Spaces:
Sleeping
Sleeping
File size: 7,258 Bytes
00613e2 |
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 |
import torch.nn as nn
import torch
import torch.nn.functional as F
import copy
class MlpProjector(nn.Module):
def __init__(self, cfg):
super().__init__()
self.cfg = cfg
if cfg.projector_type == "identity":
modules = nn.Identity()
elif cfg.projector_type == "linear":
modules = nn.Linear(cfg.input_dim, cfg.n_embed)
elif cfg.projector_type == "mlp_gelu":
mlp_depth = cfg.get("depth", 1)
modules = [nn.Linear(cfg.input_dim, cfg.n_embed)]
for _ in range(1, mlp_depth):
modules.append(nn.GELU())
modules.append(nn.Linear(cfg.n_embed, cfg.n_embed))
modules = nn.Sequential(*modules)
elif cfg.projector_type == "normlayer_downsample_mlp_gelu":
mlp_depth = cfg.get("depth", 1)
mlp_ratio = cfg.get("mlp_ratio", 1)
modules = [
nn.LayerNorm(cfg.input_dim * cfg.downsample_ratio * cfg.downsample_ratio),
nn.Linear(cfg.input_dim * cfg.downsample_ratio * cfg.downsample_ratio, cfg.n_embed * mlp_ratio)
]
for _ in range(1, mlp_depth - 1):
modules.append(nn.GELU())
modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed * mlp_ratio))
modules.append(nn.GELU())
modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed))
modules = nn.Sequential(*modules)
elif cfg.projector_type == "downsample_mlp_gelu":
mlp_depth = cfg.get("depth", 1)
mlp_ratio = cfg.get("mlp_ratio", 1)
modules = [nn.Linear(cfg.input_dim * cfg.downsample_ratio * cfg.downsample_ratio, cfg.n_embed * mlp_ratio)]
for _ in range(1, mlp_depth - 1):
modules.append(nn.GELU())
modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed * mlp_ratio))
modules.append(nn.GELU())
modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed))
modules = nn.Sequential(*modules)
elif cfg.projector_type == "low_high_hybrid_split_mlp_gelu":
mlp_depth = cfg.get("depth", 1)
self.high_up_proj = nn.Linear(cfg.input_dim, cfg.n_embed // 2)
self.low_up_proj = nn.Linear(cfg.input_dim, cfg.n_embed // 2)
modules = []
for _ in range(1, mlp_depth):
modules.append(nn.GELU())
modules.append(nn.Linear(cfg.n_embed, cfg.n_embed))
modules = nn.Sequential(*modules)
elif cfg.projector_type == "hybrid_split_feature_mlp_gelu":
mlp_depth = cfg.get("depth", 1)
channel_div = cfg.get("channel_div", 0.5)
self.high_up_proj = nn.Linear(cfg.input_dim[0], int(cfg.n_embed * channel_div))
self.low_up_proj = nn.Linear(cfg.input_dim[1], cfg.n_embed - int(cfg.n_embed * channel_div))
modules = []
for _ in range(1, mlp_depth):
modules.append(nn.GELU())
modules.append(nn.Linear(cfg.n_embed, cfg.n_embed))
modules = nn.Sequential(*modules)
elif cfg.projector_type == "low_high_split_mlp_gelu":
mlp_depth = cfg.get("depth", 1)
modules = []
for _ in range(1, mlp_depth):
modules.append(nn.GELU())
modules.append(nn.Linear(cfg.n_embed // 2, cfg.n_embed // 2))
modules = nn.Sequential(*modules)
self.high_layers = nn.Sequential(*modules)
self.low_layers = copy.deepcopy(modules)
else:
raise ValueError(f"Unknown projector type: {cfg.projector_type}")
if cfg.get("token_pooling", False):
self.token_pooling_layer = nn.Linear(cfg.input_dim * 4, cfg.input_dim)
if cfg.get("conv_fusion_high_low_features", False):
self.fusion_layer = nn.Linear(cfg.input_dim, cfg.input_dim)
self.layers = modules
def forward(self, x):
if self.cfg.get("token_pooling", False):
batch_size, wxh, channels = x.shape
w = h = int(wxh**0.5)
x = x.view(batch_size, w, h, channels)
x = x.permute(0, 3, 1, 2)
# import ipdb; ipdb.set_trace()
patches = x.unfold(2, 2, 2).unfold(3, 2, 2)
batch_size, channels, h_patches, w_patches, _, _ = patches.size()
# 在通道维度上拼接
patches = patches.contiguous().view(batch_size, channels, h_patches * w_patches, -1)
# 通过线性层
patches = patches.permute(0, 2, 1, 3).contiguous()
patches = patches.view(batch_size, h_patches * w_patches, channels * 4)
x = self.token_pooling_layer(patches)
if self.cfg.get("conv_fusion_high_low_features", False):
x = self.fusion_layer(x[:, 0]) + x[:, 1]
if self.cfg.projector_type == 'low_high_hybrid_split_mlp_gelu':
high_x, low_x = x[0], x[1]
high_x = self.high_up_proj(high_x)
low_x = self.low_up_proj(low_x)
x = torch.concat([high_x, low_x], dim=-1)
if self.cfg.projector_type == 'hybrid_split_feature_mlp_gelu':
high_x = x[...,:self.cfg.input_dim[0]]
low_x = x[...,self.cfg.input_dim[0]:]
high_x = self.high_up_proj(high_x)
low_x = self.low_up_proj(low_x)
x = torch.concat([high_x, low_x], dim=-1)
if self.cfg.projector_type == 'low_high_split_mlp_gelu':
high_x, low_x = x[0], x[1]
high_x = self.high_layers(high_x)
low_x = self.low_layers(low_x)
x = torch.concat([high_x, low_x], dim=-1)
return x
if self.cfg.projector_type == 'downsample_mlp_gelu' or self.cfg.projector_type == 'normlayer_downsample_mlp_gelu':
bs, hw, input_dim = x.shape
h = w = int((hw) ** 0.5)
"""compute padding"""
if h % self.cfg.downsample_ratio:
pad = self.cfg.downsample_ratio - h % self.cfg.downsample_ratio
else:
pad = 0
x = x.reshape(bs, h, w, input_dim)
if pad > 0:
x = F.pad(x, (0, 0, 0, pad, 0, pad), "constant", 0)
"""4 to 1 concat"""
x = x.permute(0, 3, 1, 2) # B, C, H, W
x = F.unfold(x, kernel_size=self.cfg.downsample_ratio, stride=self.cfg.downsample_ratio, padding=0) # B, C*4, HW // 4
x = x.permute(0, 2, 1)
return self.layers(x)
@staticmethod
def get_flops_per_sample(cfg):
if cfg.projector_type == "linear":
fwd = 2 * cfg.input_dim * cfg.n_embed
elif "mlp_gelu" in cfg.projector_type :
mlp_depth = cfg.get("depth", 1)
downsample_ratio = cfg.get("downsample_ratio", 1)
input_dim = sum(cfg.input_dim) if isinstance(cfg.input_dim, list) else cfg.input_dim
input_dim = input_dim * downsample_ratio * downsample_ratio
fwd = 2 * input_dim * cfg.n_embed + (mlp_depth - 1) * 2 * cfg.n_embed * cfg.n_embed
else:
fwd = 0
return fwd * 3
|