imaginAIry/imaginairy/modules/diffusion/ddpm.py

"""
wild mixture of
https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
https://github.com/CompVis/taming-transformers
-- merci
"""
import logging
from functools import partial

import numpy as np
import pytorch_lightning as pl
import torch
import torch.nn as nn
from einops import rearrange
from torchvision.utils import make_grid
from tqdm import tqdm

from imaginairy.modules.diffusion.util import (
    extract_into_tensor,
    make_beta_schedule,
    noise_like,
)
from imaginairy.modules.distributions import DiagonalGaussianDistribution
from imaginairy.utils import instantiate_from_config, log_params

logger = logging.getLogger(__name__)
__conditioning_keys__ = {"concat": "c_concat", "crossattn": "c_crossattn", "adm": "y"}


def disabled_train(self, mode=True):
    """Overwrite model.train with this function to make sure train/eval mode
    does not change anymore."""
    return self


def uniform_on_device(r1, r2, shape, device):
    return (r1 - r2) * torch.rand(*shape, device=device) + r2


class DDPM(pl.LightningModule):
    """
    classic DDPM with Gaussian diffusion, in image space

    Denoising diffusion probabilistic models
    """

    def __init__(
        self,
        unet_config,
        timesteps=1000,
        beta_schedule="linear",
        loss_type="l2",
        ckpt_path=None,
        ignore_keys=[],
        load_only_unet=False,
        monitor="val/loss",
        first_stage_key="image",
        image_size=256,
        channels=3,
        log_every_t=100,
        clip_denoised=True,
        linear_start=1e-4,
        linear_end=2e-2,
        cosine_s=8e-3,
        given_betas=None,
        original_elbo_weight=0.0,
        v_posterior=0.0,  # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
        l_simple_weight=1.0,
        conditioning_key=None,
        parameterization="eps",  # all assuming fixed variance schedules
        scheduler_config=None,
        use_positional_encodings=False,
        learn_logvar=False,
        logvar_init=0.0,
    ):
        super().__init__()
        assert parameterization in [
            "eps",
            "x0",
        ], 'currently only supporting "eps" and "x0"'
        self.parameterization = parameterization
        logger.debug(
            f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode"
        )
        self.cond_stage_model = None
        self.clip_denoised = clip_denoised
        self.log_every_t = log_every_t
        self.first_stage_key = first_stage_key
        self.image_size = image_size  # try conv?
        self.channels = channels
        self.use_positional_encodings = use_positional_encodings
        self.model = DiffusionWrapper(unet_config, conditioning_key)
        log_params(self.model)

        self.use_scheduler = scheduler_config is not None
        if self.use_scheduler:
            self.scheduler_config = scheduler_config

        self.v_posterior = v_posterior
        self.original_elbo_weight = original_elbo_weight
        self.l_simple_weight = l_simple_weight

        if monitor is not None:
            self.monitor = monitor
        if ckpt_path is not None:
            self.init_from_ckpt(
                ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet
            )

        self.register_schedule(
            given_betas=given_betas,
            beta_schedule=beta_schedule,
            timesteps=timesteps,
            linear_start=linear_start,
            linear_end=linear_end,
            cosine_s=cosine_s,
        )

        self.loss_type = loss_type

        self.learn_logvar = learn_logvar
        self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
        if self.learn_logvar:
            self.logvar = nn.Parameter(self.logvar, requires_grad=True)

    def register_schedule(
        self,
        given_betas=None,
        beta_schedule="linear",
        timesteps=1000,
        linear_start=1e-4,
        linear_end=2e-2,
        cosine_s=8e-3,
    ):
        if given_betas is not None:
            betas = given_betas
        else:
            betas = make_beta_schedule(
                beta_schedule,
                timesteps,
                linear_start=linear_start,
                linear_end=linear_end,
                cosine_s=cosine_s,
            )
        alphas = 1.0 - betas
        alphas_cumprod = np.cumprod(alphas, axis=0)
        alphas_cumprod_prev = np.append(1.0, alphas_cumprod[:-1])

        (timesteps,) = betas.shape
        self.num_timesteps = int(timesteps)
        self.linear_start = linear_start
        self.linear_end = linear_end
        assert (
            alphas_cumprod.shape[0] == self.num_timesteps
        ), "alphas have to be defined for each timestep"

        to_torch = partial(torch.tensor, dtype=torch.float32)

        self.register_buffer("betas", to_torch(betas))
        self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod))
        self.register_buffer("alphas_cumprod_prev", to_torch(alphas_cumprod_prev))

        # calculations for diffusion q(x_t | x_{t-1}) and others
        self.register_buffer("sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod)))
        self.register_buffer(
            "sqrt_one_minus_alphas_cumprod", to_torch(np.sqrt(1.0 - alphas_cumprod))
        )
        self.register_buffer(
            "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod))
        )
        self.register_buffer(
            "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod))
        )
        self.register_buffer(
            "sqrt_recipm1_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod - 1))
        )

        # calculations for posterior q(x_{t-1} | x_t, x_0)
        posterior_variance = (1 - self.v_posterior) * betas * (
            1.0 - alphas_cumprod_prev
        ) / (1.0 - alphas_cumprod) + self.v_posterior * betas
        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
        self.register_buffer("posterior_variance", to_torch(posterior_variance))
        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
        self.register_buffer(
            "posterior_log_variance_clipped",
            to_torch(np.log(np.maximum(posterior_variance, 1e-20))),
        )
        self.register_buffer(
            "posterior_mean_coef1",
            to_torch(betas * np.sqrt(alphas_cumprod_prev) / (1.0 - alphas_cumprod)),
        )
        self.register_buffer(
            "posterior_mean_coef2",
            to_torch(
                (1.0 - alphas_cumprod_prev) * np.sqrt(alphas) / (1.0 - alphas_cumprod)
            ),
        )

        if self.parameterization == "eps":
            lvlb_weights = self.betas**2 / (
                2
                * self.posterior_variance
                * to_torch(alphas)
                * (1 - self.alphas_cumprod)
            )
        elif self.parameterization == "x0":
            lvlb_weights = (
                0.5
                * np.sqrt(torch.Tensor(alphas_cumprod))
                / (2.0 * 1 - torch.Tensor(alphas_cumprod))
            )
        else:
            raise NotImplementedError("mu not supported")
        # TODO how to choose this term
        lvlb_weights[0] = lvlb_weights[1]
        self.register_buffer("lvlb_weights", lvlb_weights, persistent=False)
        assert not torch.isnan(self.lvlb_weights).all()


class LatentDiffusion(DDPM):
    """main class"""

    def __init__(
        self,
        first_stage_config,
        cond_stage_config,
        num_timesteps_cond=None,
        cond_stage_key="image",
        cond_stage_trainable=False,
        concat_mode=True,
        cond_stage_forward=None,
        conditioning_key=None,
        scale_factor=1.0,
        scale_by_std=False,
        *args,
        **kwargs,
    ):
        self.num_timesteps_cond = (
            1 if num_timesteps_cond is None else num_timesteps_cond
        )
        self.scale_by_std = scale_by_std
        assert self.num_timesteps_cond <= kwargs["timesteps"]
        # for backwards compatibility after implementation of DiffusionWrapper
        if conditioning_key is None:
            conditioning_key = "concat" if concat_mode else "crossattn"
        if cond_stage_config == "__is_unconditional__":
            conditioning_key = None
        ckpt_path = kwargs.pop("ckpt_path", None)
        ignore_keys = kwargs.pop("ignore_keys", [])
        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
        self.concat_mode = concat_mode
        self.cond_stage_trainable = cond_stage_trainable
        self.cond_stage_key = cond_stage_key
        try:
            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
        except:  # noqa
            logger.exception("Bad num downs?")
            self.num_downs = 0
        if not scale_by_std:
            self.scale_factor = scale_factor
        else:
            self.register_buffer("scale_factor", torch.tensor(scale_factor))
        self.instantiate_first_stage(first_stage_config)
        self.instantiate_cond_stage(cond_stage_config)
        self.cond_stage_forward = cond_stage_forward
        self.clip_denoised = False
        self.bbox_tokenizer = None

        self.restarted_from_ckpt = False
        if ckpt_path is not None:
            self.init_from_ckpt(ckpt_path, ignore_keys)
            self.restarted_from_ckpt = True

    def make_cond_schedule(
        self,
    ):
        self.cond_ids = torch.full(
            size=(self.num_timesteps,),
            fill_value=self.num_timesteps - 1,
            dtype=torch.long,
        )
        ids = torch.round(
            torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)
        ).long()
        self.cond_ids[: self.num_timesteps_cond] = ids

    def register_schedule(
        self,
        given_betas=None,
        beta_schedule="linear",
        timesteps=1000,
        linear_start=1e-4,
        linear_end=2e-2,
        cosine_s=8e-3,
    ):
        super().register_schedule(
            given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s
        )

        self.shorten_cond_schedule = self.num_timesteps_cond > 1
        if self.shorten_cond_schedule:
            self.make_cond_schedule()

    def instantiate_first_stage(self, config):
        model = instantiate_from_config(config)
        self.first_stage_model = model.eval()
        self.first_stage_model.train = disabled_train
        for param in self.first_stage_model.parameters():
            param.requires_grad = False

    def instantiate_cond_stage(self, config):
        if not self.cond_stage_trainable:
            if config == "__is_first_stage__":
                logger.debug("Using first stage also as cond stage.")
                self.cond_stage_model = self.first_stage_model
            elif config == "__is_unconditional__":
                logger.debug(
                    f"Training {self.__class__.__name__} as an unconditional model."
                )
                self.cond_stage_model = None
                # self.be_unconditional = True
            else:
                model = instantiate_from_config(config)
                self.cond_stage_model = model.eval()
                self.cond_stage_model.train = disabled_train
                for param in self.cond_stage_model.parameters():
                    param.requires_grad = False
        else:
            assert config != "__is_first_stage__"
            assert config != "__is_unconditional__"
            model = instantiate_from_config(config)
            self.cond_stage_model = model

    def _get_denoise_row_from_list(
        self, samples, desc="", force_no_decoder_quantization=False
    ):
        denoise_row = []
        for zd in tqdm(samples, desc=desc):
            denoise_row.append(
                self.decode_first_stage(
                    zd.to(self.device), force_not_quantize=force_no_decoder_quantization
                )
            )
        n_imgs_per_row = len(denoise_row)
        denoise_row = torch.stack(denoise_row)  # n_log_step, n_row, C, H, W
        denoise_grid = rearrange(denoise_row, "n b c h w -> b n c h w")
        denoise_grid = rearrange(denoise_grid, "b n c h w -> (b n) c h w")
        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
        return denoise_grid

    def get_first_stage_encoding(self, encoder_posterior):
        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
            z = encoder_posterior.sample()
        elif isinstance(encoder_posterior, torch.Tensor):
            z = encoder_posterior
        else:
            raise NotImplementedError(
                f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented"
            )
        return self.scale_factor * z

    def get_learned_conditioning(self, c):
        if self.cond_stage_forward is None:
            if hasattr(self.cond_stage_model, "encode") and callable(
                self.cond_stage_model.encode
            ):
                c = self.cond_stage_model.encode(c)
                if isinstance(c, DiagonalGaussianDistribution):
                    c = c.mode()
            else:
                c = self.cond_stage_model(c)
        else:
            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
        return c

    def meshgrid(self, h, w):
        y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
        x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)

        arr = torch.cat([y, x], dim=-1)
        return arr

    def delta_border(self, h, w):
        """
        :param h: height
        :param w: width
        :return: normalized distance to image border,
         wtith min distance = 0 at border and max dist = 0.5 at image center
        """
        lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
        arr = self.meshgrid(h, w) / lower_right_corner
        dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
        dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
        edge_dist = torch.min(
            torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1
        )[0]
        return edge_dist

    def get_weighting(self, h, w, Ly, Lx, device):
        weighting = self.delta_border(h, w)
        weighting = torch.clip(
            weighting,
            self.split_input_params["clip_min_weight"],
            self.split_input_params["clip_max_weight"],
        )
        weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)

        if self.split_input_params["tie_braker"]:
            L_weighting = self.delta_border(Ly, Lx)
            L_weighting = torch.clip(
                L_weighting,
                self.split_input_params["clip_min_tie_weight"],
                self.split_input_params["clip_max_tie_weight"],
            )

            L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
            weighting = weighting * L_weighting
        return weighting

    def get_fold_unfold(
        self, x, kernel_size, stride, uf=1, df=1
    ):  # todo load once not every time, shorten code
        """
        :param x: img of size (bs, c, h, w)
        :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
        """
        bs, nc, h, w = x.shape

        # number of crops in image
        Ly = (h - kernel_size[0]) // stride[0] + 1
        Lx = (w - kernel_size[1]) // stride[1] + 1

        if uf == 1 and df == 1:
            fold_params = dict(
                kernel_size=kernel_size, dilation=1, padding=0, stride=stride
            )
            unfold = torch.nn.Unfold(**fold_params)

            fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)

            weighting = self.get_weighting(
                kernel_size[0], kernel_size[1], Ly, Lx, x.device
            ).to(x.dtype)
            normalization = fold(weighting).view(1, 1, h, w)  # normalizes the overlap
            weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))

        elif uf > 1 and df == 1:
            fold_params = dict(
                kernel_size=kernel_size, dilation=1, padding=0, stride=stride
            )
            unfold = torch.nn.Unfold(**fold_params)

            fold_params2 = dict(
                kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
                dilation=1,
                padding=0,
                stride=(stride[0] * uf, stride[1] * uf),
            )
            fold = torch.nn.Fold(
                output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2
            )

            weighting = self.get_weighting(
                kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device
            ).to(x.dtype)
            normalization = fold(weighting).view(
                1, 1, h * uf, w * uf
            )  # normalizes the overlap
            weighting = weighting.view(
                (1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx)
            )

        elif df > 1 and uf == 1:
            fold_params = dict(
                kernel_size=kernel_size, dilation=1, padding=0, stride=stride
            )
            unfold = torch.nn.Unfold(**fold_params)

            fold_params2 = dict(
                kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
                dilation=1,
                padding=0,
                stride=(stride[0] // df, stride[1] // df),
            )
            fold = torch.nn.Fold(
                output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2
            )

            weighting = self.get_weighting(
                kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device
            ).to(x.dtype)
            normalization = fold(weighting).view(
                1, 1, h // df, w // df
            )  # normalizes the overlap
            weighting = weighting.view(
                (1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx)
            )

        else:
            raise NotImplementedError

        return fold, unfold, normalization, weighting

    @torch.no_grad()
    def get_input(
        self,
        batch,
        k,
        return_first_stage_outputs=False,
        force_c_encode=False,
        cond_key=None,
        return_original_cond=False,
        bs=None,
    ):
        x = super().get_input(batch, k)
        if bs is not None:
            x = x[:bs]
        x = x.to(self.device)
        encoder_posterior = self.encode_first_stage(x)
        z = self.get_first_stage_encoding(encoder_posterior).detach()

        if self.model.conditioning_key is not None:
            if cond_key is None:
                cond_key = self.cond_stage_key
            if cond_key != self.first_stage_key:
                if cond_key in ["caption", "coordinates_bbox"]:
                    xc = batch[cond_key]
                elif cond_key == "class_label":
                    xc = batch
                else:
                    xc = super().get_input(batch, cond_key).to(self.device)
            else:
                xc = x
            if not self.cond_stage_trainable or force_c_encode:
                if isinstance(xc, dict) or isinstance(xc, list):
                    # import pudb; pudb.set_trace()
                    c = self.get_learned_conditioning(xc)
                else:
                    c = self.get_learned_conditioning(xc.to(self.device))
            else:
                c = xc
            if bs is not None:
                c = c[:bs]

            if self.use_positional_encodings:
                pos_x, pos_y = self.compute_latent_shifts(batch)
                ckey = __conditioning_keys__[self.model.conditioning_key]
                c = {ckey: c, "pos_x": pos_x, "pos_y": pos_y}

        else:
            c = None
            xc = None
            if self.use_positional_encodings:
                pos_x, pos_y = self.compute_latent_shifts(batch)
                c = {"pos_x": pos_x, "pos_y": pos_y}
        out = [z, c]
        if return_first_stage_outputs:
            xrec = self.decode_first_stage(z)
            out.extend([x, xrec])
        if return_original_cond:
            out.append(xc)
        return out

    @torch.no_grad()
    def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
        if predict_cids:
            if z.dim() == 4:
                z = torch.argmax(z.exp(), dim=1).long()
            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
            z = rearrange(z, "b h w c -> b c h w").contiguous()

        z = 1.0 / self.scale_factor * z

        return self.first_stage_model.decode(z)

    @torch.no_grad()
    def encode_first_stage(self, x):
        if (
            hasattr(self, "split_input_params")
            and self.split_input_params["patch_distributed_vq"]
        ):
            ks = self.split_input_params["ks"]  # eg. (128, 128)
            stride = self.split_input_params["stride"]  # eg. (64, 64)
            df = self.split_input_params["vqf"]
            self.split_input_params["original_image_size"] = x.shape[-2:]
            bs, nc, h, w = x.shape
            if ks[0] > h or ks[1] > w:
                ks = (min(ks[0], h), min(ks[1], w))
                logger.info("reducing Kernel")

            if stride[0] > h or stride[1] > w:
                stride = (min(stride[0], h), min(stride[1], w))
                logger.info("reducing stride")

            fold, unfold, normalization, weighting = self.get_fold_unfold(
                x, ks, stride, df=df
            )
            z = unfold(x)  # (bn, nc * prod(**ks), L)
            # Reshape to img shape
            z = z.view(
                (z.shape[0], -1, ks[0], ks[1], z.shape[-1])
            )  # (bn, nc, ks[0], ks[1], L )

            output_list = [
                self.first_stage_model.encode(z[:, :, :, :, i])
                for i in range(z.shape[-1])
            ]

            o = torch.stack(output_list, axis=-1)
            o = o * weighting

            # Reverse reshape to img shape
            o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
            # stitch crops together
            decoded = fold(o)
            decoded = decoded / normalization
            return decoded

        return self.first_stage_model.encode(x)

    def apply_model(self, x_noisy, t, cond, return_ids=False):

        if isinstance(cond, dict):
            # hybrid case, cond is exptected to be a dict
            pass
        else:
            if not isinstance(cond, list):
                cond = [cond]
            key = (
                "c_concat" if self.model.conditioning_key == "concat" else "c_crossattn"
            )
            cond = {key: cond}

        if hasattr(self, "split_input_params"):
            assert len(cond) == 1  # todo can only deal with one conditioning atm
            assert not return_ids
            ks = self.split_input_params["ks"]  # eg. (128, 128)
            stride = self.split_input_params["stride"]  # eg. (64, 64)

            h, w = x_noisy.shape[-2:]  # noqa

            fold, unfold, normalization, weighting = self.get_fold_unfold(
                x_noisy, ks, stride
            )

            z = unfold(x_noisy)  # (bn, nc * prod(**ks), L)
            # Reshape to img shape
            z = z.view(
                (z.shape[0], -1, ks[0], ks[1], z.shape[-1])
            )  # (bn, nc, ks[0], ks[1], L )
            z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]

            if (
                self.cond_stage_key in ["image", "LR_image", "segmentation", "bbox_img"]
                and self.model.conditioning_key
            ):  # todo check for completeness
                c_key = next(iter(cond.keys()))  # get key
                c = next(iter(cond.values()))  # get value
                assert len(c) == 1  # todo extend to list with more than one elem
                c = c[0]  # get element

                c = unfold(c)
                c = c.view(
                    (c.shape[0], -1, ks[0], ks[1], c.shape[-1])
                )  # (bn, nc, ks[0], ks[1], L )

                cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]

            elif self.cond_stage_key == "coordinates_bbox":
                assert (
                    "original_image_size" in self.split_input_params
                ), "BoudingBoxRescaling is missing original_image_size"

                # assuming padding of unfold is always 0 and its dilation is always 1
                n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
                full_img_h, full_img_w = self.split_input_params["original_image_size"]
                # as we are operating on latents, we need the factor from the original image size to the
                # spatial latent size to properly rescale the crops for regenerating the bbox annotations
                num_downs = self.first_stage_model.encoder.num_resolutions - 1
                rescale_latent = 2 ** (num_downs)

                # get top left postions of patches as conforming for the bbbox tokenizer, therefore we
                # need to rescale the tl patch coordinates to be in between (0,1)
                tl_patch_coordinates = [
                    (
                        rescale_latent
                        * stride[0]
                        * (patch_nr % n_patches_per_row)
                        / full_img_w,
                        rescale_latent
                        * stride[1]
                        * (patch_nr // n_patches_per_row)
                        / full_img_h,
                    )
                    for patch_nr in range(z.shape[-1])
                ]

                # patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
                patch_limits = [
                    (
                        x_tl,
                        y_tl,
                        rescale_latent * ks[0] / full_img_w,
                        rescale_latent * ks[1] / full_img_h,
                    )
                    for x_tl, y_tl in tl_patch_coordinates
                ]
                # patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]

                # tokenize crop coordinates for the bounding boxes of the respective patches
                patch_limits_tknzd = [
                    torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[  # noqa
                        None
                    ].to(  # noqa
                        self.device
                    )
                    for bbox in patch_limits
                ]  # list of length l with tensors of shape (1, 2)

                # cut tknzd crop position from conditioning
                assert isinstance(cond, dict), "cond must be dict to be fed into model"
                cut_cond = cond["c_crossattn"][0][..., :-2].to(self.device)

                adapted_cond = torch.stack(
                    [torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd]
                )
                adapted_cond = rearrange(adapted_cond, "l b n -> (l b) n")

                adapted_cond = self.get_learned_conditioning(adapted_cond)

                adapted_cond = rearrange(
                    adapted_cond, "(l b) n d -> l b n d", l=z.shape[-1]
                )

                cond_list = [{"c_crossattn": [e]} for e in adapted_cond]

            else:
                cond_list = [
                    cond for i in range(z.shape[-1])
                ]  # Todo make this more efficient

            # apply model by loop over crops
            output_list = [
                self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])
            ]
            assert not isinstance(
                output_list[0], tuple
            )  # todo cant deal with multiple model outputs check this never happens

            o = torch.stack(output_list, axis=-1)
            o = o * weighting
            # Reverse reshape to img shape
            o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
            # stitch crops together
            x_recon = fold(o) / normalization

        else:
            x_recon = self.model(x_noisy, t, **cond)

        if isinstance(x_recon, tuple) and not return_ids:
            return x_recon[0]

        return x_recon

    def p_mean_variance(
        self,
        x,
        c,
        t,
        clip_denoised: bool,
        return_codebook_ids=False,
        quantize_denoised=False,
        return_x0=False,
        score_corrector=None,
        corrector_kwargs=None,
    ):
        t_in = t
        model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)

        if score_corrector is not None:
            assert self.parameterization == "eps"
            model_out = score_corrector.modify_score(
                self, model_out, x, t, c, **corrector_kwargs
            )

        if return_codebook_ids:
            model_out, logits = model_out

        if self.parameterization == "eps":
            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
        elif self.parameterization == "x0":
            x_recon = model_out
        else:
            raise NotImplementedError()

        if clip_denoised:
            x_recon.clamp_(-1.0, 1.0)
        if quantize_denoised:
            x_recon, _, _ = self.first_stage_model.quantize(x_recon)
        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(
            x_start=x_recon, x_t=x, t=t
        )
        if return_codebook_ids:
            return model_mean, posterior_variance, posterior_log_variance, logits
        if return_x0:
            return model_mean, posterior_variance, posterior_log_variance, x_recon

        return model_mean, posterior_variance, posterior_log_variance

    @torch.no_grad()
    def p_sample(
        self,
        x,
        c,
        t,
        clip_denoised=False,
        repeat_noise=False,
        return_codebook_ids=False,
        quantize_denoised=False,
        return_x0=False,
        temperature=1.0,
        noise_dropout=0.0,
        score_corrector=None,
        corrector_kwargs=None,
    ):
        b, *_, device = *x.shape, x.device
        outputs = self.p_mean_variance(
            x=x,
            c=c,
            t=t,
            clip_denoised=clip_denoised,
            return_codebook_ids=return_codebook_ids,
            quantize_denoised=quantize_denoised,
            return_x0=return_x0,
            score_corrector=score_corrector,
            corrector_kwargs=corrector_kwargs,
        )
        if return_x0:
            model_mean, _, model_log_variance, x0 = outputs
        else:
            model_mean, _, model_log_variance = outputs

        noise = noise_like(x.shape, device, repeat_noise) * temperature
        if noise_dropout > 0.0:
            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
        # no noise when t == 0
        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))

        if return_x0:
            return (
                model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise,
                x0,
            )

        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise

    def q_sample(self, x_start, t, noise=None):
        noise = (
            noise
            if noise is not None
            else torch.randn_like(x_start, device="cpu").to(x_start.device)
        )
        return (
            extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
            + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape)
            * noise
        )


class DiffusionWrapper(pl.LightningModule):
    def __init__(self, diff_model_config, conditioning_key):
        super().__init__()
        self.diffusion_model = instantiate_from_config(diff_model_config)
        self.conditioning_key = conditioning_key
        assert self.conditioning_key in [None, "concat", "crossattn", "hybrid", "adm"]

    def forward(self, x, t, c_concat: list = None, c_crossattn: list = None):
        if self.conditioning_key is None:
            out = self.diffusion_model(x, t)
        elif self.conditioning_key == "concat":
            xc = torch.cat([x] + c_concat, dim=1)
            out = self.diffusion_model(xc, t)
        elif self.conditioning_key == "crossattn":
            cc = torch.cat(c_crossattn, 1)
            out = self.diffusion_model(x, t, context=cc)
        elif self.conditioning_key == "hybrid":
            xc = torch.cat([x] + c_concat, dim=1)
            cc = torch.cat(c_crossattn, 1)
            out = self.diffusion_model(xc, t, context=cc)
        elif self.conditioning_key == "adm":
            cc = c_crossattn[0]
            out = self.diffusion_model(x, t, y=cc)
        else:
            raise NotImplementedError()

        return out