i

sgm/models/__init__.py

@@ -0,0 +1,2 @@
+from .autoencoder import AutoencodingEngine
+from .diffusion import DiffusionEngine

sgm/models/autoencoder.py

@@ -0,0 +1,335 @@
+import re
+from abc import abstractmethod
+from contextlib import contextmanager
+from typing import Any, Dict, Tuple, Union
+
+import pytorch_lightning as pl
+import torch
+from omegaconf import ListConfig
+from packaging import version
+from safetensors.torch import load_file as load_safetensors
+
+from ..modules.diffusionmodules.model import Decoder, Encoder
+from ..modules.distributions.distributions import DiagonalGaussianDistribution
+from ..modules.ema import LitEma
+from ..util import default, get_obj_from_str, instantiate_from_config
+
+
+class AbstractAutoencoder(pl.LightningModule):
+    """
+    This is the base class for all autoencoders, including image autoencoders, image autoencoders with discriminators,
+    unCLIP models, etc. Hence, it is fairly general, and specific features
+    (e.g. discriminator training, encoding, decoding) must be implemented in subclasses.
+    """
+
+    def __init__(
+        self,
+        ema_decay: Union[None, float] = None,
+        monitor: Union[None, str] = None,
+        input_key: str = "jpg",
+        ckpt_path: Union[None, str] = None,
+        ignore_keys: Union[Tuple, list, ListConfig] = (),
+    ):
+        super().__init__()
+        self.input_key = input_key
+        self.use_ema = ema_decay is not None
+        if monitor is not None:
+            self.monitor = monitor
+
+        if self.use_ema:
+            self.model_ema = LitEma(self, decay=ema_decay)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+        if version.parse(torch.__version__) >= version.parse("2.0.0"):
+            self.automatic_optimization = False
+
+    def init_from_ckpt(
+        self, path: str, ignore_keys: Union[Tuple, list, ListConfig] = tuple()
+    ) -> None:
+        if path.endswith("ckpt"):
+            sd = torch.load(path, map_location="cpu")["state_dict"]
+        elif path.endswith("safetensors"):
+            sd = load_safetensors(path)
+        else:
+            raise NotImplementedError
+
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if re.match(ik, k):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False)
+        print(
+            f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys"
+        )
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    @abstractmethod
+    def get_input(self, batch) -> Any:
+        raise NotImplementedError()
+
+    def on_train_batch_end(self, *args, **kwargs):
+        # for EMA computation
+        if self.use_ema:
+            self.model_ema(self)
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.parameters())
+            self.model_ema.copy_to(self)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    @abstractmethod
+    def encode(self, *args, **kwargs) -> torch.Tensor:
+        raise NotImplementedError("encode()-method of abstract base class called")
+
+    @abstractmethod
+    def decode(self, *args, **kwargs) -> torch.Tensor:
+        raise NotImplementedError("decode()-method of abstract base class called")
+
+    def instantiate_optimizer_from_config(self, params, lr, cfg):
+        print(f"loading >>> {cfg['target']} <<< optimizer from config")
+        return get_obj_from_str(cfg["target"])(
+            params, lr=lr, **cfg.get("params", dict())
+        )
+
+    def configure_optimizers(self) -> Any:
+        raise NotImplementedError()
+
+
+class AutoencodingEngine(AbstractAutoencoder):
+    """
+    Base class for all image autoencoders that we train, like VQGAN or AutoencoderKL
+    (we also restore them explicitly as special cases for legacy reasons).
+    Regularizations such as KL or VQ are moved to the regularizer class.
+    """
+
+    def __init__(
+        self,
+        *args,
+        encoder_config: Dict,
+        decoder_config: Dict,
+        loss_config: Dict,
+        regularizer_config: Dict,
+        optimizer_config: Union[Dict, None] = None,
+        lr_g_factor: float = 1.0,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        # todo: add options to freeze encoder/decoder
+        self.encoder = instantiate_from_config(encoder_config)
+        self.decoder = instantiate_from_config(decoder_config)
+        self.loss = instantiate_from_config(loss_config)
+        self.regularization = instantiate_from_config(regularizer_config)
+        self.optimizer_config = default(
+            optimizer_config, {"target": "torch.optim.Adam"}
+        )
+        self.lr_g_factor = lr_g_factor
+
+    def get_input(self, batch: Dict) -> torch.Tensor:
+        # assuming unified data format, dataloader returns a dict.
+        # image tensors should be scaled to -1 ... 1 and in channels-first format (e.g., bchw instead if bhwc)
+        return batch[self.input_key]
+
+    def get_autoencoder_params(self) -> list:
+        params = (
+            list(self.encoder.parameters())
+            + list(self.decoder.parameters())
+            + list(self.regularization.get_trainable_parameters())
+            + list(self.loss.get_trainable_autoencoder_parameters())
+        )
+        return params
+
+    def get_discriminator_params(self) -> list:
+        params = list(self.loss.get_trainable_parameters())  # e.g., discriminator
+        return params
+
+    def get_last_layer(self):
+        return self.decoder.get_last_layer()
+
+    def encode(self, x: Any, return_reg_log: bool = False) -> Any:
+        z = self.encoder(x)
+        z, reg_log = self.regularization(z)
+        if return_reg_log:
+            return z, reg_log
+        return z
+
+    def decode(self, z: Any) -> torch.Tensor:
+        x = self.decoder(z)
+        return x
+
+    def forward(self, x: Any) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        z, reg_log = self.encode(x, return_reg_log=True)
+        dec = self.decode(z)
+        return z, dec, reg_log
+
+    def training_step(self, batch, batch_idx, optimizer_idx) -> Any:
+        x = self.get_input(batch)
+        z, xrec, regularization_log = self(x)
+
+        if optimizer_idx == 0:
+            # autoencode
+            aeloss, log_dict_ae = self.loss(
+                regularization_log,
+                x,
+                xrec,
+                optimizer_idx,
+                self.global_step,
+                last_layer=self.get_last_layer(),
+                split="train",
+            )
+
+            self.log_dict(
+                log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True
+            )
+            return aeloss
+
+        if optimizer_idx == 1:
+            # discriminator
+            discloss, log_dict_disc = self.loss(
+                regularization_log,
+                x,
+                xrec,
+                optimizer_idx,
+                self.global_step,
+                last_layer=self.get_last_layer(),
+                split="train",
+            )
+            self.log_dict(
+                log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True
+            )
+            return discloss
+
+    def validation_step(self, batch, batch_idx) -> Dict:
+        log_dict = self._validation_step(batch, batch_idx)
+        with self.ema_scope():
+            log_dict_ema = self._validation_step(batch, batch_idx, postfix="_ema")
+            log_dict.update(log_dict_ema)
+        return log_dict
+
+    def _validation_step(self, batch, batch_idx, postfix="") -> Dict:
+        x = self.get_input(batch)
+
+        z, xrec, regularization_log = self(x)
+        aeloss, log_dict_ae = self.loss(
+            regularization_log,
+            x,
+            xrec,
+            0,
+            self.global_step,
+            last_layer=self.get_last_layer(),
+            split="val" + postfix,
+        )
+
+        discloss, log_dict_disc = self.loss(
+            regularization_log,
+            x,
+            xrec,
+            1,
+            self.global_step,
+            last_layer=self.get_last_layer(),
+            split="val" + postfix,
+        )
+        self.log(f"val{postfix}/rec_loss", log_dict_ae[f"val{postfix}/rec_loss"])
+        log_dict_ae.update(log_dict_disc)
+        self.log_dict(log_dict_ae)
+        return log_dict_ae
+
+    def configure_optimizers(self) -> Any:
+        ae_params = self.get_autoencoder_params()
+        disc_params = self.get_discriminator_params()
+
+        opt_ae = self.instantiate_optimizer_from_config(
+            ae_params,
+            default(self.lr_g_factor, 1.0) * self.learning_rate,
+            self.optimizer_config,
+        )
+        opt_disc = self.instantiate_optimizer_from_config(
+            disc_params, self.learning_rate, self.optimizer_config
+        )
+
+        return [opt_ae, opt_disc], []
+
+    @torch.no_grad()
+    def log_images(self, batch: Dict, **kwargs) -> Dict:
+        log = dict()
+        x = self.get_input(batch)
+        _, xrec, _ = self(x)
+        log["inputs"] = x
+        log["reconstructions"] = xrec
+        with self.ema_scope():
+            _, xrec_ema, _ = self(x)
+            log["reconstructions_ema"] = xrec_ema
+        return log
+
+
+class AutoencoderKL(AutoencodingEngine):
+    def __init__(self, embed_dim: int, **kwargs):
+        ddconfig = kwargs.pop("ddconfig")
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", ())
+        super().__init__(
+            encoder_config={"target": "torch.nn.Identity"},
+            decoder_config={"target": "torch.nn.Identity"},
+            regularizer_config={"target": "torch.nn.Identity"},
+            loss_config=kwargs.pop("lossconfig"),
+            **kwargs,
+        )
+        assert ddconfig["double_z"]
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.quant_conv = torch.nn.Conv2d(2 * ddconfig["z_channels"], 2 * embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def encode(self, x):
+        assert (
+            not self.training
+        ), f"{self.__class__.__name__} only supports inference currently"
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z, **decoder_kwargs):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z, **decoder_kwargs)
+        return dec
+
+
+class AutoencoderKLInferenceWrapper(AutoencoderKL):
+    def encode(self, x):
+        return super().encode(x).sample()
+
+
+class IdentityFirstStage(AbstractAutoencoder):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def get_input(self, x: Any) -> Any:
+        return x
+
+    def encode(self, x: Any, *args, **kwargs) -> Any:
+        return x
+
+    def decode(self, x: Any, *args, **kwargs) -> Any:
+        return x

sgm/models/diffusion.py

@@ -0,0 +1,320 @@
+from contextlib import contextmanager
+from typing import Any, Dict, List, Tuple, Union
+
+import pytorch_lightning as pl
+import torch
+from omegaconf import ListConfig, OmegaConf
+from safetensors.torch import load_file as load_safetensors
+from torch.optim.lr_scheduler import LambdaLR
+
+from ..modules import UNCONDITIONAL_CONFIG
+from ..modules.diffusionmodules.wrappers import OPENAIUNETWRAPPER
+from ..modules.ema import LitEma
+from ..util import (
+    default,
+    disabled_train,
+    get_obj_from_str,
+    instantiate_from_config,
+    log_txt_as_img,
+)
+
+
+class DiffusionEngine(pl.LightningModule):
+    def __init__(
+        self,
+        network_config,
+        denoiser_config,
+        first_stage_config,
+        conditioner_config: Union[None, Dict, ListConfig, OmegaConf] = None,
+        sampler_config: Union[None, Dict, ListConfig, OmegaConf] = None,
+        optimizer_config: Union[None, Dict, ListConfig, OmegaConf] = None,
+        scheduler_config: Union[None, Dict, ListConfig, OmegaConf] = None,
+        loss_fn_config: Union[None, Dict, ListConfig, OmegaConf] = None,
+        network_wrapper: Union[None, str] = None,
+        ckpt_path: Union[None, str] = None,
+        use_ema: bool = False,
+        ema_decay_rate: float = 0.9999,
+        scale_factor: float = 1.0,
+        disable_first_stage_autocast=False,
+        input_key: str = "jpg",
+        log_keys: Union[List, None] = None,
+        no_cond_log: bool = False,
+        compile_model: bool = False,
+    ):
+        super().__init__()
+        self.log_keys = log_keys
+        self.input_key = input_key
+        self.optimizer_config = default(
+            optimizer_config, {"target": "torch.optim.AdamW"}
+        )
+        model = instantiate_from_config(network_config)
+        self.model = get_obj_from_str(default(network_wrapper, OPENAIUNETWRAPPER))(
+            model, compile_model=compile_model
+        )
+
+        self.denoiser = instantiate_from_config(denoiser_config)
+        self.sampler = (
+            instantiate_from_config(sampler_config)
+            if sampler_config is not None
+            else None
+        )
+        self.conditioner = instantiate_from_config(
+            default(conditioner_config, UNCONDITIONAL_CONFIG)
+        )
+        self.scheduler_config = scheduler_config
+        self._init_first_stage(first_stage_config)
+
+        self.loss_fn = (
+            instantiate_from_config(loss_fn_config)
+            if loss_fn_config is not None
+            else None
+        )
+
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self.model, decay=ema_decay_rate)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        self.scale_factor = scale_factor
+        self.disable_first_stage_autocast = disable_first_stage_autocast
+        self.no_cond_log = no_cond_log
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path)
+
+    def init_from_ckpt(
+        self,
+        path: str,
+    ) -> None:
+        if path.endswith("ckpt"):
+            sd = torch.load(path, map_location="cpu")["state_dict"]
+        elif path.endswith("safetensors"):
+            sd = load_safetensors(path)
+        else:
+            raise NotImplementedError
+
+        missing, unexpected = self.load_state_dict(sd, strict=False)
+        print(
+            f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys"
+        )
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    def _init_first_stage(self, config):
+        model = instantiate_from_config(config).eval()
+        model.train = disabled_train
+        for param in model.parameters():
+            param.requires_grad = False
+        self.first_stage_model = model
+
+    def get_input(self, batch):
+        # assuming unified data format, dataloader returns a dict.
+        # image tensors should be scaled to -1 ... 1 and in bchw format
+        return batch[self.input_key]
+
+    @torch.no_grad()
+    def decode_first_stage(self, z):
+        z = 1.0 / self.scale_factor * z
+        with torch.autocast("cuda", enabled=not self.disable_first_stage_autocast):
+            out = self.first_stage_model.decode(z)
+        return out
+
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        with torch.autocast("cuda", enabled=not self.disable_first_stage_autocast):
+            z = self.first_stage_model.encode(x)
+        z = self.scale_factor * z
+        return z
+
+    def forward(self, x, batch):
+        loss = self.loss_fn(self.model, self.denoiser, self.conditioner, x, batch)
+        loss_mean = loss.mean()
+        loss_dict = {"loss": loss_mean}
+        return loss_mean, loss_dict
+
+    def shared_step(self, batch: Dict) -> Any:
+        x = self.get_input(batch)
+        x = self.encode_first_stage(x)
+        batch["global_step"] = self.global_step
+        loss, loss_dict = self(x, batch)
+        return loss, loss_dict
+
+    def training_step(self, batch, batch_idx):
+        loss, loss_dict = self.shared_step(batch)
+
+        self.log_dict(
+            loss_dict, prog_bar=True, logger=True, on_step=True, on_epoch=False
+        )
+
+        self.log(
+            "global_step",
+            self.global_step,
+            prog_bar=True,
+            logger=True,
+            on_step=True,
+            on_epoch=False,
+        )
+
+        if self.scheduler_config is not None:
+            lr = self.optimizers().param_groups[0]["lr"]
+            self.log(
+                "lr_abs", lr, prog_bar=True, logger=True, on_step=True, on_epoch=False
+            )
+
+        return loss
+
+    def on_train_start(self, *args, **kwargs):
+        if self.sampler is None or self.loss_fn is None:
+            raise ValueError("Sampler and loss function need to be set for training.")
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self.model)
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.model.parameters())
+            self.model_ema.copy_to(self.model)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.model.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def instantiate_optimizer_from_config(self, params, lr, cfg):
+        return get_obj_from_str(cfg["target"])(
+            params, lr=lr, **cfg.get("params", dict())
+        )
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        for embedder in self.conditioner.embedders:
+            if embedder.is_trainable:
+                params = params + list(embedder.parameters())
+        opt = self.instantiate_optimizer_from_config(params, lr, self.optimizer_config)
+        if self.scheduler_config is not None:
+            scheduler = instantiate_from_config(self.scheduler_config)
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    "scheduler": LambdaLR(opt, lr_lambda=scheduler.schedule),
+                    "interval": "step",
+                    "frequency": 1,
+                }
+            ]
+            return [opt], scheduler
+        return opt
+
+    @torch.no_grad()
+    def sample(
+        self,
+        cond: Dict,
+        uc: Union[Dict, None] = None,
+        batch_size: int = 16,
+        shape: Union[None, Tuple, List] = None,
+        **kwargs,
+    ):
+        randn = torch.randn(batch_size, *shape).to(self.device)
+
+        denoiser = lambda input, sigma, c: self.denoiser(
+            self.model, input, sigma, c, **kwargs
+        )
+        samples = self.sampler(denoiser, randn, cond, uc=uc)
+        return samples
+
+    @torch.no_grad()
+    def log_conditionings(self, batch: Dict, n: int) -> Dict:
+        """
+        Defines heuristics to log different conditionings.
+        These can be lists of strings (text-to-image), tensors, ints, ...
+        """
+        image_h, image_w = batch[self.input_key].shape[2:]
+        log = dict()
+
+        for embedder in self.conditioner.embedders:
+            if (
+                (self.log_keys is None) or (embedder.input_key in self.log_keys)
+            ) and not self.no_cond_log:
+                x = batch[embedder.input_key][:n]
+                if isinstance(x, torch.Tensor):
+                    if x.dim() == 1:
+                        # class-conditional, convert integer to string
+                        x = [str(x[i].item()) for i in range(x.shape[0])]
+                        xc = log_txt_as_img((image_h, image_w), x, size=image_h // 4)
+                    elif x.dim() == 2:
+                        # size and crop cond and the like
+                        x = [
+                            "x".join([str(xx) for xx in x[i].tolist()])
+                            for i in range(x.shape[0])
+                        ]
+                        xc = log_txt_as_img((image_h, image_w), x, size=image_h // 20)
+                    else:
+                        raise NotImplementedError()
+                elif isinstance(x, (List, ListConfig)):
+                    if isinstance(x[0], str):
+                        # strings
+                        xc = log_txt_as_img((image_h, image_w), x, size=image_h // 20)
+                    else:
+                        raise NotImplementedError()
+                else:
+                    raise NotImplementedError()
+                log[embedder.input_key] = xc
+        return log
+
+    @torch.no_grad()
+    def log_images(
+        self,
+        batch: Dict,
+        N: int = 8,
+        sample: bool = True,
+        ucg_keys: List[str] = None,
+        **kwargs,
+    ) -> Dict:
+        conditioner_input_keys = [e.input_key for e in self.conditioner.embedders]
+        if ucg_keys:
+            assert all(map(lambda x: x in conditioner_input_keys, ucg_keys)), (
+                "Each defined ucg key for sampling must be in the provided conditioner input keys,"
+                f"but we have {ucg_keys} vs. {conditioner_input_keys}"
+            )
+        else:
+            ucg_keys = conditioner_input_keys
+        log = dict()
+
+        x = self.get_input(batch)
+
+        c, uc = self.conditioner.get_unconditional_conditioning(
+            batch,
+            force_uc_zero_embeddings=ucg_keys
+            if len(self.conditioner.embedders) > 0
+            else [],
+        )
+
+        sampling_kwargs = {}
+
+        N = min(x.shape[0], N)
+        x = x.to(self.device)[:N]
+        log["inputs"] = x
+        z = self.encode_first_stage(x)
+        log["reconstructions"] = self.decode_first_stage(z)
+        log.update(self.log_conditionings(batch, N))
+
+        for k in c:
+            if isinstance(c[k], torch.Tensor):
+                c[k], uc[k] = map(lambda y: y[k][:N].to(self.device), (c, uc))
+
+        if sample:
+            with self.ema_scope("Plotting"):
+                samples = self.sample(
+                    c, shape=z.shape[1:], uc=uc, batch_size=N, **sampling_kwargs
+                )
+            samples = self.decode_first_stage(samples)
+            log["samples"] = samples
+        return log