | [model](https://dl.fbaipublicfiles.com/deit/XS24_384.pth) |
| S24 | 85.1 | 384 | 32.2B |47M | [model](https://dl.fbaipublicfiles.com/deit/S24_384.pth) |
| S36 | 85.4 | 384 | 48.0B| 68M| [model](https://dl.fbaipublicfiles.com/deit/S36_384.pth) |
| M36 | 86.1 | 384 | 173.3B| 271M | [model](https://dl.fbaipublicfiles.com/deit/M36_384.pth) |
| M48 | 86.5 | 448 | 329.6B| 356M | [model](https://dl.fbaipublicfiles.com/deit/M48_448.pth) |
The models are also available via torch hub.
Before using it, make sure you have the pytorch-image-models package [`timm==0.3.2`](https://github.com/rwightman/pytorch-image-models) by [Ross Wightman](https://github.com/rwightman) installed.
# Evaluation transforms
CaiT employs a slightly different pre-processing, in particular a crop-ratio of 1.0 at test time. To reproduce the results of our paper please use the following pre-processing:
```
def get_test_transforms(input_size):
mean, std = [0.485, 0.456, 0.406],[0.229, 0.224, 0.225]
transformations = {}
transformations= transforms.Compose(
[transforms.Resize(input_size, interpolation=3),
transforms.CenterCrop(input_size),
transforms.ToTensor(),
transforms.Normalize(mean, std)])
return transformations
```
Remark: for CaiT M48 it is best to evaluate with FP32 precision
# License
This repository is released under the Apache 2.0 license as found in the [LICENSE](LICENSE) file.
# Contributing
We actively welcome your pull requests! Please see [CONTRIBUTING.md](.github/CONTRIBUTING.md) and [CODE_OF_CONDUCT.md](.github/CODE_OF_CONDUCT.md) for more info.

395
cait_models.py Normal file
View File

@ -0,0 +1,395 @@
# Copyright (c) 2015-present, Facebook, Inc.
# All rights reserved.
import torch
import torch.nn as nn
from functools import partial
from timm.models.vision_transformer import Mlp, PatchEmbed , _cfg
from timm.models.registry import register_model
from timm.models.layers import trunc_normal_
__all__ = [
'cait_M48', 'cait_M36', 'cait_M4',
'cait_S36', 'cait_S24','cait_S24_224',
'cait_XS24'
]
class Class_Attention(nn.Module):
# taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
# with slight modifications to do CA
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim ** -0.5
self.q = nn.Linear(dim, dim, bias=qkv_bias)
self.k = nn.Linear(dim, dim, bias=qkv_bias)
self.v = nn.Linear(dim, dim, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x ):
B, N, C = x.shape
q = self.q(x[:,0]).unsqueeze(1).reshape(B, 1, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
k = self.k(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
q = q * self.scale
v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
attn = (q @ k.transpose(-2, -1))
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x_cls = (attn @ v).transpose(1, 2).reshape(B, 1, C)
x_cls = self.proj(x_cls)
x_cls = self.proj_drop(x_cls)
return x_cls
class LayerScale_Block_CA(nn.Module):
# taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
# with slight modifications to add CA and LayerScale
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, Attention_block = Class_Attention,
Mlp_block=Mlp,init_values=1e-4):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = Attention_block(
dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp_block(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
def forward(self, x, x_cls):
u = torch.cat((x_cls,x),dim=1)
x_cls = x_cls + self.drop_path(self.gamma_1 * self.attn(self.norm1(u)))
x_cls = x_cls + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x_cls)))
return x_cls
class Attention_talking_head(nn.Module):
# taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
# with slight modifications to add Talking Heads Attention (https://arxiv.org/pdf/2003.02436v1.pdf)
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim ** -0.5
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_l = nn.Linear(num_heads, num_heads)
self.proj_w = nn.Linear(num_heads, num_heads)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x):
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0] * self.scale , qkv[1], qkv[2]
attn = (q @ k.transpose(-2, -1))
attn = self.proj_l(attn.permute(0,2,3,1)).permute(0,3,1,2)
attn = attn.softmax(dim=-1)
attn = self.proj_w(attn.permute(0,2,3,1)).permute(0,3,1,2)
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_Block(nn.Module):
# taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
# with slight modifications to add layerScale
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,Attention_block = Attention_talking_head,
Mlp_block=Mlp,init_values=1e-4):
super().__init__()
self.norm1 = norm_layer(dim)
self.attn = Attention_block(
dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp_block(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
def forward(self, x):
x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x)))
x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
return x
class cait_models(nn.Module):
# taken from https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py
# with slight modifications to adapt to our cait models
def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
drop_path_rate=0., norm_layer=nn.LayerNorm, global_pool=None,
block_layers = LayerScale_Block,
block_layers_token = LayerScale_Block_CA,
Patch_layer=PatchEmbed,act_layer=nn.GELU,
Attention_block = Attention_talking_head,Mlp_block=Mlp,
init_scale=1e-4,
Attention_block_token_only=Class_Attention,
Mlp_block_token_only= Mlp,
depth_token_only=2,
mlp_ratio_clstk = 4.0):
super().__init__()
self.num_classes = num_classes
self.num_features = self.embed_dim = embed_dim
self.patch_embed = Patch_layer(
img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
num_patches = self.patch_embed.num_patches
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
self.pos_drop = nn.Dropout(p=drop_rate)
dpr = [drop_path_rate for i in range(depth)]
self.blocks = nn.ModuleList([
block_layers(
dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
act_layer=act_layer,Attention_block=Attention_block,Mlp_block=Mlp_block,init_values=init_scale)
for i in range(depth)])
self.blocks_token_only = nn.ModuleList([
block_layers_token(
dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio_clstk, qkv_bias=qkv_bias, qk_scale=qk_scale,
drop=0.0, attn_drop=0.0, drop_path=0.0, norm_layer=norm_layer,
act_layer=act_layer,Attention_block=Attention_block_token_only,
Mlp_block=Mlp_block_token_only,init_values=init_scale)
for i in range(depth_token_only)])
self.norm = norm_layer(embed_dim)
self.feature_info = [dict(num_chs=embed_dim, reduction=0, module='head')]
self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()
trunc_normal_(self.pos_embed, std=.02)
trunc_normal_(self.cls_token, 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):
return {'pos_embed', 'cls_token'}
def forward_features(self, x):
B = x.shape[0]
x = self.patch_embed(x)
cls_tokens = self.cls_token.expand(B, -1, -1)
x = x + self.pos_embed
x = self.pos_drop(x)
for i , blk in enumerate(self.blocks):
x = blk(x)
for i , blk in enumerate(self.blocks_token_only):
cls_tokens = blk(x,cls_tokens)
x = torch.cat((cls_tokens, x), dim=1)
x = self.norm(x)
return x[:, 0]
def forward(self, x):
x = self.forward_features(x)
x = self.head(x)
return x
@register_model
def cait_XS24(pretrained=False, **kwargs):
model = cait_models(
img_size= 384,patch_size=16, embed_dim=288, depth=24, num_heads=6, mlp_ratio=4, qkv_bias=True,
norm_layer=partial(nn.LayerNorm, eps=1e-6),
init_scale=1e-5,
depth_token_only=2,**kwargs)
model.default_cfg = _cfg()
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/deit/XS24_384.pth",
map_location="cpu", check_hash=True
)
checkpoint_no_module = {}
for k in model.state_dict().keys():
checkpoint_no_module[k] = checkpoint["model"]['module.'+k]
model.load_state_dict(checkpoint_no_module)
return model
@register_model
def cait_S24_224(pretrained=False, **kwargs):
model = cait_models(
img_size= 224,patch_size=16, embed_dim=384, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True,
norm_layer=partial(nn.LayerNorm, eps=1e-6),
init_scale=1e-5,
depth_token_only=2,**kwargs)
model.default_cfg = _cfg()
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/deit/S24_224.pth",
map_location="cpu", check_hash=True
)
checkpoint_no_module = {}
for k in model.state_dict().keys():
checkpoint_no_module[k] = checkpoint["model"]['module.'+k]
model.load_state_dict(checkpoint_no_module)
return model
@register_model
def cait_S24(pretrained=False, **kwargs):
model = cait_models(
img_size= 384,patch_size=16, embed_dim=384, depth=24, num_heads=8, mlp_ratio=4, qkv_bias=True,
norm_layer=partial(nn.LayerNorm, eps=1e-6),
init_scale=1e-5,
depth_token_only=2,**kwargs)
model.default_cfg = _cfg()
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/deit/S24_384.pth",
map_location="cpu", check_hash=True
)
checkpoint_no_module = {}
for k in model.state_dict().keys():
checkpoint_no_module[k] = checkpoint["model"]['module.'+k]
model.load_state_dict(checkpoint_no_module)
return model
@register_model
def cait_S36(pretrained=False, **kwargs):
model = cait_models(
img_size= 384,patch_size=16, embed_dim=384, depth=36, num_heads=8, mlp_ratio=4, qkv_bias=True,
norm_layer=partial(nn.LayerNorm, eps=1e-6),
dpr_constant=True,
init_scale=1e-6,
depth_token_only=2,**kwargs)
model.default_cfg = _cfg()
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/deit/S36_384.pth",
map_location="cpu", check_hash=True
)
checkpoint_no_module = {}
for k in model.state_dict().keys():
checkpoint_no_module[k] = checkpoint["model"]['module.'+k]
model.load_state_dict(checkpoint_no_module)
return model
@register_model
def cait_M36(pretrained=False, **kwargs):
model = cait_models(
img_size= 384, patch_size=16, embed_dim=768, depth=36, num_heads=16, mlp_ratio=4, qkv_bias=True,
norm_layer=partial(nn.LayerNorm, eps=1e-6),
init_scale=1e-6,
depth_token_only=2,**kwargs)
model.default_cfg = _cfg()
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/deit/M36_384.pth",
map_location="cpu", check_hash=True
)
checkpoint_no_module = {}
for k in model.state_dict().keys():
checkpoint_no_module[k] = checkpoint["model"]['module.'+k]
model.load_state_dict(checkpoint_no_module)
return model
@register_model
def cait_M48(pretrained=False, **kwargs):
model = cait_models(
img_size= 448 , patch_size=16, embed_dim=768, depth=48, num_heads=16, mlp_ratio=4, qkv_bias=True,
norm_layer=partial(nn.LayerNorm, eps=1e-6),
init_scale=1e-6,
depth_token_only=2,**kwargs)
model.default_cfg = _cfg()
if pretrained:
checkpoint = torch.hub.load_state_dict_from_url(
url="https://dl.fbaipublicfiles.com/deit/M48_448.pth",
map_location="cpu", check_hash=True
)
checkpoint_no_module = {}
for k in model.state_dict().keys():
checkpoint_no_module[k] = checkpoint["model"]['module.'+k]
model.load_state_dict(checkpoint_no_module)
return model