(self,
ddconfig,
lossconfig,
embed_dim,
ckpt_path=None,
ignore_keys=[],
image_key="image",
colorize_nlabels=None,
monitor=None,
ema_decay=None,
learn_logvar=False,
)
| 17 | |
| 18 | class AutoencoderKL(pl.LightningModule): |
| 19 | def __init__(self, |
| 20 | ddconfig, |
| 21 | lossconfig, |
| 22 | embed_dim, |
| 23 | ckpt_path=None, |
| 24 | ignore_keys=[], |
| 25 | image_key="image", |
| 26 | colorize_nlabels=None, |
| 27 | monitor=None, |
| 28 | ema_decay=None, |
| 29 | learn_logvar=False, |
| 30 | ): |
| 31 | super().__init__() |
| 32 | self.learn_logvar = learn_logvar |
| 33 | self.image_key = image_key |
| 34 | self.encoder = Encoder(**ddconfig) |
| 35 | self.decoder = Decoder(**ddconfig) |
| 36 | self.loss = instantiate_from_config(lossconfig) |
| 37 | assert ddconfig["double_z"] |
| 38 | self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1) |
| 39 | self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) |
| 40 | self.embed_dim = embed_dim |
| 41 | if colorize_nlabels is not None: |
| 42 | assert type(colorize_nlabels)==int |
| 43 | self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1)) |
| 44 | if monitor is not None: |
| 45 | self.monitor = monitor |
| 46 | |
| 47 | self.use_ema = ema_decay is not None |
| 48 | if self.use_ema: |
| 49 | self.ema_decay = ema_decay |
| 50 | assert 0. < ema_decay < 1. |
| 51 | self.model_ema = LitEma(self, decay=ema_decay) |
| 52 | print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") |
| 53 | |
| 54 | if ckpt_path is not None: |
| 55 | self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) |
| 56 | |
| 57 | # self.register_buffer("mu_sig_bank", torch.empty(0)) |
| 58 | |
| 59 | def init_from_ckpt(self, path, ignore_keys=list()): |
| 60 | sd = torch.load(path, map_location="cpu")["state_dict"] |
no test coverage detected