manbeast3b
commited on
Commit
•
7e1e9b6
1
Parent(s):
71d2928
Create output.py
Browse files- src/output.py +70 -0
src/output.py
ADDED
@@ -0,0 +1,70 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
|
4 |
+
def conv(n_in, n_out, **kwargs):
|
5 |
+
return nn.Conv2d(n_in, n_out, 3, padding=1, **kwargs)
|
6 |
+
|
7 |
+
class Clamp(nn.Module):
|
8 |
+
def forward(self, x):
|
9 |
+
return torch.tanh(x / 3) * 3
|
10 |
+
|
11 |
+
class Block(nn.Module):
|
12 |
+
def __init__(self, n_in, n_out):
|
13 |
+
super().__init__()
|
14 |
+
self.conv = nn.Sequential(conv(n_in, n_out), nn.ReLU(), conv(n_out, n_out), nn.ReLU(), conv(n_out, n_out))
|
15 |
+
self.skip = nn.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity()
|
16 |
+
self.fuse = nn.ReLU()
|
17 |
+
def forward(self, x):
|
18 |
+
return self.fuse(self.conv(x) + self.skip(x))
|
19 |
+
|
20 |
+
def Encoder(latent_channels=4):
|
21 |
+
return nn.Sequential(
|
22 |
+
conv(3, 64), Block(64, 64),
|
23 |
+
conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
|
24 |
+
conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
|
25 |
+
conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
|
26 |
+
conv(64, latent_channels),
|
27 |
+
)
|
28 |
+
|
29 |
+
def Decoder(latent_channels=4):
|
30 |
+
return nn.Sequential(
|
31 |
+
Clamp(), conv(latent_channels, 64), nn.ReLU(),
|
32 |
+
Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
|
33 |
+
Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
|
34 |
+
Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
|
35 |
+
Block(64, 64), conv(64, 3),
|
36 |
+
)
|
37 |
+
|
38 |
+
class TAESD(nn.Module):
|
39 |
+
latent_magnitude = 3
|
40 |
+
latent_shift = 0.5
|
41 |
+
|
42 |
+
def __init__(self, encoder_path="taesd_encoder.pth", decoder_path="taesd_decoder.pth", latent_channels=None):
|
43 |
+
"""Initialize pretrained TAESD on the given device from the given checkpoints."""
|
44 |
+
super().__init__()
|
45 |
+
if latent_channels is None:
|
46 |
+
latent_channels = self.guess_latent_channels(str(encoder_path))
|
47 |
+
self.encoder = Encoder(latent_channels)
|
48 |
+
self.decoder = Decoder(latent_channels)
|
49 |
+
if encoder_path is not None:
|
50 |
+
self.encoder.load_state_dict(torch.load(encoder_path, map_location="cpu", weights_only=True))
|
51 |
+
if decoder_path is not None:
|
52 |
+
self.decoder.load_state_dict(torch.load(decoder_path, map_location="cpu", weights_only=True))
|
53 |
+
|
54 |
+
def guess_latent_channels(self, encoder_path):
|
55 |
+
"""guess latent channel count based on encoder filename"""
|
56 |
+
if "taef1" in encoder_path:
|
57 |
+
return 16
|
58 |
+
if "taesd3" in encoder_path:
|
59 |
+
return 16
|
60 |
+
return 4
|
61 |
+
|
62 |
+
@staticmethod
|
63 |
+
def scale_latents(x):
|
64 |
+
"""raw latents -> [0, 1]"""
|
65 |
+
return x.div(2 * TAESD.latent_magnitude).add(TAESD.latent_shift).clamp(0, 1)
|
66 |
+
|
67 |
+
@staticmethod
|
68 |
+
def unscale_latents(x):
|
69 |
+
"""[0, 1] -> raw latents"""
|
70 |
+
return x.sub(TAESD.latent_shift).mul(2 * TAESD.latent_magnitude)
|