Spaces:
Sleeping
Sleeping
File size: 6,203 Bytes
297dd64 | 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 | """
DenseNet encoder from ICAL.
Copied from: https://github.com/qingzhenduyu/ICAL
"""
import math
from typing import Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
from torch import FloatTensor, LongTensor
from .pos_enc import ImgPosEnc
# DenseNet-B
class _Bottleneck(nn.Module):
def __init__(self, n_channels: int, growth_rate: int, use_dropout: bool):
super(_Bottleneck, self).__init__()
interChannels = 4 * growth_rate
self.bn1 = nn.BatchNorm2d(interChannels)
self.conv1 = nn.Conv2d(n_channels, interChannels,
kernel_size=1, bias=False)
self.bn2 = nn.BatchNorm2d(growth_rate)
self.conv2 = nn.Conv2d(
interChannels, growth_rate, kernel_size=3, padding=1, bias=False
)
self.use_dropout = use_dropout
self.dropout = nn.Dropout(p=0.2)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)), inplace=True)
if self.use_dropout:
out = self.dropout(out)
out = F.relu(self.bn2(self.conv2(out)), inplace=True)
if self.use_dropout:
out = self.dropout(out)
out = torch.cat((x, out), 1)
return out
# single layer
class _SingleLayer(nn.Module):
def __init__(self, n_channels: int, growth_rate: int, use_dropout: bool):
super(_SingleLayer, self).__init__()
self.bn1 = nn.BatchNorm2d(n_channels)
self.conv1 = nn.Conv2d(
n_channels, growth_rate, kernel_size=3, padding=1, bias=False
)
self.use_dropout = use_dropout
self.dropout = nn.Dropout(p=0.2)
def forward(self, x):
out = self.conv1(F.relu(x, inplace=True))
if self.use_dropout:
out = self.dropout(out)
out = torch.cat((x, out), 1)
return out
# transition layer
class _Transition(nn.Module):
def __init__(self, n_channels: int, n_out_channels: int, use_dropout: bool):
super(_Transition, self).__init__()
self.bn1 = nn.BatchNorm2d(n_out_channels)
self.conv1 = nn.Conv2d(n_channels, n_out_channels,
kernel_size=1, bias=False)
self.use_dropout = use_dropout
self.dropout = nn.Dropout(p=0.2)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)), inplace=True)
if self.use_dropout:
out = self.dropout(out)
out = F.avg_pool2d(out, 2, ceil_mode=True)
return out
class DenseNet(nn.Module):
def __init__(
self,
growth_rate: int,
num_layers: int,
reduction: float = 0.5,
bottleneck: bool = True,
use_dropout: bool = True,
):
super(DenseNet, self).__init__()
n_dense_blocks = num_layers
n_channels = 2 * growth_rate
self.conv1 = nn.Conv2d(
1, n_channels, kernel_size=7, padding=3, stride=2, bias=False
)
self.norm1 = nn.BatchNorm2d(n_channels)
self.dense1 = self._make_dense(
n_channels, growth_rate, n_dense_blocks, bottleneck, use_dropout
)
n_channels += n_dense_blocks * growth_rate
n_out_channels = int(math.floor(n_channels * reduction))
self.trans1 = _Transition(n_channels, n_out_channels, use_dropout)
n_channels = n_out_channels
self.dense2 = self._make_dense(
n_channels, growth_rate, n_dense_blocks, bottleneck, use_dropout
)
n_channels += n_dense_blocks * growth_rate
n_out_channels = int(math.floor(n_channels * reduction))
self.trans2 = _Transition(n_channels, n_out_channels, use_dropout)
n_channels = n_out_channels
self.dense3 = self._make_dense(
n_channels, growth_rate, n_dense_blocks, bottleneck, use_dropout
)
self.out_channels = n_channels + n_dense_blocks * growth_rate
self.post_norm = nn.BatchNorm2d(self.out_channels)
@staticmethod
def _make_dense(n_channels, growth_rate, n_dense_blocks, bottleneck, use_dropout):
layers = []
for _ in range(int(n_dense_blocks)):
if bottleneck:
layers.append(_Bottleneck(
n_channels, growth_rate, use_dropout))
else:
layers.append(_SingleLayer(
n_channels, growth_rate, use_dropout))
n_channels += growth_rate
return nn.Sequential(*layers)
def forward(self, x, x_mask):
out = self.conv1(x)
out = self.norm1(out)
out_mask = x_mask[:, 0::2, 0::2]
out = F.relu(out, inplace=True)
out = F.max_pool2d(out, 2, ceil_mode=True)
out_mask = out_mask[:, 0::2, 0::2]
out = self.dense1(out)
out = self.trans1(out)
out_mask = out_mask[:, 0::2, 0::2]
out = self.dense2(out)
out = self.trans2(out)
out_mask = out_mask[:, 0::2, 0::2]
out = self.dense3(out)
out = self.post_norm(out)
return out, out_mask
class Encoder(nn.Module):
def __init__(self, d_model: int, growth_rate: int, num_layers: int):
super().__init__()
self.model = DenseNet(growth_rate=growth_rate, num_layers=num_layers)
self.feature_proj = nn.Conv2d(
self.model.out_channels, d_model, kernel_size=1)
self.pos_enc_2d = ImgPosEnc(d_model, normalize=True)
self.norm = nn.LayerNorm(d_model)
def forward(
self, img: FloatTensor, img_mask: LongTensor
) -> Tuple[FloatTensor, LongTensor]:
"""encode image to feature
Parameters
----------
img : FloatTensor
[b, 1, h', w']
img_mask: LongTensor
[b, h', w']
Returns
-------
Tuple[FloatTensor, LongTensor]
[b, h, w, d], [b, h, w]
"""
# extract feature
feature, mask = self.model(img, img_mask)
feature = self.feature_proj(feature)
# proj
feature = rearrange(feature, "b d h w -> b h w d")
# positional encoding
feature = self.pos_enc_2d(feature, mask)
feature = self.norm(feature)
return feature, mask
|