import ast import gc import random import cv2 import numpy as np import torch import torch.nn.functional as F from diffusers.models.attention_processor import AttnProcessor2_0 from diffusers.models.attention import BasicTransformerBlock from decord import VideoReader import wandb def extract_into_tensor(a, t, x_shape): b, *_ = t.shape out = a.gather(-1, t) return out.reshape(b, *((1,) * (len(x_shape) - 1))) def is_attn(name): return "attn1" or "attn2" == name.split(".")[-1] def set_processors(attentions): for attn in attentions: attn.set_processor(AttnProcessor2_0()) def set_torch_2_attn(unet): optim_count = 0 for name, module in unet.named_modules(): if is_attn(name): if isinstance(module, torch.nn.ModuleList): for m in module: if isinstance(m, BasicTransformerBlock): set_processors([m.attn1, m.attn2]) optim_count += 1 if optim_count > 0: print(f"{optim_count} Attention layers using Scaled Dot Product Attention.") # From LatentConsistencyModel.get_guidance_scale_embedding def guidance_scale_embedding(w, embedding_dim=512, dtype=torch.float32): """ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 Args: timesteps (`torch.Tensor`): generate embedding vectors at these timesteps embedding_dim (`int`, *optional*, defaults to 512): dimension of the embeddings to generate dtype: data type of the generated embeddings Returns: `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)` """ assert len(w.shape) == 1 w = w * 1000.0 half_dim = embedding_dim // 2 emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) emb = w.to(dtype)[:, None] * emb[None, :] emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) if embedding_dim % 2 == 1: # zero pad emb = torch.nn.functional.pad(emb, (0, 1)) assert emb.shape == (w.shape[0], embedding_dim) return emb def append_dims(x, target_dims): """Appends dimensions to the end of a tensor until it has target_dims dimensions.""" dims_to_append = target_dims - x.ndim if dims_to_append < 0: raise ValueError( f"input has {x.ndim} dims but target_dims is {target_dims}, which is less" ) return x[(...,) + (None,) * dims_to_append] # From LCMScheduler.get_scalings_for_boundary_condition_discrete def scalings_for_boundary_conditions(timestep, sigma_data=0.5, timestep_scaling=10.0): scaled_timestep = timestep_scaling * timestep c_skip = sigma_data**2 / (scaled_timestep**2 + sigma_data**2) c_out = scaled_timestep / (scaled_timestep**2 + sigma_data**2) ** 0.5 return c_skip, c_out # Compare LCMScheduler.step, Step 4 def get_predicted_original_sample( model_output, timesteps, sample, prediction_type, alphas, sigmas ): alphas = extract_into_tensor(alphas, timesteps, sample.shape) sigmas = extract_into_tensor(sigmas, timesteps, sample.shape) if prediction_type == "epsilon": pred_x_0 = (sample - sigmas * model_output) / alphas elif prediction_type == "sample": pred_x_0 = model_output elif prediction_type == "v_prediction": pred_x_0 = alphas * sample - sigmas * model_output else: raise ValueError( f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`" f" are supported." ) return pred_x_0 # Based on step 4 in DDIMScheduler.step def get_predicted_noise( model_output, timesteps, sample, prediction_type, alphas, sigmas ): alphas = extract_into_tensor(alphas, timesteps, sample.shape) sigmas = extract_into_tensor(sigmas, timesteps, sample.shape) if prediction_type == "epsilon": pred_epsilon = model_output elif prediction_type == "sample": pred_epsilon = (sample - alphas * model_output) / sigmas elif prediction_type == "v_prediction": pred_epsilon = alphas * model_output + sigmas * sample else: raise ValueError( f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`" f" are supported." ) return pred_epsilon # From LatentConsistencyModel.get_guidance_scale_embedding def guidance_scale_embedding(w, embedding_dim=512, dtype=torch.float32): """ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 Args: timesteps (`torch.Tensor`): generate embedding vectors at these timesteps embedding_dim (`int`, *optional*, defaults to 512): dimension of the embeddings to generate dtype: data type of the generated embeddings Returns: `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)` """ assert len(w.shape) == 1 w = w * 1000.0 half_dim = embedding_dim // 2 emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) emb = w.to(dtype)[:, None] * emb[None, :] emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) if embedding_dim % 2 == 1: # zero pad emb = torch.nn.functional.pad(emb, (0, 1)) assert emb.shape == (w.shape[0], embedding_dim) return emb def append_dims(x, target_dims): """Appends dimensions to the end of a tensor until it has target_dims dimensions.""" dims_to_append = target_dims - x.ndim if dims_to_append < 0: raise ValueError( f"input has {x.ndim} dims but target_dims is {target_dims}, which is less" ) return x[(...,) + (None,) * dims_to_append] # From LCMScheduler.get_scalings_for_boundary_condition_discrete def scalings_for_boundary_conditions(timestep, sigma_data=0.5, timestep_scaling=10.0): scaled_timestep = timestep_scaling * timestep c_skip = sigma_data**2 / (scaled_timestep**2 + sigma_data**2) c_out = scaled_timestep / (scaled_timestep**2 + sigma_data**2) ** 0.5 return c_skip, c_out # Compare LCMScheduler.step, Step 4 def get_predicted_original_sample( model_output, timesteps, sample, prediction_type, alphas, sigmas ): alphas = extract_into_tensor(alphas, timesteps, sample.shape) sigmas = extract_into_tensor(sigmas, timesteps, sample.shape) if prediction_type == "epsilon": pred_x_0 = (sample - sigmas * model_output) / alphas elif prediction_type == "sample": pred_x_0 = model_output elif prediction_type == "v_prediction": pred_x_0 = alphas * sample - sigmas * model_output else: raise ValueError( f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`" f" are supported." ) return pred_x_0 # Based on step 4 in DDIMScheduler.step def get_predicted_noise( model_output, timesteps, sample, prediction_type, alphas, sigmas ): alphas = extract_into_tensor(alphas, timesteps, sample.shape) sigmas = extract_into_tensor(sigmas, timesteps, sample.shape) if prediction_type == "epsilon": pred_epsilon = model_output elif prediction_type == "sample": pred_epsilon = (sample - alphas * model_output) / sigmas elif prediction_type == "v_prediction": pred_epsilon = alphas * model_output + sigmas * sample else: raise ValueError( f"Prediction type {prediction_type} is not supported; currently, `epsilon`, `sample`, and `v_prediction`" f" are supported." ) return pred_epsilon def param_optim(model, condition, extra_params=None, is_lora=False, negation=None): extra_params = extra_params if len(extra_params.keys()) > 0 else None return { "model": model, "condition": condition, "extra_params": extra_params, "is_lora": is_lora, "negation": negation, } def create_optim_params(name="param", params=None, lr=5e-6, extra_params=None): params = {"name": name, "params": params, "lr": lr} if extra_params is not None: for k, v in extra_params.items(): params[k] = v return params def create_optimizer_params(model_list, lr): import itertools optimizer_params = [] for optim in model_list: model, condition, extra_params, is_lora, negation = optim.values() # Check if we are doing LoRA training. if is_lora and condition and isinstance(model, list): params = create_optim_params( params=itertools.chain(*model), extra_params=extra_params ) optimizer_params.append(params) continue if is_lora and condition and not isinstance(model, list): for n, p in model.named_parameters(): if "lora" in n: params = create_optim_params(n, p, lr, extra_params) optimizer_params.append(params) continue # If this is true, we can train it. if condition: for n, p in model.named_parameters(): should_negate = "lora" in n and not is_lora if should_negate: continue params = create_optim_params(n, p, lr, extra_params) optimizer_params.append(params) return optimizer_params def handle_trainable_modules( model, trainable_modules=None, is_enabled=True, negation=None ): acc = [] unfrozen_params = 0 if trainable_modules is not None: unlock_all = any([name == "all" for name in trainable_modules]) if unlock_all: model.requires_grad_(True) unfrozen_params = len(list(model.parameters())) else: model.requires_grad_(False) for name, param in model.named_parameters(): for tm in trainable_modules: if all([tm in name, name not in acc, "lora" not in name]): param.requires_grad_(is_enabled) acc.append(name) unfrozen_params += 1 def huber_loss(pred, target, huber_c=0.001): loss = torch.sqrt((pred.float() - target.float()) ** 2 + huber_c**2) - huber_c return loss.mean() @torch.no_grad() def update_ema(target_params, source_params, rate=0.99): """ Update target parameters to be closer to those of source parameters using an exponential moving average. :param target_params: the target parameter sequence. :param source_params: the source parameter sequence. :param rate: the EMA rate (closer to 1 means slower). """ for targ, src in zip(target_params, source_params): src_to_dtype = src.to(targ.dtype) targ.detach().mul_(rate).add_(src_to_dtype, alpha=1 - rate) def log_validation_video(pipeline, args, accelerator, save_fps): if args.seed is None: generator = None else: generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) validation_prompts = [ "An astronaut riding a horse.", "Darth vader surfing in waves.", "Robot dancing in times square.", "Clown fish swimming through the coral reef.", "A child excitedly swings on a rusty swing set, laughter filling the air.", "With the style of van gogh, A young couple dances under the moonlight by the lake.", "A young woman with glasses is jogging in the park wearing a pink headband.", "Impressionist style, a yellow rubber duck floating on the wave on the sunset", "Wolf, turns its head, in the wild", "Iron man, walks, on the moon, 8k, high detailed, best quality", "With the style of low-poly game art, A majestic, white horse gallops gracefully", "a rabbit, low-poly game art style", ] video_logs = [] if getattr(args, "use_motion_cond", False): use_motion_cond = True else: use_motion_cond = False for _, prompt in enumerate(validation_prompts): if use_motion_cond: motin_gs_unit = (args.motion_gs_max - args.motion_gs_min) / 2 for i in range(3): with torch.autocast("cuda"): videos = pipeline( prompt=prompt, frames=args.n_frames, num_inference_steps=8, num_videos_per_prompt=1, fps=args.fps, use_motion_cond=True, motion_gs=motin_gs_unit * i, lcm_origin_steps=args.num_ddim_timesteps, generator=generator, ) videos = (videos.clamp(-1.0, 1.0) + 1.0) / 2.0 videos = ( (videos * 255) .to(torch.uint8) .permute(0, 2, 1, 3, 4) .cpu() .numpy() ) video_logs.append( { "validation_prompt": f"GS={i * motin_gs_unit}, {prompt}", "videos": videos, } ) else: for i in range(2): with torch.autocast("cuda"): videos = pipeline( prompt=prompt, frames=args.n_frames, num_inference_steps=4 * (i + 1), num_videos_per_prompt=1, fps=args.fps, use_motion_cond=False, lcm_origin_steps=args.num_ddim_timesteps, generator=generator, ) videos = (videos.clamp(-1.0, 1.0) + 1.0) / 2.0 videos = ( (videos * 255) .to(torch.uint8) .permute(0, 2, 1, 3, 4) .cpu() .numpy() ) video_logs.append( { "validation_prompt": f"Steps={4 * (i + 1)}, {prompt}", "videos": videos, } ) for tracker in accelerator.trackers: if tracker.name == "wandb": formatted_videos = [] for log in video_logs: videos = log["videos"] validation_prompt = log["validation_prompt"] for video in videos: video = wandb.Video(video, caption=validation_prompt, fps=save_fps) formatted_videos.append(video) tracker.log({f"validation": formatted_videos}) del pipeline gc.collect() def tuple_type(s): if isinstance(s, tuple): return s value = ast.literal_eval(s) if isinstance(value, tuple): return value raise TypeError("Argument must be a tuple") def load_model_checkpoint(model, ckpt): def load_checkpoint(model, ckpt, full_strict): state_dict = torch.load(ckpt, map_location="cpu", weights_only=True) if "state_dict" in list(state_dict.keys()): state_dict = state_dict["state_dict"] model.load_state_dict(state_dict, strict=full_strict) del state_dict gc.collect() return model load_checkpoint(model, ckpt, full_strict=True) print(">>> model checkpoint loaded.") return model def read_video_to_tensor( path_to_video, sample_fps, sample_frames, uniform_sampling=False ): video_reader = VideoReader(path_to_video) video_fps = video_reader.get_avg_fps() video_frames = video_reader._num_frame video_duration = video_frames / video_fps sample_duration = sample_frames / sample_fps stride = video_fps / sample_fps if uniform_sampling or video_duration <= sample_duration: index_range = np.linspace(0, video_frames - 1, sample_frames).astype(np.int32) else: max_start_frame = video_frames - np.ceil(sample_frames * stride).astype( np.int32 ) if max_start_frame > 0: start_frame = random.randint(0, max_start_frame) else: start_frame = 0 index_range = start_frame + np.arange(sample_frames) * stride index_range = np.round(index_range).astype(np.int32) sampled_frames = video_reader.get_batch(index_range).asnumpy() pixel_values = torch.from_numpy(sampled_frames).permute(0, 3, 1, 2).contiguous() pixel_values = pixel_values / 255.0 del video_reader return pixel_values def calculate_motion_rank_new(tensor_ref, tensor_gen, rank_k=1): if rank_k == 0: loss = torch.tensor(0.0, device=tensor_ref.device) elif rank_k > tensor_ref.shape[-1]: raise ValueError( "The value of rank_k cannot be larger than the number of frames" ) else: # Sort the reference tensor along the frames dimension _, sorted_indices = torch.sort(tensor_ref, dim=-1) # Create a mask to select the top rank_k frames mask = torch.zeros_like(tensor_ref, dtype=torch.bool) mask.scatter_(-1, sorted_indices[..., -rank_k:], True) # Compute the mean squared error loss only on the masked elements loss = F.mse_loss(tensor_ref[mask].detach(), tensor_gen[mask]) return loss def compute_temp_loss(attention_prob, attention_prob_example): temp_attn_prob_loss = [] # 1. Loop though all layers to get the query, key, and Compute the PCA loss for name in attention_prob.keys(): attn_prob_example = attention_prob_example[name] attn_prob = attention_prob[name] module_attn_loss = calculate_motion_rank_new( attn_prob_example.detach(), attn_prob, rank_k=1 ) temp_attn_prob_loss.append(module_attn_loss) loss_temp = torch.stack(temp_attn_prob_loss) * 100 loss = loss_temp.mean() return loss