pytorch实现 vision_transformer
Posted AI浩
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了pytorch实现 vision_transformer相关的知识，希望对你有一定的参考价值。
github地址：https://github.com/rwightman/pytorch-image-models/
记录一下模型，方便查阅
""" Vision Transformer (ViT) in PyTorch
A PyTorch implement of Vision Transformers as described in:
'An Image Is Worth 16 x 16 Words: Transformers for Image Recognition at Scale'
- https://arxiv.org/abs/2010.11929
`How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers`
- https://arxiv.org/abs/2106.10270
The official jax code is released and available at https://github.com/google-research/vision_transformer
DeiT model defs and weights from https://github.com/facebookresearch/deit,
paper `DeiT: Data-efficient Image Transformers` - https://arxiv.org/abs/2012.12877
Acknowledgments:
* The paper authors for releasing code and weights, thanks!
* I fixed my class token impl based on Phil Wang's https://github.com/lucidrains/vit-pytorch ... check it out
for some einops/einsum fun
* Simple transformer style inspired by Andrej Karpathy's https://github.com/karpathy/minGPT
* Bert reference code checks against Huggingface Transformers and Tensorflow Bert
Hacked together by / Copyright 2020, Ross Wightman
"""
import math
import logging
from functools import partial
from collections import OrderedDict
from copy import deepcopy
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD, IMAGENET_INCEPTION_MEAN, IMAGENET_INCEPTION_STD
from models.weight_init import trunc_normal_, lecun_normal_
_logger = logging.getLogger(__name__)
def adapt_input_conv(in_chans, conv_weight):
conv_type = conv_weight.dtype
conv_weight = conv_weight.float() # Some weights are in torch.half, ensure it's float for sum on CPU
O, I, J, K = conv_weight.shape
if in_chans == 1:
if I > 3:
assert conv_weight.shape[1] % 3 == 0
# For models with space2depth stems
conv_weight = conv_weight.reshape(O, I // 3, 3, J, K)
conv_weight = conv_weight.sum(dim=2, keepdim=False)
else:
conv_weight = conv_weight.sum(dim=1, keepdim=True)
elif in_chans != 3:
if I != 3:
raise NotImplementedError('Weight format not supported by conversion.')
else:
# NOTE this strategy should be better than random init, but there could be other combinations of
# the original RGB input layer weights that'd work better for specific cases.
repeat = int(math.ceil(in_chans / 3))
conv_weight = conv_weight.repeat(1, repeat, 1, 1)[:, :in_chans, :, :]
conv_weight *= (3 / float(in_chans))
conv_weight = conv_weight.to(conv_type)
return conv_weight
def drop_path(x, drop_prob: float = 0., training: bool = False):
"""
Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
'survival rate' as the argument.
"""
if drop_prob == 0. or not training:
return x
keep_prob = 1 - drop_prob
shape = (x.shape[0],) + (1,) * (x.ndim - 1) # work with diff dim tensors, not just 2D ConvNets
random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
random_tensor.floor_() # binarize
output = x.div(keep_prob) * random_tensor
return output
class DropPath(nn.Module):
"""
Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
"""
def __init__(self, drop_prob=None):
super(DropPath, self).__init__()
self.drop_prob = drop_prob
def forward(self, x):
return drop_path(x, self.drop_prob, self.training)
class PatchEmbed(nn.Module):
"""
2D Image to Patch Embedding
"""
def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None):
super().__init__()
img_size = (img_size, img_size)
patch_size = (patch_size, patch_size)
self.img_size = img_size
self.patch_size = patch_size
self.grid_size = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
self.num_patches = self.grid_size[0] * self.grid_size[1]
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
def forward(self, x):
B, C, H, W = x.shape
assert H == self.img_size[0] and W == self.img_size[1], \\
f"Input image size (H*W) doesn't match model (self.img_size[0]*self.img_size[1])."
# flatten: [B, C, H, W] -> [B, C, HW]
# transpose: [B, C, HW] -> [B, HW, C]
x = self.proj(x).flatten(2).transpose(1, 2)
x = self.norm(x)
return x
class Attention(nn.Module):
def __init__(self,
dim, # 输入token的dim
num_heads=8,
qkv_bias=False,
qk_scale=None,
attn_drop_ratio=0.,
proj_drop_ratio=0.):
super(Attention, self).__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_ratio)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop_ratio)
def forward(self, x):
# [batch_size, num_patches + 1, total_embed_dim]
B, N, C = x.shape
# qkv(): -> [batch_size, num_patches + 1, 3 * total_embed_dim]
# reshape: -> [batch_size, num_patches + 1, 3, num_heads, embed_dim_per_head]
# permute: -> [3, batch_size, num_heads, num_patches + 1, embed_dim_per_head]
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
# [batch_size, num_heads, num_patches + 1, embed_dim_per_head]
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
# transpose: -> [batch_size, num_heads, embed_dim_per_head, num_patches + 1]
# @: multiply -> [batch_size, num_heads, num_patches + 1, num_patches + 1]
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
# @: multiply -> [batch_size, num_heads, num_patches + 1, embed_dim_per_head]
# transpose: -> [batch_size, num_patches + 1, num_heads, embed_dim_per_head]
# reshape: -> [batch_size, num_patches + 1, total_embed_dim]
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class Mlp(nn.Module):
"""
MLP as used in Vision Transformer, MLP-Mixer and related networks
"""
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x
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 =
# patch models (weights from official Google JAX impl)
'vit_tiny_patch16_224': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_224.npz'),
'vit_tiny_patch16_384': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz',
input_size=(3, 384, 384), crop_pct=1.0),
'vit_small_patch32_224': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'S_32-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_224.npz'),
'vit_small_patch32_384': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'S_32-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz',
input_size=(3, 384, 384), crop_pct=1.0),
'vit_small_patch16_224': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'S_16-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_224.npz'),
'vit_small_patch16_384': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'S_16-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz',
input_size=(3, 384, 384), crop_pct=1.0),
'vit_base_patch32_224': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'B_32-i21k-300ep-lr_0.001-aug_medium1-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_224.npz'),
'vit_base_patch32_384': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'B_32-i21k-300ep-lr_0.001-aug_light1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz',
input_size=(3, 384, 384), crop_pct=1.0),
'vit_base_patch16_224': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_224.npz'),
'vit_base_patch16_384': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_384.npz',
input_size=(3, 384, 384), crop_pct=1.0),
'vit_base_patch8_224': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'B_8-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_224.npz'),
'vit_large_patch32_224': _cfg(
url='', # no official model weights for this combo, only for in21k
),
'vit_large_patch32_384': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_p32_384-9b920ba8.pth',
input_size=(3, 384, 384), crop_pct=1.0),
'vit_large_patch16_224': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'L_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.1-sd_0.1--imagenet2012-steps_20k-lr_0.01-res_224.npz'),
'vit_large_patch16_384': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/'
'L_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.1-sd_0.1--imagenet2012-steps_20k-lr_0.01-res_384.npz',
input_size=(3, 384, 384), crop_pct=1.0),
'vit_huge_patch14_224': _cfg(url=''),
'vit_giant_patch14_224': _cfg(url=''),
'vit_gigantic_patch14_224': _cfg(url=''),
'vit_base2_patch32_256': _cfg(url='', input_size=(3, 256, 256), crop_pct=0.95),
# patch models, imagenet21k (weights from official Google JAX impl)
'vit_tiny_patch16_224_in21k': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0.npz',
num_classes=21843),
'vit_small_patch32_224_in21k': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/S_32-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0.npz',
num_classes=21843),
'vit_small_patch16_224_in21k': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/S_16-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0.npz',
num_classes=21843),
'vit_base_patch32_224_in21k': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/B_32-i21k-300ep-lr_0.001-aug_medium1-wd_0.03-do_0.0-sd_0.0.npz',
num_classes=21843),
'vit_base_patch16_224_in21k': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0.npz',
num_classes=21843),
'vit_base_patch8_224_in21k': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/B_8-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0.npz',
num_classes=21843),
'vit_large_patch32_224_in21k': _cfg(
url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_large_patch32_224_in21k-9046d2e7.pth',
num_classes=21843),
'vit_large_patch16_224_in21k': _cfg(
url='https://storage.googleapis.com/vit_models/augreg/L_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.1-sd_0.1.npz',
num_classes=21843),
'vit_huge_patch14_224_in21k': _cfg(
url='https://storage.googleapis.com/vit_models/imagenet21k/ViT-H_14.npz',
hf_hub='timm/vit_huge_patch14_224_in21k',
num_classes=21843),
# SAM trained models (https://arxiv.org/abs/2106.01548)
'vit_base_patch32_224_sam': _cfg(
url='https://storage.googleapis.com/vit_models/sam/ViT-B_32.npz'),
'vit_base_patch16_224_sam': _cfg(
url='https://storage.googleapis.com/vit_models/sam/ViT-B_16.npz'),
# DINO pretrained - https://arxiv.org/abs/2104.14294 (no classifier head, for fine-tune only)
'vit_small_patch16_224_dino': _cfg(
url='https://dl.fbaipublicfiles.com/dino/dino_deitsmall16_pretrain/dino_deitsmall16_pretrain.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD, num_classes=0),
'vit_small_patch8_224_dino': _cfg(
url='https://dl.fbaipublicfiles.com/dino/dino_deitsmall8_pretrain/dino_deitsmall8_pretrain.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD, num_classes=0),
'vit_base_patch16_224_dino': _cfg(
url='https://dl.fbaipublicfiles.com/dino/dino_vitbase16_pretrain/dino_vitbase16_pretrain.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD, num_classes=0),
'vit_base_patch8_224_dino': _cfg(
url='https://dl.fbaipublicfiles.com/dino/dino_vitbase8_pretrain/dino_vitbase8_pretrain.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD, num_classes=0),
# deit models (FB weights)
'deit_tiny_patch16_224': _cfg(
url='https://dl.fbaipublicfiles.com/deit/deit_tiny_patch16_224-a1311bcf.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD),
'deit_small_patch16_224': _cfg(
url='https://dl.fbaipublicfiles.com/deit/deit_small_patch16_224-cd65a155.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD),
'deit_base_patch16_224': _cfg(
url='https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD),
'deit_base_patch16_384': _cfg(
url='https://dl.fbaipublicfiles.com/deit/deit_base_patch16_384-8de9b5d1.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD, input_size=(3, 384, 384), crop_pct=1.0),
'deit_tiny_distilled_patch16_224': _cfg(
url='https://dl.fbaipublicfiles.com/deit/deit_tiny_distilled_patch16_224-b40b3cf7.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD, classifier=('head', 'head_dist')),
'deit_small_distilled_patch16_224': _cfg(
url='https://dl.fbaipublicfiles.com/deit/deit_small_distilled_patch16_224-649709d9.pth',
mean=IMAGENET_DEFAULT_MEAN, std=IMAGENET_DEFAULT_STD, classifier=('head', 'head_dist')),
'deit_base_distilled_patch16_224': _cfgPytorch实现GAT(基于PyTorch实现)
教程 | PyTorch内部机制解析：如何通过PyTorch实现Tensor