import einops import torch import torch as th import torch.nn as nn import math from ldm.modules.diffusionmodules.util import ( conv_nd, linear, zero_module, timestep_embedding, ) import torchvision from einops import rearrange, repeat from torchvision.utils import make_grid from ldm.modules.attention import SpatialTransformer from ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample, AttentionBlock from ldm.models.diffusion.ddpm import LatentDiffusion from ldm.util import log_txt_as_img, exists, instantiate_from_config from ldm.models.diffusion.ddim import DDIMSampler class VGGPerceptualLoss(torch.nn.Module): def __init__(self, resize=True): super(VGGPerceptualLoss, self).__init__() blocks = [] vgg_model = torchvision.models.vgg16(pretrained=True) print('Loaded VGG weights') blocks.append(vgg_model.features[:4].eval()) blocks.append(vgg_model.features[4:9].eval()) blocks.append(vgg_model.features[9:16].eval()) blocks.append(vgg_model.features[16:23].eval()) for bl in blocks: for p in bl.parameters(): p.requires_grad = False self.blocks = torch.nn.ModuleList(blocks) self.transform = torch.nn.functional.interpolate self.resize = resize self.register_buffer("mean", torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1)) self.register_buffer("std", torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1)) print('Initialized VGG model') def forward(self, input, feature_layers=[0, 1, 2, 3], style_layers=[1,]): if input.shape[1] != 3: input = input.repeat(1, 3, 1, 1) target = target.repeat(1, 3, 1, 1) input = (input-self.mean) / self.std if self.resize: input = self.transform(input, mode='bilinear', size=(224, 224), align_corners=False) x = input gram_matrices_all = [] feats = [] for i, block in enumerate(self.blocks): x = block(x) if i in style_layers: feats.append(x) return feats class ControlledUnetModel(UNetModel): def forward(self, x, timesteps=None, context=None, control=None, only_mid_control=False, **kwargs): hs = [] with torch.no_grad(): t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) emb = self.time_embed(t_emb) h = x.type(self.dtype) for module in self.input_blocks: h = module(h, emb, context) hs.append(h) h = self.middle_block(h, emb, context) if control is not None: h += control.pop() for i, module in enumerate(self.output_blocks): if only_mid_control or control is None: h = torch.cat([h, hs.pop()], dim=1) else: h = torch.cat([h, hs.pop() + control.pop()], dim=1) h = module(h, emb, context) h = h.type(x.dtype) return self.out(h) class ControlNet(nn.Module): def __init__( self, image_size, in_channels, model_channels, hint_channels, num_res_blocks, attention_resolutions, dropout=0, channel_mult=(1, 2, 4, 8), conv_resample=True, dims=2, use_checkpoint=False, use_fp16=False, num_heads=-1, num_head_channels=-1, num_heads_upsample=-1, use_scale_shift_norm=False, resblock_updown=False, use_new_attention_order=False, use_spatial_transformer=False, # custom transformer support transformer_depth=1, # custom transformer support context_dim=None, # custom transformer support n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model legacy=True, disable_self_attentions=None, num_attention_blocks=None, disable_middle_self_attn=False, use_linear_in_transformer=False, ): super().__init__() if use_spatial_transformer: assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' if context_dim is not None: assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' from omegaconf.listconfig import ListConfig if type(context_dim) == ListConfig: context_dim = list(context_dim) if num_heads_upsample == -1: num_heads_upsample = num_heads if num_heads == -1: assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' if num_head_channels == -1: assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' self.dims = dims self.image_size = image_size self.in_channels = in_channels self.model_channels = model_channels if isinstance(num_res_blocks, int): self.num_res_blocks = len(channel_mult) * [num_res_blocks] else: if len(num_res_blocks) != len(channel_mult): raise ValueError("provide num_res_blocks either as an int (globally constant) or " "as a list/tuple (per-level) with the same length as channel_mult") self.num_res_blocks = num_res_blocks if disable_self_attentions is not None: # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not assert len(disable_self_attentions) == len(channel_mult) if num_attention_blocks is not None: assert len(num_attention_blocks) == len(self.num_res_blocks) assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks)))) print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " f"This option has LESS priority than attention_resolutions {attention_resolutions}, " f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " f"attention will still not be set.") self.attention_resolutions = attention_resolutions self.dropout = dropout self.channel_mult = channel_mult self.conv_resample = conv_resample self.use_checkpoint = use_checkpoint self.dtype = th.float16 if use_fp16 else th.float32 self.num_heads = num_heads self.num_head_channels = num_head_channels self.num_heads_upsample = num_heads_upsample self.predict_codebook_ids = n_embed is not None time_embed_dim = model_channels * 4 self.time_embed = nn.Sequential( linear(model_channels, time_embed_dim), nn.SiLU(), linear(time_embed_dim, time_embed_dim), ) self.input_blocks = nn.ModuleList( [ TimestepEmbedSequential( conv_nd(dims, in_channels, model_channels, 3, padding=1) ) ] ) self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels)]) self.input_hint_block = TimestepEmbedSequential( conv_nd(dims, hint_channels, 16, 3, padding=1), nn.SiLU(), conv_nd(dims, 16, 16, 3, padding=1), nn.SiLU(), conv_nd(dims, 16, 32, 3, padding=1, stride=2), nn.SiLU(), conv_nd(dims, 32, 32, 3, padding=1), nn.SiLU(), conv_nd(dims, 32, 96, 3, padding=1, stride=2), nn.SiLU(), conv_nd(dims, 96, 96, 3, padding=1), nn.SiLU(), conv_nd(dims, 96, 256, 3, padding=1, stride=2), nn.SiLU(), zero_module(conv_nd(dims, 256, model_channels, 3, padding=1)) ) self._feature_size = model_channels input_block_chans = [model_channels] ch = model_channels ds = 1 for level, mult in enumerate(channel_mult): for nr in range(self.num_res_blocks[level]): layers = [ ResBlock( ch, time_embed_dim, dropout, out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, ) ] ch = mult * model_channels if ds in attention_resolutions: if num_head_channels == -1: dim_head = ch // num_heads else: num_heads = ch // num_head_channels dim_head = num_head_channels if legacy: # num_heads = 1 dim_head = ch // num_heads if use_spatial_transformer else num_head_channels if exists(disable_self_attentions): disabled_sa = disable_self_attentions[level] else: disabled_sa = False if not exists(num_attention_blocks) or nr < num_attention_blocks[level]: layers.append( AttentionBlock( ch, use_checkpoint=use_checkpoint, num_heads=num_heads, num_head_channels=dim_head, use_new_attention_order=use_new_attention_order, ) if not use_spatial_transformer else SpatialTransformer( ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, use_checkpoint=use_checkpoint ) ) self.input_blocks.append(TimestepEmbedSequential(*layers)) self.zero_convs.append(self.make_zero_conv(ch)) self._feature_size += ch input_block_chans.append(ch) if level != len(channel_mult) - 1: out_ch = ch self.input_blocks.append( TimestepEmbedSequential( ResBlock( ch, time_embed_dim, dropout, out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, down=True, ) if resblock_updown else Downsample( ch, conv_resample, dims=dims, out_channels=out_ch ) ) ) ch = out_ch input_block_chans.append(ch) self.zero_convs.append(self.make_zero_conv(ch)) ds *= 2 self._feature_size += ch if num_head_channels == -1: dim_head = ch // num_heads else: num_heads = ch // num_head_channels dim_head = num_head_channels if legacy: # num_heads = 1 dim_head = ch // num_heads if use_spatial_transformer else num_head_channels self.middle_block = TimestepEmbedSequential( ResBlock( ch, time_embed_dim, dropout, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, ), AttentionBlock( ch, use_checkpoint=use_checkpoint, num_heads=num_heads, num_head_channels=dim_head, use_new_attention_order=use_new_attention_order, ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, use_checkpoint=use_checkpoint ), ResBlock( ch, time_embed_dim, dropout, dims=dims, use_checkpoint=use_checkpoint, use_scale_shift_norm=use_scale_shift_norm, ), ) self.middle_block_out = self.make_zero_conv(ch) self._feature_size += ch def make_zero_conv(self, channels): return TimestepEmbedSequential(zero_module(conv_nd(self.dims, channels, channels, 1, padding=0))) def forward(self, x, hint, timesteps, context, **kwargs): t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) emb = self.time_embed(t_emb) guided_hint = self.input_hint_block(hint, emb, context, x.shape) outs = [] h = x.type(self.dtype) for module, zero_conv in zip(self.input_blocks, self.zero_convs): if guided_hint is not None: h = module(h, emb, context) h += guided_hint guided_hint = None else: h = module(h, emb, context) outs.append(zero_conv(h, emb, context)) h = self.middle_block(h, emb, context) outs.append(self.middle_block_out(h, emb, context)) return outs class Interpolate(nn.Module): def __init__(self, size, mode): super(Interpolate, self).__init__() self.interp = torch.nn.functional.interpolate self.size = size self.mode = mode self.factor = 8 def forward(self, x): h,w = x.shape[2]//self.factor, x.shape[3]//self.factor x = self.interp(x, size=(h,w), mode=self.mode) return x class ControlNetSAP(ControlNet): def __init__( self, hint_channels, model_channels, input_hint_block='fixed', size = 64, mode='nearest', *args, **kwargs ): super().__init__( hint_channels=hint_channels, model_channels=model_channels, *args, **kwargs) #hint channels are atleast 128 dims if input_hint_block == 'learnable': ch = 2 ** (int(math.log2(hint_channels))) self.input_hint_block = TimestepEmbedSequential( conv_nd(self.dims, hint_channels, hint_channels, 3, padding=1), nn.SiLU(), conv_nd(self.dims, hint_channels, 2*ch, 3, padding=1, stride=2), nn.SiLU(), conv_nd(self.dims, 2*ch, 2*ch, 3, padding=1), nn.SiLU(), conv_nd(self.dims, 2*ch, 2*ch, 3, padding=1, stride=2), nn.SiLU(), conv_nd(self.dims, 2*ch, 2*ch, 3, padding=1), nn.SiLU(), conv_nd(self.dims, 2*ch, model_channels, 3, padding=1, stride=2), nn.SiLU(), zero_module(conv_nd(self.dims, model_channels, model_channels, 3, padding=1)) ) else: print("Only interpolation") self.input_hint_block = TimestepEmbedSequential( Interpolate(size, mode), zero_module(conv_nd(self.dims, hint_channels, model_channels, 3, padding=1))) class ControlLDM(LatentDiffusion): def __init__(self, control_stage_config, control_key, only_mid_control, *args, **kwargs): super().__init__(*args, **kwargs) self.control_model = instantiate_from_config(control_stage_config) self.control_key = control_key self.only_mid_control = only_mid_control self.control_scales = [1.0] * 13 @torch.no_grad() def get_input(self, batch, k, bs=None, *args, **kwargs): x, c = super().get_input(batch, self.first_stage_key, *args, **kwargs) control = batch[self.control_key] if bs is not None: control = control[:bs] control = control.to(self.device) control = einops.rearrange(control, 'b h w c -> b c h w') control = control.to(memory_format=torch.contiguous_format).float() return x, dict(c_crossattn=[c], c_concat=[control]) def apply_model(self, x_noisy, t, cond, *args, **kwargs): assert isinstance(cond, dict) diffusion_model = self.model.diffusion_model cond_txt = torch.cat(cond['c_crossattn'], 1) if cond['c_concat'] is None: eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=None, only_mid_control=self.only_mid_control) else: control = self.control_model(x=x_noisy, hint=torch.cat(cond['c_concat'], 1), timesteps=t, context=cond_txt) control = [c * scale for c, scale in zip(control, self.control_scales)] eps = diffusion_model(x=x_noisy, timesteps=t, context=cond_txt, control=control, only_mid_control=self.only_mid_control) return eps @torch.no_grad() def get_unconditional_conditioning(self, N): return self.get_learned_conditioning([""] * N) @torch.no_grad() def log_images(self, batch, N=4, n_row=2, sample=False, ddim_steps=50, ddim_eta=0.0, return_keys=None, quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=False, unconditional_guidance_scale=9.0, unconditional_guidance_label=None, use_ema_scope=True, **kwargs): use_ddim = ddim_steps is not None log = dict() z, c = self.get_input(batch, self.first_stage_key, bs=N) c_cat, c = c["c_concat"][0][:N], c["c_crossattn"][0][:N] N = min(z.shape[0], N) n_row = min(z.shape[0], n_row) log["reconstruction"] = self.decode_first_stage(z) log["control"] = c_cat * 2.0 - 1.0 log["conditioning"] = log_txt_as_img((512, 512), batch[self.cond_stage_key], size=16) if plot_diffusion_rows: # get diffusion row diffusion_row = list() z_start = z[:n_row] for t in range(self.num_timesteps): if t % self.log_every_t == 0 or t == self.num_timesteps - 1: t = repeat(torch.tensor([t]), '1 -> b', b=n_row) t = t.to(self.device).long() noise = torch.randn_like(z_start) z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) diffusion_row.append(self.decode_first_stage(z_noisy)) diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) log["diffusion_row"] = diffusion_grid if sample: # get denoise row samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, batch_size=N, ddim=use_ddim, ddim_steps=ddim_steps, eta=ddim_eta) x_samples = self.decode_first_stage(samples) log["samples"] = x_samples if plot_denoise_rows: denoise_grid = self._get_denoise_row_from_list(z_denoise_row) log["denoise_row"] = denoise_grid if unconditional_guidance_scale > 1.0: uc_cross = self.get_unconditional_conditioning(N) uc_cat = c_cat # torch.zeros_like(c_cat) uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]} samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, batch_size=N, ddim=use_ddim, ddim_steps=ddim_steps, eta=ddim_eta, unconditional_guidance_scale=unconditional_guidance_scale, unconditional_conditioning=uc_full, ) x_samples_cfg = self.decode_first_stage(samples_cfg) log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg return log @torch.no_grad() def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs): ddim_sampler = DDIMSampler(self) b, c, h, w = cond["c_concat"][0].shape shape = (self.channels, h // 8, w // 8) samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, shape, cond, verbose=False, **kwargs) return samples, intermediates def configure_optimizers(self): lr = self.learning_rate params = list(self.control_model.parameters()) if not self.sd_locked: params += list(self.model.diffusion_model.output_blocks.parameters()) params += list(self.model.diffusion_model.out.parameters()) opt = torch.optim.AdamW(params, lr=lr) return opt def low_vram_shift(self, is_diffusing): if is_diffusing: self.model = self.model.cuda() self.control_model = self.control_model.cuda() self.first_stage_model = self.first_stage_model.cpu() self.cond_stage_model = self.cond_stage_model.cpu() else: self.model = self.model.cpu() self.control_model = self.control_model.cpu() self.first_stage_model = self.first_stage_model.cuda() self.cond_stage_model = self.cond_stage_model.cuda() class SAP(ControlLDM): @torch.no_grad() def __init__(self,control_stage_config, control_key, only_mid_control, *args, **kwargs): super().__init__(control_stage_config=control_stage_config, control_key=control_key, only_mid_control=only_mid_control, *args, **kwargs) self.appearance_net = VGGPerceptualLoss().to(self.device) print("Loaded VGG model") def get_appearance(self, img, mask, return_all=False): img = (img + 1) * 0.5 feat = self.appearance_net(img)[0] empty_mask_flag = torch.sum(mask, dim=(1,2,3)) == 0 empty_appearance = torch.zeros(feat.shape).to(self.device) mask = torch.nn.functional.interpolate(mask.float(), (feat.shape[2:])).long() one_hot = torch.nn.functional.one_hot(mask[:,0]).permute(0,3,1,2).float() feat = torch.einsum('nchw, nmhw->nmchw', feat, one_hot) feat = torch.sum(feat, dim=(3,4)) norm = torch.sum(one_hot, dim=(2,3)) + 1e-6 #nm mean_feat = feat/norm[:,:,None] #nmc mean_feat[:, 0] = torch.zeros(mean_feat[:,0].shape).to(self.device) #set edges in panopitc mask to empty appearance feature splatted_feat = torch.einsum('nmc, nmhw->nchw', mean_feat, one_hot) splatted_feat[empty_mask_flag] = empty_appearance[empty_mask_flag] splatted_feat = torch.nn.functional.normalize(splatted_feat) #l2 normalize on c dim if return_all: return splatted_feat, mean_feat, one_hot, empty_mask_flag return splatted_feat def get_input(self, batch, k, bs=None, *args, **kwargs): z, c, x_orig, x_recon = super(ControlLDM, self).get_input(batch, self.first_stage_key, return_first_stage_outputs=True , *args, **kwargs) structure = batch['seg'].unsqueeze(1) mask = batch['mask'].unsqueeze(1).to(self.device) appearance = self.get_appearance(x_orig, mask) if bs is not None: structure = structure[:bs] appearance = appearance[:bs] structure = structure.to(self.device) appearance = appearance.to(self.device) structure = structure.to(memory_format=torch.contiguous_format).float() appearance = appearance.to(memory_format=torch.contiguous_format).float() structure = torch.nn.functional.interpolate(structure, x_orig.shape[2:]) appearance = torch.nn.functional.interpolate(appearance, x_orig.shape[2:]) control = torch.cat([structure, appearance], dim=1) return z, dict(c_crossattn=[c], c_concat=[control]) @torch.no_grad() def log_images(self, batch, N=4, n_row=2, sample=False, ddim_steps=50, ddim_eta=0.0, return_keys=None, quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=False, plot_diffusion_rows=False, unconditional_guidance_scale=9.0, unconditional_guidance_label=None, use_ema_scope=True, **kwargs): use_ddim = ddim_steps is not None log = dict() z, c = self.get_input(batch, self.first_stage_key, bs=N) c_cat, c = c["c_concat"][0][:N,], c["c_crossattn"][0][:N] N = min(z.shape[0], N) n_row = min(z.shape[0], n_row) log["reconstruction"] = self.decode_first_stage(z) log["control"] = c_cat[:, :1] log["conditioning"] = log_txt_as_img((512, 512), batch[self.cond_stage_key], size=16) if plot_diffusion_rows: # get diffusion row diffusion_row = list() z_start = z[:n_row] for t in range(self.num_timesteps): if t % self.log_every_t == 0 or t == self.num_timesteps - 1: t = repeat(torch.tensor([t]), '1 -> b', b=n_row) t = t.to(self.device).long() noise = torch.randn_like(z_start) z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) diffusion_row.append(self.decode_first_stage(z_noisy)) diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) log["diffusion_row"] = diffusion_grid if plot_progressive_rows: with self.ema_scope("Plotting Progressives"): img, progressives = self.progressive_denoising({"c_concat": [c_cat], "c_crossattn": [c]}, shape=(self.channels, self.image_size, self.image_size), batch_size=N) prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation") log["progressive_row"] = prog_row if sample: # get denoise row samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, batch_size=N, ddim=use_ddim, ddim_steps=ddim_steps, eta=ddim_eta) x_samples = self.decode_first_stage(samples) log["samples"] = x_samples if plot_denoise_rows: denoise_grid = self._get_denoise_row_from_list(z_denoise_row) log["denoise_row"] = denoise_grid if unconditional_guidance_scale > 1.0: uc_cross = self.get_unconditional_conditioning(N) uc_cat = c_cat # torch.zeros_like(c_cat) uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]} samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, batch_size=N, ddim=use_ddim, ddim_steps=ddim_steps, eta=ddim_eta, unconditional_guidance_scale=unconditional_guidance_scale, unconditional_conditioning=uc_full, ) x_samples_cfg = self.decode_first_stage(samples_cfg) log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg return log