michaelriedl
commited on
Commit
•
8f71eda
1
Parent(s):
4fe7b39
Initial dump
Browse files- LightweightGANConfig.py +31 -0
- LightweightGANModel.py +29 -0
- config.json +24 -0
- deploy.py +385 -0
- pytorch_model.bin +3 -0
LightweightGANConfig.py
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import PretrainedConfig
|
2 |
+
|
3 |
+
|
4 |
+
class LightweightGANConfig(PretrainedConfig):
|
5 |
+
model_type = "lightweight-gan"
|
6 |
+
|
7 |
+
def __init__(
|
8 |
+
self,
|
9 |
+
image_size=64,
|
10 |
+
latent_dim=256,
|
11 |
+
fmap_max=512,
|
12 |
+
fmap_inverse_coef=12,
|
13 |
+
transparent=False,
|
14 |
+
greyscale=False,
|
15 |
+
attn_res_layers=[32],
|
16 |
+
freq_chan_attn=False,
|
17 |
+
syncbatchnorm=False,
|
18 |
+
antialias=False,
|
19 |
+
**kwargs,
|
20 |
+
):
|
21 |
+
self.image_size = image_size
|
22 |
+
self.latent_dim = latent_dim
|
23 |
+
self.fmap_max = fmap_max
|
24 |
+
self.fmap_inverse_coef = fmap_inverse_coef
|
25 |
+
self.transparent = transparent
|
26 |
+
self.greyscale = greyscale
|
27 |
+
self.attn_res_layers = attn_res_layers
|
28 |
+
self.freq_chan_attn = freq_chan_attn
|
29 |
+
self.syncbatchnorm = syncbatchnorm
|
30 |
+
self.antialias = antialias
|
31 |
+
super().__init__(**kwargs)
|
LightweightGANModel.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from transformers import PreTrainedModel
|
3 |
+
from .LightweightGANConfig import LightweightGANConfig
|
4 |
+
from .deploy import Generator
|
5 |
+
|
6 |
+
|
7 |
+
class LightweightGANModel(PreTrainedModel):
|
8 |
+
config_class = LightweightGANConfig
|
9 |
+
|
10 |
+
def __init__(self, config):
|
11 |
+
super().__init__(config)
|
12 |
+
self.model = Generator(
|
13 |
+
image_size=config.image_size,
|
14 |
+
latent_dim=config.latent_dim,
|
15 |
+
fmap_max=config.fmap_max,
|
16 |
+
fmap_inverse_coef=config.fmap_inverse_coef,
|
17 |
+
transparent=config.transparent,
|
18 |
+
greyscale=config.greyscale,
|
19 |
+
attn_res_layers=config.attn_res_layers,
|
20 |
+
freq_chan_attn=config.freq_chan_attn,
|
21 |
+
syncbatchnorm=config.syncbatchnorm,
|
22 |
+
antialias=config.antialias,
|
23 |
+
)
|
24 |
+
|
25 |
+
def forward(self, tensor):
|
26 |
+
return self.model(tensor)
|
27 |
+
|
28 |
+
def load_params(self, pt_file):
|
29 |
+
self.model.load_state_dict(torch.load(pt_file))
|
config.json
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"antialias": false,
|
3 |
+
"architectures": [
|
4 |
+
"LightweightGANModel"
|
5 |
+
],
|
6 |
+
"attn_res_layers": [
|
7 |
+
32
|
8 |
+
],
|
9 |
+
"auto_map": {
|
10 |
+
"AutoConfig": "LightweightGANConfig.LightweightGANConfig",
|
11 |
+
"AutoModel": "LightweightGANModel.LightweightGANModel"
|
12 |
+
},
|
13 |
+
"fmap_inverse_coef": 12,
|
14 |
+
"fmap_max": 512,
|
15 |
+
"freq_chan_attn": false,
|
16 |
+
"greyscale": false,
|
17 |
+
"image_size": 256,
|
18 |
+
"latent_dim": 256,
|
19 |
+
"model_type": "lightweight-gan",
|
20 |
+
"syncbatchnorm": false,
|
21 |
+
"torch_dtype": "float32",
|
22 |
+
"transformers_version": "4.31.0",
|
23 |
+
"transparent": true
|
24 |
+
}
|
deploy.py
ADDED
@@ -0,0 +1,385 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import torch
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from math import log2
|
5 |
+
from torch import nn, einsum
|
6 |
+
from kornia.filters import filter2d
|
7 |
+
from einops import reduce, rearrange, repeat
|
8 |
+
|
9 |
+
|
10 |
+
def exists(val):
|
11 |
+
return val is not None
|
12 |
+
|
13 |
+
|
14 |
+
def is_power_of_two(val):
|
15 |
+
return log2(val).is_integer()
|
16 |
+
|
17 |
+
|
18 |
+
def default(val, d):
|
19 |
+
return val if exists(val) else d
|
20 |
+
|
21 |
+
|
22 |
+
def get_1d_dct(i, freq, L):
|
23 |
+
result = math.cos(math.pi * freq * (i + 0.5) / L) / math.sqrt(L)
|
24 |
+
return result * (1 if freq == 0 else math.sqrt(2))
|
25 |
+
|
26 |
+
|
27 |
+
def get_dct_weights(width, channel, fidx_u, fidx_v):
|
28 |
+
dct_weights = torch.zeros(1, channel, width, width)
|
29 |
+
c_part = channel // len(fidx_u)
|
30 |
+
|
31 |
+
for i, (u_x, v_y) in enumerate(zip(fidx_u, fidx_v)):
|
32 |
+
for x in range(width):
|
33 |
+
for y in range(width):
|
34 |
+
coor_value = get_1d_dct(x, u_x, width) * get_1d_dct(y, v_y, width)
|
35 |
+
dct_weights[:, i * c_part : (i + 1) * c_part, x, y] = coor_value
|
36 |
+
|
37 |
+
return dct_weights
|
38 |
+
|
39 |
+
|
40 |
+
class Blur(nn.Module):
|
41 |
+
def __init__(self):
|
42 |
+
super().__init__()
|
43 |
+
f = torch.Tensor([1, 2, 1])
|
44 |
+
self.register_buffer("f", f)
|
45 |
+
|
46 |
+
def forward(self, x):
|
47 |
+
f = self.f
|
48 |
+
f = f[None, None, :] * f[None, :, None]
|
49 |
+
return filter2d(x, f, normalized=True)
|
50 |
+
|
51 |
+
|
52 |
+
class ChanNorm(nn.Module):
|
53 |
+
def __init__(self, dim, eps=1e-5):
|
54 |
+
super().__init__()
|
55 |
+
self.eps = eps
|
56 |
+
self.g = nn.Parameter(torch.ones(1, dim, 1, 1))
|
57 |
+
self.b = nn.Parameter(torch.zeros(1, dim, 1, 1))
|
58 |
+
|
59 |
+
def forward(self, x):
|
60 |
+
var = torch.var(x, dim=1, unbiased=False, keepdim=True)
|
61 |
+
mean = torch.mean(x, dim=1, keepdim=True)
|
62 |
+
return (x - mean) / (var + self.eps).sqrt() * self.g + self.b
|
63 |
+
|
64 |
+
|
65 |
+
def Conv2dSame(dim_in, dim_out, kernel_size, bias=True):
|
66 |
+
pad_left = kernel_size // 2
|
67 |
+
pad_right = (pad_left - 1) if (kernel_size % 2) == 0 else pad_left
|
68 |
+
|
69 |
+
return nn.Sequential(
|
70 |
+
nn.ZeroPad2d((pad_left, pad_right, pad_left, pad_right)),
|
71 |
+
nn.Conv2d(dim_in, dim_out, kernel_size, bias=bias),
|
72 |
+
)
|
73 |
+
|
74 |
+
|
75 |
+
class DepthWiseConv2d(nn.Module):
|
76 |
+
def __init__(self, dim_in, dim_out, kernel_size, padding=0, stride=1, bias=True):
|
77 |
+
super().__init__()
|
78 |
+
self.net = nn.Sequential(
|
79 |
+
nn.Conv2d(
|
80 |
+
dim_in,
|
81 |
+
dim_in,
|
82 |
+
kernel_size=kernel_size,
|
83 |
+
padding=padding,
|
84 |
+
groups=dim_in,
|
85 |
+
stride=stride,
|
86 |
+
bias=bias,
|
87 |
+
),
|
88 |
+
nn.Conv2d(dim_in, dim_out, kernel_size=1, bias=bias),
|
89 |
+
)
|
90 |
+
|
91 |
+
def forward(self, x):
|
92 |
+
return self.net(x)
|
93 |
+
|
94 |
+
|
95 |
+
class FCANet(nn.Module):
|
96 |
+
def __init__(self, *, chan_in, chan_out, reduction=4, width):
|
97 |
+
super().__init__()
|
98 |
+
|
99 |
+
freq_w, freq_h = ([0] * 8), list(
|
100 |
+
range(8)
|
101 |
+
) # in paper, it seems 16 frequencies was ideal
|
102 |
+
dct_weights = get_dct_weights(
|
103 |
+
width, chan_in, [*freq_w, *freq_h], [*freq_h, *freq_w]
|
104 |
+
)
|
105 |
+
self.register_buffer("dct_weights", dct_weights)
|
106 |
+
|
107 |
+
chan_intermediate = max(3, chan_out // reduction)
|
108 |
+
|
109 |
+
self.net = nn.Sequential(
|
110 |
+
nn.Conv2d(chan_in, chan_intermediate, 1),
|
111 |
+
nn.LeakyReLU(0.1),
|
112 |
+
nn.Conv2d(chan_intermediate, chan_out, 1),
|
113 |
+
nn.Sigmoid(),
|
114 |
+
)
|
115 |
+
|
116 |
+
def forward(self, x):
|
117 |
+
x = reduce(
|
118 |
+
x * self.dct_weights, "b c (h h1) (w w1) -> b c h1 w1", "sum", h1=1, w1=1
|
119 |
+
)
|
120 |
+
return self.net(x)
|
121 |
+
|
122 |
+
|
123 |
+
class Generator(nn.Module):
|
124 |
+
def __init__(
|
125 |
+
self,
|
126 |
+
*,
|
127 |
+
image_size,
|
128 |
+
latent_dim=256,
|
129 |
+
fmap_max=512,
|
130 |
+
fmap_inverse_coef=12,
|
131 |
+
transparent=False,
|
132 |
+
greyscale=False,
|
133 |
+
attn_res_layers=[],
|
134 |
+
freq_chan_attn=False,
|
135 |
+
syncbatchnorm=False,
|
136 |
+
antialias=False,
|
137 |
+
):
|
138 |
+
super().__init__()
|
139 |
+
resolution = log2(image_size)
|
140 |
+
assert is_power_of_two(image_size), "image size must be a power of 2"
|
141 |
+
|
142 |
+
# Set the normalization and blur
|
143 |
+
norm_class = nn.SyncBatchNorm if syncbatchnorm else nn.BatchNorm2d
|
144 |
+
Blur = nn.Identity if not antialias else Blur
|
145 |
+
|
146 |
+
if transparent:
|
147 |
+
init_channel = 4
|
148 |
+
elif greyscale:
|
149 |
+
init_channel = 1
|
150 |
+
else:
|
151 |
+
init_channel = 3
|
152 |
+
|
153 |
+
self.latent_dim = latent_dim
|
154 |
+
|
155 |
+
fmap_max = default(fmap_max, latent_dim)
|
156 |
+
|
157 |
+
self.initial_conv = nn.Sequential(
|
158 |
+
nn.ConvTranspose2d(latent_dim, latent_dim * 2, 4),
|
159 |
+
norm_class(latent_dim * 2),
|
160 |
+
nn.GLU(dim=1),
|
161 |
+
)
|
162 |
+
|
163 |
+
num_layers = int(resolution) - 2
|
164 |
+
features = list(
|
165 |
+
map(lambda n: (n, 2 ** (fmap_inverse_coef - n)), range(2, num_layers + 2))
|
166 |
+
)
|
167 |
+
features = list(map(lambda n: (n[0], min(n[1], fmap_max)), features))
|
168 |
+
features = list(map(lambda n: 3 if n[0] >= 8 else n[1], features))
|
169 |
+
features = [latent_dim, *features]
|
170 |
+
|
171 |
+
in_out_features = list(zip(features[:-1], features[1:]))
|
172 |
+
|
173 |
+
self.res_layers = range(2, num_layers + 2)
|
174 |
+
self.layers = nn.ModuleList([])
|
175 |
+
self.res_to_feature_map = dict(zip(self.res_layers, in_out_features))
|
176 |
+
|
177 |
+
self.sle_map = ((3, 7), (4, 8), (5, 9), (6, 10))
|
178 |
+
self.sle_map = list(
|
179 |
+
filter(lambda t: t[0] <= resolution and t[1] <= resolution, self.sle_map)
|
180 |
+
)
|
181 |
+
self.sle_map = dict(self.sle_map)
|
182 |
+
|
183 |
+
self.num_layers_spatial_res = 1
|
184 |
+
|
185 |
+
for res, (chan_in, chan_out) in zip(self.res_layers, in_out_features):
|
186 |
+
image_width = 2**res
|
187 |
+
|
188 |
+
attn = None
|
189 |
+
if image_width in attn_res_layers:
|
190 |
+
attn = PreNorm(chan_in, LinearAttention(chan_in))
|
191 |
+
|
192 |
+
sle = None
|
193 |
+
if res in self.sle_map:
|
194 |
+
residual_layer = self.sle_map[res]
|
195 |
+
sle_chan_out = self.res_to_feature_map[residual_layer - 1][-1]
|
196 |
+
|
197 |
+
if freq_chan_attn:
|
198 |
+
sle = FCANet(
|
199 |
+
chan_in=chan_out, chan_out=sle_chan_out, width=2 ** (res + 1)
|
200 |
+
)
|
201 |
+
else:
|
202 |
+
sle = GlobalContext(chan_in=chan_out, chan_out=sle_chan_out)
|
203 |
+
|
204 |
+
layer = nn.ModuleList(
|
205 |
+
[
|
206 |
+
nn.Sequential(
|
207 |
+
PixelShuffleUpsample(chan_in),
|
208 |
+
Blur(),
|
209 |
+
Conv2dSame(chan_in, chan_out * 2, 4),
|
210 |
+
Noise(),
|
211 |
+
norm_class(chan_out * 2),
|
212 |
+
nn.GLU(dim=1),
|
213 |
+
),
|
214 |
+
sle,
|
215 |
+
attn,
|
216 |
+
]
|
217 |
+
)
|
218 |
+
self.layers.append(layer)
|
219 |
+
|
220 |
+
self.out_conv = nn.Conv2d(features[-1], init_channel, 3, padding=1)
|
221 |
+
|
222 |
+
def forward(self, x):
|
223 |
+
x = rearrange(x, "b c -> b c () ()")
|
224 |
+
x = self.initial_conv(x)
|
225 |
+
x = F.normalize(x, dim=1)
|
226 |
+
|
227 |
+
residuals = dict()
|
228 |
+
|
229 |
+
for res, (up, sle, attn) in zip(self.res_layers, self.layers):
|
230 |
+
if exists(attn):
|
231 |
+
x = attn(x) + x
|
232 |
+
|
233 |
+
x = up(x)
|
234 |
+
|
235 |
+
if exists(sle):
|
236 |
+
out_res = self.sle_map[res]
|
237 |
+
residual = sle(x)
|
238 |
+
residuals[out_res] = residual
|
239 |
+
|
240 |
+
next_res = res + 1
|
241 |
+
if next_res in residuals:
|
242 |
+
x = x * residuals[next_res]
|
243 |
+
|
244 |
+
return self.out_conv(x)
|
245 |
+
|
246 |
+
|
247 |
+
class GlobalContext(nn.Module):
|
248 |
+
def __init__(self, *, chan_in, chan_out):
|
249 |
+
super().__init__()
|
250 |
+
self.to_k = nn.Conv2d(chan_in, 1, 1)
|
251 |
+
chan_intermediate = max(3, chan_out // 2)
|
252 |
+
|
253 |
+
self.net = nn.Sequential(
|
254 |
+
nn.Conv2d(chan_in, chan_intermediate, 1),
|
255 |
+
nn.LeakyReLU(0.1),
|
256 |
+
nn.Conv2d(chan_intermediate, chan_out, 1),
|
257 |
+
nn.Sigmoid(),
|
258 |
+
)
|
259 |
+
|
260 |
+
def forward(self, x):
|
261 |
+
context = self.to_k(x)
|
262 |
+
context = context.flatten(2).softmax(dim=-1)
|
263 |
+
out = einsum("b i n, b c n -> b c i", context, x.flatten(2))
|
264 |
+
out = out.unsqueeze(-1)
|
265 |
+
return self.net(out)
|
266 |
+
|
267 |
+
|
268 |
+
class LinearAttention(nn.Module):
|
269 |
+
def __init__(self, dim, dim_head=64, heads=8, kernel_size=3):
|
270 |
+
super().__init__()
|
271 |
+
self.scale = dim_head**-0.5
|
272 |
+
self.heads = heads
|
273 |
+
self.dim_head = dim_head
|
274 |
+
inner_dim = dim_head * heads
|
275 |
+
|
276 |
+
self.kernel_size = kernel_size
|
277 |
+
self.nonlin = nn.GELU()
|
278 |
+
|
279 |
+
self.to_lin_q = nn.Conv2d(dim, inner_dim, 1, bias=False)
|
280 |
+
self.to_lin_kv = DepthWiseConv2d(dim, inner_dim * 2, 3, padding=1, bias=False)
|
281 |
+
|
282 |
+
self.to_q = nn.Conv2d(dim, inner_dim, 1, bias=False)
|
283 |
+
self.to_kv = nn.Conv2d(dim, inner_dim * 2, 1, bias=False)
|
284 |
+
|
285 |
+
self.to_out = nn.Conv2d(inner_dim * 2, dim, 1)
|
286 |
+
|
287 |
+
def forward(self, fmap):
|
288 |
+
h, x, y = self.heads, *fmap.shape[-2:]
|
289 |
+
|
290 |
+
# linear attention
|
291 |
+
|
292 |
+
lin_q, lin_k, lin_v = (
|
293 |
+
self.to_lin_q(fmap),
|
294 |
+
*self.to_lin_kv(fmap).chunk(2, dim=1),
|
295 |
+
)
|
296 |
+
lin_q, lin_k, lin_v = map(
|
297 |
+
lambda t: rearrange(t, "b (h c) x y -> (b h) (x y) c", h=h),
|
298 |
+
(lin_q, lin_k, lin_v),
|
299 |
+
)
|
300 |
+
|
301 |
+
lin_q = lin_q.softmax(dim=-1)
|
302 |
+
lin_k = lin_k.softmax(dim=-2)
|
303 |
+
|
304 |
+
lin_q = lin_q * self.scale
|
305 |
+
|
306 |
+
context = einsum("b n d, b n e -> b d e", lin_k, lin_v)
|
307 |
+
lin_out = einsum("b n d, b d e -> b n e", lin_q, context)
|
308 |
+
lin_out = rearrange(lin_out, "(b h) (x y) d -> b (h d) x y", h=h, x=x, y=y)
|
309 |
+
|
310 |
+
# conv-like full attention
|
311 |
+
|
312 |
+
q, k, v = (self.to_q(fmap), *self.to_kv(fmap).chunk(2, dim=1))
|
313 |
+
q, k, v = map(
|
314 |
+
lambda t: rearrange(t, "b (h c) x y -> (b h) c x y", h=h), (q, k, v)
|
315 |
+
)
|
316 |
+
|
317 |
+
k = F.unfold(k, kernel_size=self.kernel_size, padding=self.kernel_size // 2)
|
318 |
+
v = F.unfold(v, kernel_size=self.kernel_size, padding=self.kernel_size // 2)
|
319 |
+
|
320 |
+
k, v = map(
|
321 |
+
lambda t: rearrange(t, "b (d j) n -> b n j d", d=self.dim_head), (k, v)
|
322 |
+
)
|
323 |
+
|
324 |
+
q = rearrange(q, "b c ... -> b (...) c") * self.scale
|
325 |
+
|
326 |
+
sim = einsum("b i d, b i j d -> b i j", q, k)
|
327 |
+
sim = sim - sim.amax(dim=-1, keepdim=True).detach()
|
328 |
+
|
329 |
+
attn = sim.softmax(dim=-1)
|
330 |
+
|
331 |
+
full_out = einsum("b i j, b i j d -> b i d", attn, v)
|
332 |
+
full_out = rearrange(full_out, "(b h) (x y) d -> b (h d) x y", h=h, x=x, y=y)
|
333 |
+
|
334 |
+
# add outputs of linear attention + conv like full attention
|
335 |
+
|
336 |
+
lin_out = self.nonlin(lin_out)
|
337 |
+
out = torch.cat((lin_out, full_out), dim=1)
|
338 |
+
return self.to_out(out)
|
339 |
+
|
340 |
+
|
341 |
+
class Noise(nn.Module):
|
342 |
+
def __init__(self):
|
343 |
+
super().__init__()
|
344 |
+
self.weight = nn.Parameter(torch.zeros(1))
|
345 |
+
|
346 |
+
def forward(self, x, noise=None):
|
347 |
+
b, _, h, w, device = *x.shape, x.device
|
348 |
+
|
349 |
+
if not exists(noise):
|
350 |
+
noise = torch.randn(b, 1, h, w, device=device)
|
351 |
+
|
352 |
+
return x + self.weight * noise
|
353 |
+
|
354 |
+
|
355 |
+
class PixelShuffleUpsample(nn.Module):
|
356 |
+
def __init__(self, dim, dim_out=None):
|
357 |
+
super().__init__()
|
358 |
+
dim_out = default(dim_out, dim)
|
359 |
+
conv = nn.Conv2d(dim, dim_out * 4, 1)
|
360 |
+
|
361 |
+
self.net = nn.Sequential(conv, nn.SiLU(), nn.PixelShuffle(2))
|
362 |
+
|
363 |
+
self.init_conv_(conv)
|
364 |
+
|
365 |
+
def init_conv_(self, conv):
|
366 |
+
o, i, h, w = conv.weight.shape
|
367 |
+
conv_weight = torch.empty(o // 4, i, h, w)
|
368 |
+
nn.init.kaiming_uniform_(conv_weight)
|
369 |
+
conv_weight = repeat(conv_weight, "o ... -> (o 4) ...")
|
370 |
+
|
371 |
+
conv.weight.data.copy_(conv_weight)
|
372 |
+
nn.init.zeros_(conv.bias.data)
|
373 |
+
|
374 |
+
def forward(self, x):
|
375 |
+
return self.net(x)
|
376 |
+
|
377 |
+
|
378 |
+
class PreNorm(nn.Module):
|
379 |
+
def __init__(self, dim, fn):
|
380 |
+
super().__init__()
|
381 |
+
self.fn = fn
|
382 |
+
self.norm = ChanNorm(dim)
|
383 |
+
|
384 |
+
def forward(self, x):
|
385 |
+
return self.fn(self.norm(x))
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4cb74d9d864a2aa6256ee59c7c1e8efb6ac2f0c73d61ac987e97a28f212ff09e
|
3 |
+
size 96248639
|