File size: 2,983 Bytes
b5a6aa7 |
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 |
A=None
e_sd_pt="ko.pth"
d_sd_pt="ok.pth"
import torch as t, torch.nn as nn, torch.nn.functional as F
def C(n_in, n_out, **kwargs):
return nn.Conv2d(n_in, n_out, 3, padding=1, **kwargs)
class Clamp(nn.Module):
def forward(self, x):
return t.tanh(x / 3) * 3
class B(nn.Module):
def __init__(self, n_in, n_out):
super().__init__()
self.conv = nn.Sequential(C(n_in, n_out), nn.ReLU(), C(n_out, n_out), nn.ReLU(), C(n_out, n_out))
self.skip = nn.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity()
self.fuse = nn.ReLU()
def forward(self, x):
return self.fuse(self.conv(x) + self.skip(x))
def E(latent_channels=4):
return nn.Sequential(
C(3, 64), B(64, 64),
C(64, 64, stride=2, bias=False), B(64, 64), B(64, 64), B(64, 64),
C(64, 64, stride=2, bias=False), B(64, 64), B(64, 64), B(64, 64),
C(64, 64, stride=2, bias=False), B(64, 64), B(64, 64), B(64, 64),
C(64, latent_channels),
)
def D(latent_channels=16):
return nn.Sequential(
Clamp(),
C(latent_channels, 48),nn.ReLU(),B(48, 48), B(48, 48),
nn.Upsample(scale_factor=2), C(48, 48, bias=False),B(48, 48), B(48, 48),
nn.Upsample(scale_factor=2), C(48, 48, bias=False),B(48, 48),
nn.Upsample(scale_factor=2), C(48, 48, bias=False),B(48, 48),
C(48, 3),
)
class M(nn.Module):
lm, ls = 3, 0.5
def __init__(s, ep="encoder.pth", dp="decoder.pth", lc=None):
super().__init__()
if lc is None: lc = s.glc(str(ep))
s.e, s.d = E(lc), D(lc)
def f(sd, mod, pfx):
f_sd = {k.strip(pfx): v for k, v in sd.items() if k.strip(pfx) in mod.state_dict() and v.size() == mod.state_dict()[k.strip(pfx)].size()}
mod.load_state_dict(f_sd, strict=False)
if ep: f(t.load(ep, map_location="cpu", weights_only=True), s.e, "encoder.")
if dp: f(t.load(dp, map_location="cpu", weights_only=True), s.d, "decoder.")
s.e.requires_grad_(False)
s.d.requires_grad_(False)
def glc(s, ep): return 16 if "taef1" in ep or "taesd3" in ep else 4
@staticmethod
def sl(x): return x.div(2 * M.lm).add(M.ls).clamp(0, 1)
@staticmethod
def ul(x): return x.sub(M.ls).mul(2 * M.lm)
def forward(s, x, rl=False):
l, o = s.e(x), s.d(s.e(x))
return (o.clamp(0, 1), l) if rl else o.clamp(0, 1)
def filter_state_dict(model, name):
state_dict = t.load(e_sd_pt if name=="E" else d_sd_pt, map_location="cpu", weights_only=True)
prefix = 'encoder.' if name=="E" else 'decoder.'
return {k.strip(prefix): v for k, v in state_dict.items() if k.strip(prefix) in model.state_dict() and v.size() == model.state_dict()[k.strip(prefix)].size()}
def _load(model, name, dtype=t.bfloat16):
model = E(16) if name=="E" else D(16)
model.load_state_dict(filter_state_dict(model, name), strict=False)
model.requires_grad_(False).to(dtype=dtype)
return model |