Spaces:
Running on Zero
Running on Zero
File size: 5,244 Bytes
7405e85 | 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 | from functools import partial
from itertools import repeat
import collections.abc
import torch
import torch.nn as nn
def _ntuple(n):
def parse(x):
if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
return tuple(x)
return tuple(repeat(x, n))
return parse
to_1tuple = _ntuple(1)
to_2tuple = _ntuple(2)
to_3tuple = _ntuple(3)
to_4tuple = _ntuple(4)
to_ntuple = _ntuple
def get_conv(in_ch, out_ch, kernel_size, stride, padding, zero_bias=True, zero_weights=False):
conv = nn.Conv2d(in_ch, out_ch, kernel_size, stride, padding)
if zero_bias:
conv.bias.data.mul_(0.0)
if zero_weights:
conv.weight.data.mul_(0.0)
return conv
def conv_k1s1(in_ch, out_ch, zero_bias=True, zero_weights=False):
return get_conv(in_ch, out_ch, 1, 1, 0, zero_bias, zero_weights)
def conv_k3s1(in_ch, out_ch, zero_bias=True, zero_weights=False):
return get_conv(in_ch, out_ch, 3, 1, 1, zero_bias, zero_weights)
def patch_downsample(in_ch, out_ch, rate=2):
return get_conv(in_ch, out_ch, kernel_size=rate, stride=rate, padding=0)
def patch_upsample(in_ch, out_ch, rate=2):
conv = nn.Sequential(
get_conv(in_ch, out_ch * (rate ** 2), kernel_size=1, stride=1, padding=0),
nn.PixelShuffle(rate)
)
return conv
class SetKey(nn.Module):
""" A dummy layer that is used to mark the position of a layer in the network.
"""
def __init__(self, key):
super().__init__()
self.key = key
def forward(self, x):
return x
class CompressionStopFlag(nn.Module):
""" A dummy layer that is used to mark the stop position of encoding bits.
"""
def __init__(self):
super().__init__()
def forward(self, x):
return x
class FeatureExtractorWithEmbedding(nn.Module):
def __init__(self, blocks):
super().__init__()
self.enc_blocks = nn.ModuleList(blocks)
def forward(self, x, emb=None):
features = {}
for i, block in enumerate(self.enc_blocks):
if isinstance(block, SetKey):
features[block.key] = x
elif getattr(block, 'requires_embedding', False):
x = block(x, emb)
else:
x = block(x)
return x, features
class Mlp(nn.Module):
def __init__(
self,
in_features,
hidden_features=None,
out_features=None,
act_layer=nn.GELU,
norm_layer=None,
bias=True,
use_conv=False,
):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
bias = to_2tuple(bias)
linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
self.fc1 = linear_layer(in_features, hidden_features, bias=bias[0])
self.act = act_layer()
self.norm = norm_layer(hidden_features) if norm_layer is not None else nn.Identity()
self.fc2 = linear_layer(hidden_features, out_features, bias=bias[1])
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.norm(x)
x = self.fc2(x)
return x
class ConvNeXtBlockAdaLN(nn.Module):
def __init__(self, dim, embed_dim=256, out_dim=None, kernel_size=7, mlp_ratio: float = 2,
residual=True, ls_init_value=1e-6):
super().__init__()
# depthwise conv
pad = (kernel_size - 1) // 2
self.conv_dw = nn.Conv2d(dim, dim, kernel_size=kernel_size, padding=pad, groups=dim)
# layer norm
self.norm = nn.LayerNorm(dim, eps=1e-6, elementwise_affine=False)
self.norm.affine = False # for FLOPs computing
# AdaLN
self.embedding_layer = nn.Sequential(
nn.Sequential(),
nn.Linear(embed_dim, 2*dim),
nn.Unflatten(1, unflattened_size=(1, 1, 2*dim))
)
# MLP
hidden = int(mlp_ratio * dim)
out_dim = out_dim or dim
self.mlp = Mlp(dim, hidden_features=hidden, out_features=out_dim, act_layer=nn.GELU)
# layer scaling
if ls_init_value >= 0:
self.gamma = nn.Parameter(torch.full(size=(1, out_dim, 1, 1), fill_value=1e-6))
else:
self.gamma = None
self.residual = residual
self.requires_embedding = True
for m in self.modules():
if isinstance(m, (nn.Conv2d, nn.Linear)):
nn.init.trunc_normal_(m.weight, std=0.02)
if m.bias is not None:
nn.init.zeros_(m.bias)
def forward(self, x, emb):
shortcut = x
# depthwise conv
x = self.conv_dw(x)
# layer norm
x = x.permute(0, 2, 3, 1).contiguous()
x = self.norm(x)
# AdaLN
embedding = self.embedding_layer(emb)
shift, scale = torch.chunk(embedding, chunks=2, dim=-1)
x = x * (1 + scale) + shift
# MLP
x = self.mlp(x)
# scaling
if self.gamma is not None:
x = x.mul(self.gamma.reshape(-1))
x = x.permute(0, 3, 1, 2).contiguous()
if self.residual:
x = x + shortcut
return x
|