latent diffusion 原理+代码

latent diffusion - Github

以下代码来自
作者： 李宝璐
链接： https://libaolu312.github.io/2023/11/27/Latent-Diffusion-Models-原理和代码/
版权声明： 本博客所有文章除特别声明外，均采用 MIT 许可协议。转载请注明出处！

核心代码：

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 = default(num_timesteps_cond, 1)
        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:
            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
 
    @rank_zero_only
    @torch.no_grad()
    def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
        # only for very first batch
        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
            assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
            # set rescale weight to 1./std of encodings
            print("### USING STD-RESCALING ###")
            x = super().get_input(batch, self.first_stage_key)
            x = x.to(self.device)
            encoder_posterior = self.encode_first_stage(x)
            z = self.get_first_stage_encoding(encoder_posterior).detach()
            del self.scale_factor
            self.register_buffer('scale_factor', 1. / z.flatten().std())
            print(f"setting self.scale_factor to {
     self.scale_factor}")
            print("### USING STD-RESCALING ###")
 
    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__":
                print("Using first stage also as cond stage.")
                self.cond_stage_model = self.first_stage_model
            elif config == "__is_unconditional__":
                print(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))
            fol