| import torch |
| import torch as th |
| import torch.nn as nn |
| import torch.nn.functional as F |
|
|
| def conv(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 torch.tanh(x / 3) * 3 |
|
|
| class Block(nn.Module): |
| def __init__(self, n_in, n_out): |
| super().__init__() |
| self.conv = nn.Sequential(conv(n_in, n_out), nn.ReLU(), conv(n_out, n_out), nn.ReLU(), conv(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 Encoder(latent_channels=4): |
| return nn.Sequential( |
| conv(3, 64), Block(64, 64), |
| conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64), |
| conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64), |
| conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64), |
| conv(64, latent_channels), |
| ) |
|
|
| def Decoder(latent_channels=16): |
| return nn.Sequential( |
| Clamp(), |
| conv(latent_channels, 48), |
| nn.ReLU(), |
| Block(48, 48), Block(48, 48), |
| nn.Upsample(scale_factor=2), conv(48, 48, bias=False), |
| Block(48, 48), Block(48, 48), |
| nn.Upsample(scale_factor=2), conv(48, 48, bias=False), |
| Block(48, 48), |
| nn.Upsample(scale_factor=2), conv(48, 48, bias=False), |
| Block(48, 48), |
| conv(48, 3), |
| ) |
|
|
| |
|
|
| class Model(nn.Module): |
| latent_magnitude = 3 |
| latent_shift = 0.5 |
|
|
| def __init__(self, encoder_path="encoder.pth", decoder_path="decoder.pth", latent_channels=None): |
| super().__init__() |
| if latent_channels is None: |
| latent_channels = self.guess_latent_channels(str(encoder_path)) |
| self.encoder = Encoder(latent_channels) |
| self.decoder = Decoder(latent_channels) |
| |
| if encoder_path is not None: |
| encoder_state_dict = torch.load(encoder_path, map_location="cpu", weights_only=True) |
| filtered_state_dict = {k.strip('encoder.'): v for k, v in encoder_state_dict.items() if k.strip('encoder.') in self.encoder.state_dict() and v.size() == self.encoder.state_dict()[k.strip('encoder.')].size()} |
| print(f" num of keys in filtered: {len(filtered_state_dict)} and in decoder: {len(self.encoder.state_dict())}") |
| self.encoder.load_state_dict(filtered_state_dict, strict=False) |
| |
| if decoder_path is not None: |
| decoder_state_dict = torch.load(decoder_path, map_location="cpu", weights_only=True) |
| filtered_state_dict = {k.strip('decoder.'): v for k, v in decoder_state_dict.items() if k.strip('decoder.') in self.decoder.state_dict() and v.size() == self.decoder.state_dict()[k.strip('decoder.')].size()} |
| print(f" num of keys in filtered: {len(filtered_state_dict)} and in decoder: {len(self.decoder.state_dict())}") |
| self.decoder.load_state_dict(filtered_state_dict, strict=False) |
| |
| self.encoder.requires_grad_(False) |
| self.decoder.requires_grad_(False) |
|
|
| def guess_latent_channels(self, encoder_path): |
| if "taef1" in encoder_path:return 16 |
| if "taesd3" in encoder_path:return 16 |
| return 4 |
| |
| @staticmethod |
| def scale_latents(x): |
| return x.div(2 * Model.latent_magnitude).add(Model.latent_shift).clamp(0, 1) |
|
|
| @staticmethod |
| def unscale_latents(x): |
| return x.sub(Model.latent_shift).mul(2 * Model.latent_magnitude) |
|
|
| def forward(self, x, return_latent=False): |
| latent = self.encoder(x) |
| out = self.decoder(latent) |
| if return_latent: |
| return out.clamp(0, 1), latent |
| return out.clamp(0, 1) |
