Spaces:
Runtime error
Runtime error
File size: 10,089 Bytes
6163604 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 | import torch.nn as nn
import torch
from torch.nn import functional as F
from .vit.utils import trunc_normal_
from .vit.vision_transformer import VisionTransformer
from ..feature_extractor.clova_impl import ResNet_FeatureExtractor
from .addon_module import *
from ..common.mae_posembed import get_2d_sincos_pos_embed
__all__ = ['ViTEncoder', 'ViTEncoderV2', 'ViTEncoderV3', 'TRIGBaseEncoder', 'create_vit_modeling']
class ViTEncoder(VisionTransformer):
'''
'''
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
if kwargs['hybrid_backbone'] is None:
self.patch_embed = PatchEmbed(
img_size=kwargs['img_size'],
in_chans=kwargs['in_chans'],
patch_size=kwargs['patch_size'],
embed_dim=kwargs['embed_dim'],
)
else:
self.patch_embed = HybridEmbed(
backbone=kwargs['hybrid_backbone'],
img_size=kwargs['img_size'],
in_chans=kwargs['in_chans'],
patch_size=kwargs['patch_size'],
embed_dim=kwargs['embed_dim'],
)
num_patches = self.patch_embed.num_patches
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches+1, kwargs['embed_dim']))
self.emb_height = self.patch_embed.grid_size[0]
self.emb_width = self.patch_embed.grid_size[1]
trunc_normal_(self.pos_embed, std=.02)
self.apply(self._init_weights)
def reset_classifier(self, num_classes):
self.num_classes = num_classes
self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
def interpolating_pos_embedding(self, embedding, height, width):
"""
Source:
https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
"""
npatch = embedding.shape[1] - 1
N = self.pos_embed.shape[1] - 1
if npatch == N and height == width:
return self.pos_embed
class_pos_embedding = self.pos_embed[:, 0]
patch_pos_embedding = self.pos_embed[:, 1:]
dim = self.pos_embed.shape[-1]
h0 = height // self.patch_embed.patch_size[0]
w0 = width // self.patch_embed.patch_size[1]
#add a small number to avo_id floating point error
# https://github.com/facebookresearch/dino/issues/8
h0 = h0 + 0.1
w0 = w0 + 0.1
patch_pos_embedding = nn.functional.interpolate(
patch_pos_embedding.reshape(1, self.emb_height, self.emb_width, dim).permute(0, 3, 1, 2),
scale_factor=(h0 / self.emb_height, w0 / self.emb_width),
mode='bicubic',
align_corners=False
)
assert int(h0) == patch_pos_embedding.shape[-2] and int(w0) == patch_pos_embedding.shape[-1]
patch_pos_embedding = patch_pos_embedding.permute(0, 2, 3, 1).view(1, -1, dim)
class_pos_embedding = class_pos_embedding.unsqueeze(0)
return torch.cat((class_pos_embedding, patch_pos_embedding), dim=1)
def forward_features(self, x):
B, C, _, _ = x.shape
x, pad_info, size, interpolating_pos = self.patch_embed(x)
cls_tokens = self.cls_token.expand(B, -1, -1) # stole cls_tokens impl from Phil Wang, thanks
x = torch.cat((cls_tokens, x), dim=1)
if interpolating_pos:
x = x + self.interpolating_pos_embedding(x, size['height'], size['width'])
else:
x = x + self.pos_embed
x = self.pos_drop(x)
for blk in self.blocks:
x = blk(x)
x = self.norm(x)
return x, pad_info, size
class TRIGBaseEncoder(ViTEncoder):
'''
https://arxiv.org/pdf/2111.08314.pdf
'''
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.patch_embed = HybridEmbed1D(
backbone=kwargs['hybrid_backbone'],
img_size=kwargs['img_size'],
in_chans=kwargs['in_chans'],
patch_size=kwargs['patch_size'],
embed_dim=kwargs['embed_dim'],
)
num_patches = self.patch_embed.num_patches
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches+1, kwargs['embed_dim']))
self.emb_height = 1
self.emb_width = self.patch_embed.grid_size[1]
trunc_normal_(self.pos_embed, std=.02)
self.apply(self._init_weights)
def interpolating_pos_embedding(self, embedding, height, width):
"""
Source:
https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
"""
npatch = embedding.shape[1] - 1
N = self.pos_embed.shape[1] - 1
if npatch == N and height == width:
return self.pos_embed
class_pos_embedding = self.pos_embed[:, 0]
patch_pos_embedding = self.pos_embed[:, 1:]
dim = self.pos_embed.shape[-1]
w0 = width // self.patch_embed.window_width
#add a small number to avoid floating point error
# https://github.com/facebookresearch/dino/issues/8
w0 = w0 + 0.1
patch_pos_embedding = nn.functional.interpolate(
patch_pos_embedding.reshape(1, self.emb_height, self.emb_width, dim).permute(0, 3, 1, 2),
scale_factor=(1, w0 / self.emb_width),
mode='bicubic',
align_corners=False
)
assert int(w0) == patch_pos_embedding.shape[-1]
patch_pos_embedding = patch_pos_embedding.permute(0, 2, 3, 1).view(1, -1, dim)
class_pos_embedding = class_pos_embedding.unsqueeze(0)
return torch.cat((class_pos_embedding, patch_pos_embedding), dim=1)
def forward_features(self, x):
B, _, _, _ = x.shape
x, padinfo, size, interpolating_pos = self.patch_embed(x)
cls_tokens = self.cls_token.expand(B, -1, -1) # stole cls_tokens impl from Phil Wang, thanks
x = torch.cat((cls_tokens, x), dim=1) #cls_tokens is init_embedding in TRIG paper
if interpolating_pos:
x = x + self.interpolating_pos_embedding(x, size['height'], size['width'])
else:
x = x + self.pos_embed
x = self.pos_drop(x)
for blk in self.blocks:
x = blk(x)
x = self.norm(x)
return x, padinfo, size
class ViTEncoderV2(ViTEncoder):
def forward(self, x):
B, _, _, _ = x.shape
x, pad_info, size, _ = self.patch_embed(x)
_, numpatches, *_ = x.shape
cls_tokens = self.cls_token.expand(B, -1, -1) # stole cls_tokens impl from Phil Wang, thanks
x = torch.cat((cls_tokens, x), dim=1)
x = x + self.pos_embed[:, :(numpatches + 1)]
x = self.pos_drop(x)
for blk in self.blocks:
x = blk(x)
x = self.norm(x)
return x, pad_info, size
class ViTEncoderV3(ViTEncoder):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
if hasattr(self, 'pos_embed'):
del self.pos_embed
num_patches = self.patch_embed.num_patches
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches+1, kwargs['embed_dim']), requires_grad=False)
self.initialize_posembed()
def initialize_posembed(self):
pos_embed = get_2d_sincos_pos_embed(
self.pos_embed.shape[-1],
self.patch_embed.grid_size[0],
self.patch_embed.grid_size[1],
cls_token=True
)
self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
def forward(self, x):
B, _, _, _ = x.shape
x, pad_info, size, _ = self.patch_embed(x)
_, numpatches, *_ = x.shape
cls_tokens = self.cls_token.expand(B, -1, -1) # stole cls_tokens impl from Phil Wang, thanks
x = torch.cat((cls_tokens, x), dim=1)
x = x + self.pos_embed[:, :(numpatches + 1)]
x = self.pos_drop(x)
for blk in self.blocks:
x = blk(x)
x = self.norm(x)
return x, pad_info, size
def create_vit_modeling(opt):
seq_modeling = opt['SequenceModeling']['params']
if seq_modeling['backbone'] is not None:
if seq_modeling['backbone']['name'] == 'resnet':
param_kwargs = dict()
if seq_modeling['backbone'].get('pretrained', None) is not None:
param_kwargs['pretrained'] = seq_modeling['backbone']['pretrained']
if seq_modeling['backbone'].get('weight_dir', None) is not None:
param_kwargs['weight_dir'] = seq_modeling['backbone']['weight_dir']
print('kwargs', param_kwargs)
backbone = ResNet_FeatureExtractor(
seq_modeling['backbone']['input_channel'],
seq_modeling['backbone']['output_channel'],
seq_modeling['backbone']['gcb'],
**param_kwargs
)
elif seq_modeling['backbone']['name'] == 'cnn':
backbone = None
else: backbone = None
max_dimension = (opt['imgH'], opt['max_dimension'][1]) if opt['imgH'] else opt['max_dimension']
if seq_modeling['patching_style'] == '2d':
if seq_modeling.get('fix_embed', False):
encoder = ViTEncoderV3
else:
if not seq_modeling.get('interpolate_embed', True):
encoder = ViTEncoderV2
else:
encoder = ViTEncoder
else:
encoder = TRIGBaseEncoder
encoder_seq_modeling = encoder(
img_size=max_dimension,
patch_size=seq_modeling['patch_size'],
in_chans=seq_modeling['input_channel'],
depth=seq_modeling['depth'],
num_classes=0,
embed_dim=seq_modeling['hidden_size'],
num_heads=seq_modeling['num_heads'],
hybrid_backbone=backbone
)
return encoder_seq_modeling
|