Spaces:
Running on Zero
Running on Zero
File size: 4,629 Bytes
d066167 | 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 | import torch
import torch.nn as nn
from refnet.modules.layers import zero_module
from refnet.modules.attention import MemoryEfficientAttention
from refnet.modules.transformer import BasicTransformerBlock
from refnet.util import checkpoint_wrapper, exists
from refnet.util import load_weights
class NormalizedLinear(nn.Module):
def __init__(self, dim, output_dim, checkpoint=True):
super().__init__()
self.layers = nn.Sequential(
nn.Linear(dim, output_dim),
nn.LayerNorm(output_dim)
)
self.checkpoint = checkpoint
@checkpoint_wrapper
def forward(self, x):
return self.layers(x)
class GlobalProjection(nn.Module):
def __init__(self, input_dim, output_dim, heads, dim_head=128, checkpoint=True):
super().__init__()
self.c_dim = output_dim
self.dim_head = dim_head
self.head = (heads[0], heads[0] * heads[1])
self.proj1 = nn.Linear(input_dim, dim_head * heads[0])
self.proj2 = nn.Sequential(
nn.SiLU(),
zero_module(nn.Linear(dim_head, output_dim * heads[1])),
)
self.norm = nn.LayerNorm(output_dim)
self.checkpoint = checkpoint
@checkpoint_wrapper
def forward(self, x):
x = self.proj1(x).reshape(-1, self.head[0], self.dim_head).contiguous()
x = self.proj2(x).reshape(-1, self.head[1], self.c_dim).contiguous()
return self.norm(x)
class ClusterConcat(nn.Module):
def __init__(self, input_dim, c_dim, output_dim, dim_head=64, token_length=196, checkpoint=True):
super().__init__()
self.attn = MemoryEfficientAttention(input_dim, dim_head=dim_head)
self.norm = nn.LayerNorm(input_dim)
self.proj = nn.Sequential(
nn.Linear(input_dim + c_dim, output_dim),
nn.SiLU(),
nn.Linear(output_dim, output_dim),
nn.LayerNorm(output_dim)
)
self.token_length = token_length
self.checkpoint = checkpoint
@checkpoint_wrapper
def forward(self, x, emb, fgbg=False, *args, **kwargs):
x = self.attn(x)[:, :self.token_length]
x = self.norm(x)
x = torch.cat([x, emb], 2)
x = self.proj(x)
if fgbg:
x = torch.cat(torch.chunk(x, 2), 1)
return x
class RecoveryClusterConcat(ClusterConcat):
def __init__(self, input_dim, c_dim, output_dim, dim_head=64, *args, **kwargs):
super().__init__(input_dim, c_dim, output_dim, dim_head=dim_head, *args, **kwargs)
self.transformer = BasicTransformerBlock(
output_dim, output_dim//dim_head, dim_head,
disable_cross_attn=True, checkpoint=False
)
@checkpoint_wrapper
def forward(self, x, emb, bg=False):
x = self.attn(x)[:, :self.token_length]
x = self.norm(x)
x = torch.cat([x, emb], 2)
x = self.proj(x)
if bg:
x = self.transformer(x)
return x
class LogitClusterConcat(ClusterConcat):
def __init__(self, c_dim, mlp_in_dim, mlp_ckpt_path=None, *args, **kwargs):
super().__init__(c_dim=c_dim, *args, **kwargs)
self.mlp = AdaptiveMLP(c_dim, mlp_in_dim)
if exists(mlp_ckpt_path):
self.mlp.load_state_dict(load_weights(mlp_ckpt_path), strict=True)
@checkpoint_wrapper
def forward(self, x, emb, bg=False):
with torch.no_grad():
emb = self.mlp(emb).detach()
return super().forward(x, emb, bg)
class AdaptiveMLP(nn.Module):
def __init__(self, dim, in_dim, layers=4, checkpoint=True):
super().__init__()
model = [nn.Sequential(nn.Linear(in_dim, dim))]
for i in range(1, layers):
model += [nn.Sequential(
nn.SiLU(),
nn.LayerNorm(dim),
nn.Linear(dim, dim)
)]
self.mlp = nn.Sequential(*model)
self.fusion_layer = nn.Linear(dim * layers, dim, bias=False)
self.norm = nn.LayerNorm(dim)
self.checkpoint = checkpoint
@checkpoint_wrapper
def forward(self, x):
fx = []
for layer in self.mlp:
x = layer(x)
fx.append(x)
x = torch.cat(fx, dim=2)
out = self.fusion_layer(x)
out = self.norm(out)
return out
class Concat(nn.Module):
def __init__(self, *args, **kwargs):
super().__init__()
def forward(self, x, y, *args, **kwargs):
return torch.cat([x, y], dim=-1) |