#!/usr/bin/env python # coding=utf-8 # Copyright 2025 The HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """LPL training script for Stable Diffusion XL for text2image.""" import argparse import functools import gc import logging import math import os import random import re import shutil from contextlib import nullcontext from pathlib import Path from typing import Dict, List, Tuple import accelerate import datasets import numpy as np import torch import torch.nn as nn import torch.nn.functional as F import torch.utils.checkpoint import transformers from accelerate import Accelerator from accelerate.logging import get_logger from accelerate.utils import DistributedType, ProjectConfiguration, set_seed from datasets import concatenate_datasets, load_dataset from huggingface_hub import create_repo, upload_folder from lpl_loss import LatentPerceptualLoss from packaging import version from torchvision import transforms from torchvision.transforms.functional import crop from tqdm.auto import tqdm from transformers import AutoTokenizer, PretrainedConfig import diffusers from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionXLPipeline, UNet2DConditionModel from diffusers.optimization import get_scheduler from diffusers.training_utils import EMAModel, compute_snr from diffusers.utils import check_min_version, is_wandb_available from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card from diffusers.utils.import_utils import is_torch_npu_available, is_xformers_available from diffusers.utils.torch_utils import is_compiled_module # Will error if the minimal version of diffusers is not installed. Remove at your own risks. check_min_version("0.33.0.dev0") logger = get_logger(__name__) if is_torch_npu_available(): import torch_npu torch.npu.config.allow_internal_format = False DATASET_NAME_MAPPING = { "lambdalabs/naruto-blip-captions": ("image", "text"), } # Global dictionary to store intermediate features from hooks hook_features: Dict[str, torch.Tensor] = {} def get_intermediate_features_hook(name: str): """Creates a hook function that saves the output of a layer.""" def hook(model, input, output): # Some layers might return tuples (e.g., attention blocks) # We are usually interested in the first element (hidden states) if isinstance(output, tuple): hook_features[name] = output[0] else: hook_features[name] = output return hook def clear_hook_features(): """Clears the global feature dictionary.""" global hook_features hook_features = {} def normalize_features( feat1: torch.Tensor, feat2: torch.Tensor, eps: float = 1e-6 ) -> Tuple[torch.Tensor, torch.Tensor]: """ Normalizes feat1 and feat2 using the statistics of feat2 (predicted features). Normalization is done per-channel. """ # Calculate stats over spatial dimensions (H, W) dims = tuple(range(2, feat2.ndim)) # Dims to reduce over (usually 2, 3 for H, W) mean = torch.mean(feat2, dim=dims, keepdim=True) std = torch.std(feat2, dim=dims, keepdim=True) + eps feat1_norm = (feat1 - mean) / std feat2_norm = (feat2 - mean) / std return feat1_norm, feat2_norm def get_decoder_layer_names(decoder: nn.Module) -> List[str]: """Helper to get potential layer names for hooks in the VAE decoder.""" layer_names = [] for name, module in decoder.named_modules(): # Example: Target ResnetBlocks and potentially UpBlocks if isinstance(module, (diffusers.models.resnet.ResnetBlock2D, diffusers.models.unet_2d_blocks.UpBlock2D)): # Filter out redundant names if UpBlock contains ResnetBlocks already named is_child = any( name.startswith(parent + ".") for parent in layer_names if isinstance(decoder.get_submodule(parent), diffusers.models.unet_2d_blocks.UpBlock2D) ) if not is_child: layer_names.append(name) # A basic default selection if complex logic fails if not layer_names: layer_names = [ name for name, module in decoder.named_modules() if re.match(r"up_blocks\.\d+\.resnets\.\d+$", name) ] return layer_names def save_model_card( repo_id: str, images: list = None, validation_prompt: str = None, base_model: str = None, dataset_name: str = None, repo_folder: str = None, vae_path: str = None, ): img_str = "" if images is not None: for i, image in enumerate(images): image.save(os.path.join(repo_folder, f"image_{i}.png")) img_str += f"![img_{i}](./image_{i}.png)\n" model_description = f""" # Text-to-image finetuning - {repo_id} This pipeline was finetuned from **{base_model}** on the **{dataset_name}** dataset. Below are some example images generated with the finetuned pipeline using the following prompt: {validation_prompt}: \n {img_str} Special VAE used for training: {vae_path}. """ model_card = load_or_create_model_card( repo_id_or_path=repo_id, from_training=True, license="creativeml-openrail-m", base_model=base_model, model_description=model_description, inference=True, ) tags = [ "stable-diffusion-xl", "stable-diffusion-xl-diffusers", "text-to-image", "diffusers-training", "diffusers", ] model_card = populate_model_card(model_card, tags=tags) model_card.save(os.path.join(repo_folder, "README.md")) def import_model_class_from_model_name_or_path( pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder" ): text_encoder_config = PretrainedConfig.from_pretrained( pretrained_model_name_or_path, subfolder=subfolder, revision=revision ) model_class = text_encoder_config.architectures[0] if model_class == "CLIPTextModel": from transformers import CLIPTextModel return CLIPTextModel elif model_class == "CLIPTextModelWithProjection": from transformers import CLIPTextModelWithProjection return CLIPTextModelWithProjection else: raise ValueError(f"{model_class} is not supported.") def parse_args(input_args=None): parser = argparse.ArgumentParser(description="LPL based training script of Stable Diffusion XL.") parser.add_argument( "--pretrained_model_name_or_path", type=str, default=None, required=True, help="Path to pretrained model or model identifier from huggingface.co/models.", ) parser.add_argument( "--pretrained_vae_model_name_or_path", type=str, default=None, help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.", ) parser.add_argument( "--revision", type=str, default=None, required=False, help="Revision of pretrained model identifier from huggingface.co/models.", ) parser.add_argument( "--variant", type=str, default=None, help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16", ) parser.add_argument( "--dataset_name", type=str, default=None, help=( "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private," " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem," " or to a folder containing files that 🤗 Datasets can understand." ), ) parser.add_argument( "--dataset_config_name", type=str, default=None, help="The config of the Dataset, leave as None if there's only one config.", ) parser.add_argument( "--train_data_dir", type=str, default=None, help=( "A folder containing the training data. Folder contents must follow the structure described in" " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file" " must exist to provide the captions for the images. Ignored if `dataset_name` is specified." ), ) parser.add_argument( "--image_column", type=str, default="image", help="The column of the dataset containing an image." ) parser.add_argument( "--caption_column", type=str, default="text", help="The column of the dataset containing a caption or a list of captions.", ) parser.add_argument( "--validation_prompt", type=str, default=None, help="A prompt that is used during validation to verify that the model is learning.", ) parser.add_argument( "--num_validation_images", type=int, default=4, help="Number of images that should be generated during validation with `validation_prompt`.", ) parser.add_argument( "--validation_epochs", type=int, default=1, help=( "Run fine-tuning validation every X epochs. The validation process consists of running the prompt" " `args.validation_prompt` multiple times: `args.num_validation_images`." ), ) parser.add_argument( "--max_train_samples", type=int, default=None, help=( "For debugging purposes or quicker training, truncate the number of training examples to this " "value if set." ), ) parser.add_argument( "--proportion_empty_prompts", type=float, default=0, help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).", ) parser.add_argument( "--output_dir", type=str, default="sdxl-model-finetuned", help="The output directory where the model predictions and checkpoints will be written.", ) parser.add_argument( "--cache_dir", type=str, default=None, help="The directory where the downloaded models and datasets will be stored.", ) parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") parser.add_argument( "--resolution", type=int, default=1024, help=( "The resolution for input images, all the images in the train/validation dataset will be resized to this" " resolution" ), ) parser.add_argument( "--center_crop", default=False, action="store_true", help=( "Whether to center crop the input images to the resolution. If not set, the images will be randomly" " cropped. The images will be resized to the resolution first before cropping." ), ) parser.add_argument( "--random_flip", action="store_true", help="whether to randomly flip images horizontally", ) parser.add_argument( "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader." ) parser.add_argument("--num_train_epochs", type=int, default=100) parser.add_argument( "--max_train_steps", type=int, default=None, help="Total number of training steps to perform. If provided, overrides num_train_epochs.", ) parser.add_argument( "--checkpointing_steps", type=int, default=500, help=( "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final" " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming" " training using `--resume_from_checkpoint`." ), ) parser.add_argument( "--checkpoints_total_limit", type=int, default=None, help=("Max number of checkpoints to store."), ) parser.add_argument( "--resume_from_checkpoint", type=str, default=None, help=( "Whether training should be resumed from a previous checkpoint. Use a path saved by" ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.' ), ) parser.add_argument( "--gradient_accumulation_steps", type=int, default=1, help="Number of updates steps to accumulate before performing a backward/update pass.", ) parser.add_argument( "--gradient_checkpointing", action="store_true", help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", ) parser.add_argument( "--learning_rate", type=float, default=1e-4, help="Initial learning rate (after the potential warmup period) to use.", ) parser.add_argument( "--scale_lr", action="store_true", default=False, help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", ) parser.add_argument( "--lr_scheduler", type=str, default="constant", help=( 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' ' "constant", "constant_with_warmup"]' ), ) parser.add_argument( "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." ) parser.add_argument( "--timestep_bias_strategy", type=str, default="none", choices=["earlier", "later", "range", "none"], help=( "The timestep bias strategy, which may help direct the model toward learning low or high frequency details." " Choices: ['earlier', 'later', 'range', 'none']." " The default is 'none', which means no bias is applied, and training proceeds normally." " The value of 'later' will increase the frequency of the model's final training timesteps." ), ) parser.add_argument( "--timestep_bias_multiplier", type=float, default=1.0, help=( "The multiplier for the bias. Defaults to 1.0, which means no bias is applied." " A value of 2.0 will double the weight of the bias, and a value of 0.5 will halve it." ), ) parser.add_argument( "--timestep_bias_begin", type=int, default=0, help=( "When using `--timestep_bias_strategy=range`, the beginning (inclusive) timestep to bias." " Defaults to zero, which equates to having no specific bias." ), ) parser.add_argument( "--timestep_bias_end", type=int, default=1000, help=( "When using `--timestep_bias_strategy=range`, the final timestep (inclusive) to bias." " Defaults to 1000, which is the number of timesteps that Stable Diffusion is trained on." ), ) parser.add_argument( "--timestep_bias_portion", type=float, default=0.25, help=( "The portion of timesteps to bias. Defaults to 0.25, which 25% of timesteps will be biased." " A value of 0.5 will bias one half of the timesteps. The value provided for `--timestep_bias_strategy` determines" " whether the biased portions are in the earlier or later timesteps." ), ) parser.add_argument( "--snr_gamma", type=float, default=None, help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. " "More details here: https://arxiv.org/abs/2303.09556.", ) parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.") parser.add_argument( "--allow_tf32", action="store_true", help=( "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see" " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices" ), ) parser.add_argument( "--dataloader_num_workers", type=int, default=0, help=( "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." ), ) parser.add_argument( "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." ) parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.") parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer") parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.") parser.add_argument( "--prediction_type", type=str, default=None, help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediction_type` is chosen.", ) parser.add_argument( "--hub_model_id", type=str, default=None, help="The name of the repository to keep in sync with the local `output_dir`.", ) parser.add_argument( "--logging_dir", type=str, default="logs", help=( "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." ), ) parser.add_argument( "--report_to", type=str, default="tensorboard", help=( 'The integration to report the results and logs to. Supported platforms are `"tensorboard"`' ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.' ), ) parser.add_argument( "--mixed_precision", type=str, default=None, choices=["no", "fp16", "bf16"], help=( "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >=" " 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the" " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config." ), ) parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") parser.add_argument( "--enable_npu_flash_attention", action="store_true", help="Whether or not to use npu flash attention." ) parser.add_argument( "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers." ) parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.") parser.add_argument( "--use_lpl", action="store_true", help="Whether to use Latent Perceptual Loss (LPL). Increases memory usage.", ) parser.add_argument( "--lpl_weight", type=float, default=1.0, help="Weight for the Latent Perceptual Loss.", ) parser.add_argument( "--lpl_t_threshold", type=int, default=200, help="Apply LPL only for timesteps t < lpl_t_threshold. Corresponds to high SNR.", ) parser.add_argument( "--lpl_loss_type", type=str, default="mse", choices=["mse", "l1"], help="Type of loss to use for LPL.", ) parser.add_argument( "--lpl_norm_type", type=str, default="default", choices=["default", "shared", "batch"], help="Type of normalization to use for LPL features.", ) parser.add_argument( "--lpl_pow_law", action="store_true", help="Whether to use power law weighting for LPL layers.", ) parser.add_argument( "--lpl_num_blocks", type=int, default=4, help="Number of up blocks to use for LPL feature extraction.", ) parser.add_argument( "--lpl_remove_outliers", action="store_true", help="Whether to remove outliers in LPL feature maps.", ) parser.add_argument( "--lpl_scale", action="store_true", help="Whether to scale LPL loss by noise level weights.", ) parser.add_argument( "--lpl_start", type=int, default=0, help="Step to start applying LPL loss.", ) if input_args is not None: args = parser.parse_args(input_args) else: args = parser.parse_args() env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) if env_local_rank != -1 and env_local_rank != args.local_rank: args.local_rank = env_local_rank # Sanity checks if args.dataset_name is None and args.train_data_dir is None: raise ValueError("Need either a dataset name or a training folder.") if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1: raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].") return args # Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt def encode_prompt(batch, text_encoders, tokenizers, proportion_empty_prompts, caption_column, is_train=True): prompt_embeds_list = [] prompt_batch = batch[caption_column] captions = [] for caption in prompt_batch: if random.random() < proportion_empty_prompts: captions.append("") elif isinstance(caption, str): captions.append(caption) elif isinstance(caption, (list, np.ndarray)): # take a random caption if there are multiple captions.append(random.choice(caption) if is_train else caption[0]) with torch.no_grad(): for tokenizer, text_encoder in zip(tokenizers, text_encoders): text_inputs = tokenizer( captions, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt", ) text_input_ids = text_inputs.input_ids prompt_embeds = text_encoder( text_input_ids.to(text_encoder.device), output_hidden_states=True, return_dict=False, ) # We are only ALWAYS interested in the pooled output of the final text encoder pooled_prompt_embeds = prompt_embeds[0] prompt_embeds = prompt_embeds[-1][-2] bs_embed, seq_len, _ = prompt_embeds.shape prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1) prompt_embeds_list.append(prompt_embeds) prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1) return {"prompt_embeds": prompt_embeds.cpu(), "pooled_prompt_embeds": pooled_prompt_embeds.cpu()} def compute_vae_encodings(batch, vae): images = batch.pop("pixel_values") pixel_values = torch.stack(list(images)) pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float() pixel_values = pixel_values.to(vae.device, dtype=vae.dtype) with torch.no_grad(): model_input = vae.encode(pixel_values).latent_dist.sample() model_input = model_input * vae.config.scaling_factor # There might have slightly performance improvement # by changing model_input.cpu() to accelerator.gather(model_input) return {"model_input": model_input.cpu()} def generate_timestep_weights(args, num_timesteps): weights = torch.ones(num_timesteps) # Determine the indices to bias num_to_bias = int(args.timestep_bias_portion * num_timesteps) if args.timestep_bias_strategy == "later": bias_indices = slice(-num_to_bias, None) elif args.timestep_bias_strategy == "earlier": bias_indices = slice(0, num_to_bias) elif args.timestep_bias_strategy == "range": # Out of the possible 1000 timesteps, we might want to focus on eg. 200-500. range_begin = args.timestep_bias_begin range_end = args.timestep_bias_end if range_begin < 0: raise ValueError( "When using the range strategy for timestep bias, you must provide a beginning timestep greater or equal to zero." ) if range_end > num_timesteps: raise ValueError( "When using the range strategy for timestep bias, you must provide an ending timestep smaller than the number of timesteps." ) bias_indices = slice(range_begin, range_end) else: # 'none' or any other string return weights if args.timestep_bias_multiplier <= 0: return ValueError( "The parameter --timestep_bias_multiplier is not intended to be used to disable the training of specific timesteps." " If it was intended to disable timestep bias, use `--timestep_bias_strategy none` instead." " A timestep bias multiplier less than or equal to 0 is not allowed." ) # Apply the bias weights[bias_indices] *= args.timestep_bias_multiplier # Normalize weights /= weights.sum() return weights def main(args): if args.report_to == "wandb" and args.hub_token is not None: raise ValueError( "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token." " Please use `huggingface-cli login` to authenticate with the Hub." ) logging_dir = Path(args.output_dir, args.logging_dir) accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir) if torch.backends.mps.is_available() and args.mixed_precision == "bf16": # due to pytorch#99272, MPS does not yet support bfloat16. raise ValueError( "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead." ) accelerator = Accelerator( gradient_accumulation_steps=args.gradient_accumulation_steps, mixed_precision=args.mixed_precision, log_with=args.report_to, project_config=accelerator_project_config, ) # Disable AMP for MPS. if torch.backends.mps.is_available(): accelerator.native_amp = False if args.report_to == "wandb": if not is_wandb_available(): raise ImportError("Make sure to install wandb if you want to use it for logging during training.") import wandb # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) logger.info(accelerator.state, main_process_only=False) if accelerator.is_local_main_process: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_warning() diffusers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() diffusers.utils.logging.set_verbosity_error() # If passed along, set the training seed now. if args.seed is not None: set_seed(args.seed) # Handle the repository creation if accelerator.is_main_process: if args.output_dir is not None: os.makedirs(args.output_dir, exist_ok=True) if args.push_to_hub: repo_id = create_repo( repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token ).repo_id # Load the tokenizers tokenizer_one = AutoTokenizer.from_pretrained( args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision, use_fast=False, ) tokenizer_two = AutoTokenizer.from_pretrained( args.pretrained_model_name_or_path, subfolder="tokenizer_2", revision=args.revision, use_fast=False, ) # import correct text encoder classes text_encoder_cls_one = import_model_class_from_model_name_or_path( args.pretrained_model_name_or_path, args.revision ) text_encoder_cls_two = import_model_class_from_model_name_or_path( args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2" ) # Load scheduler and models noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler") # Check for terminal SNR in combination with SNR Gamma text_encoder_one = text_encoder_cls_one.from_pretrained( args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant ) text_encoder_two = text_encoder_cls_two.from_pretrained( args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant ) vae_path = ( args.pretrained_model_name_or_path if args.pretrained_vae_model_name_or_path is None else args.pretrained_vae_model_name_or_path ) vae = AutoencoderKL.from_pretrained( vae_path, subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None, revision=args.revision, variant=args.variant, ) unet = UNet2DConditionModel.from_pretrained( args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant ) # Freeze vae and text encoders. vae.requires_grad_(False) text_encoder_one.requires_grad_(False) text_encoder_two.requires_grad_(False) # Set unet as trainable. unet.train() # For mixed precision training we cast all non-trainable weights to half-precision # as these weights are only used for inference, keeping weights in full precision is not required. weight_dtype = torch.float32 if accelerator.mixed_precision == "fp16": weight_dtype = torch.float16 elif accelerator.mixed_precision == "bf16": weight_dtype = torch.bfloat16 # Move unet, vae and text_encoder to device and cast to weight_dtype # The VAE is in float32 to avoid NaN losses. vae.to(accelerator.device, dtype=torch.float32) text_encoder_one.to(accelerator.device, dtype=weight_dtype) text_encoder_two.to(accelerator.device, dtype=weight_dtype) # Create EMA for the unet. if args.use_ema: ema_unet = UNet2DConditionModel.from_pretrained( args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant ) ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config) if args.enable_npu_flash_attention: if is_torch_npu_available(): logger.info("npu flash attention enabled.") unet.enable_npu_flash_attention() else: raise ValueError("npu flash attention requires torch_npu extensions and is supported only on npu devices.") if args.enable_xformers_memory_efficient_attention: if is_xformers_available(): import xformers xformers_version = version.parse(xformers.__version__) if xformers_version == version.parse("0.0.16"): logger.warning( "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details." ) unet.enable_xformers_memory_efficient_attention() else: raise ValueError("xformers is not available. Make sure it is installed correctly") # `accelerate` 0.16.0 will have better support for customized saving if version.parse(accelerate.__version__) >= version.parse("0.16.0"): # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format def save_model_hook(models, weights, output_dir): if accelerator.is_main_process: if args.use_ema: ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema")) for i, model in enumerate(models): model.save_pretrained(os.path.join(output_dir, "unet")) # make sure to pop weight so that corresponding model is not saved again if weights: weights.pop() def load_model_hook(models, input_dir): if args.use_ema: load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel) ema_unet.load_state_dict(load_model.state_dict()) ema_unet.to(accelerator.device) del load_model for _ in range(len(models)): # pop models so that they are not loaded again model = models.pop() # load diffusers style into model load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet") model.register_to_config(**load_model.config) model.load_state_dict(load_model.state_dict()) del load_model accelerator.register_save_state_pre_hook(save_model_hook) accelerator.register_load_state_pre_hook(load_model_hook) if args.gradient_checkpointing: unet.enable_gradient_checkpointing() # Enable TF32 for faster training on Ampere GPUs, # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices if args.allow_tf32: torch.backends.cuda.matmul.allow_tf32 = True if args.scale_lr: args.learning_rate = ( args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes ) # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs if args.use_8bit_adam: try: import bitsandbytes as bnb except ImportError: raise ImportError( "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`." ) optimizer_class = bnb.optim.AdamW8bit else: optimizer_class = torch.optim.AdamW # Optimizer creation params_to_optimize = unet.parameters() optimizer = optimizer_class( params_to_optimize, lr=args.learning_rate, betas=(args.adam_beta1, args.adam_beta2), weight_decay=args.adam_weight_decay, eps=args.adam_epsilon, ) # Get the datasets: you can either provide your own training and evaluation files (see below) # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub). # In distributed training, the load_dataset function guarantees that only one local process can concurrently # download the dataset. if args.dataset_name is not None: # Downloading and loading a dataset from the hub. dataset = load_dataset( args.dataset_name, args.dataset_config_name, cache_dir=args.cache_dir, data_dir=args.train_data_dir ) else: data_files = {} if args.train_data_dir is not None: data_files["train"] = os.path.join(args.train_data_dir, "**") dataset = load_dataset( "imagefolder", data_files=data_files, cache_dir=args.cache_dir, ) # See more about loading custom images at # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder # Preprocessing the datasets. # We need to tokenize inputs and targets. column_names = dataset["train"].column_names # 6. Get the column names for input/target. dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None) if args.image_column is None: image_column = dataset_columns[0] if dataset_columns is not None else column_names[0] else: image_column = args.image_column if image_column not in column_names: raise ValueError( f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}" ) if args.caption_column is None: caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1] else: caption_column = args.caption_column if caption_column not in column_names: raise ValueError( f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}" ) # Preprocessing the datasets. train_resize = transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR) train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution) train_flip = transforms.RandomHorizontalFlip(p=1.0) train_transforms = transforms.Compose([transforms.ToTensor(), transforms.Normalize([0.5], [0.5])]) def preprocess_train(examples): images = [image.convert("RGB") for image in examples[image_column]] # image aug original_sizes = [] all_images = [] crop_top_lefts = [] for image in images: original_sizes.append((image.height, image.width)) image = train_resize(image) if args.random_flip and random.random() < 0.5: # flip image = train_flip(image) if args.center_crop: y1 = max(0, int(round((image.height - args.resolution) / 2.0))) x1 = max(0, int(round((image.width - args.resolution) / 2.0))) image = train_crop(image) else: y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution)) image = crop(image, y1, x1, h, w) crop_top_left = (y1, x1) crop_top_lefts.append(crop_top_left) image = train_transforms(image) all_images.append(image) examples["original_sizes"] = original_sizes examples["crop_top_lefts"] = crop_top_lefts examples["pixel_values"] = all_images return examples with accelerator.main_process_first(): if args.max_train_samples is not None: dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples)) # Set the training transforms train_dataset = dataset["train"].with_transform(preprocess_train) # Let's first compute all the embeddings so that we can free up the text encoders # from memory. We will pre-compute the VAE encodings too. text_encoders = [text_encoder_one, text_encoder_two] tokenizers = [tokenizer_one, tokenizer_two] compute_embeddings_fn = functools.partial( encode_prompt, text_encoders=text_encoders, tokenizers=tokenizers, proportion_empty_prompts=args.proportion_empty_prompts, caption_column=args.caption_column, ) compute_vae_encodings_fn = functools.partial(compute_vae_encodings, vae=vae) with accelerator.main_process_first(): from datasets.fingerprint import Hasher # fingerprint used by the cache for the other processes to load the result # details: https://github.com/huggingface/diffusers/pull/4038#discussion_r1266078401 new_fingerprint = Hasher.hash(args) new_fingerprint_for_vae = Hasher.hash((vae_path, args)) train_dataset_with_embeddings = train_dataset.map( compute_embeddings_fn, batched=True, new_fingerprint=new_fingerprint ) train_dataset_with_vae = train_dataset.map( compute_vae_encodings_fn, batched=True, batch_size=args.train_batch_size, new_fingerprint=new_fingerprint_for_vae, ) precomputed_dataset = concatenate_datasets( [train_dataset_with_embeddings, train_dataset_with_vae.remove_columns(["image", "text"])], axis=1 ) precomputed_dataset = precomputed_dataset.with_transform(preprocess_train) del compute_vae_encodings_fn, compute_embeddings_fn, text_encoder_one, text_encoder_two del text_encoders, tokenizers if not args.use_lpl: del vae gc.collect() if is_torch_npu_available(): torch_npu.npu.empty_cache() elif torch.cuda.is_available(): torch.cuda.empty_cache() def collate_fn(examples): model_input = torch.stack([torch.tensor(example["model_input"]) for example in examples]) original_sizes = [example["original_sizes"] for example in examples] crop_top_lefts = [example["crop_top_lefts"] for example in examples] prompt_embeds = torch.stack([torch.tensor(example["prompt_embeds"]) for example in examples]) pooled_prompt_embeds = torch.stack([torch.tensor(example["pooled_prompt_embeds"]) for example in examples]) return { "model_input": model_input, "prompt_embeds": prompt_embeds, "pooled_prompt_embeds": pooled_prompt_embeds, "original_sizes": original_sizes, "crop_top_lefts": crop_top_lefts, } # DataLoaders creation: train_dataloader = torch.utils.data.DataLoader( precomputed_dataset, shuffle=True, collate_fn=collate_fn, batch_size=args.train_batch_size, num_workers=args.dataloader_num_workers, ) # Scheduler and math around the number of training steps. overrode_max_train_steps = False num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) if args.max_train_steps is None: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch overrode_max_train_steps = True lr_scheduler = get_scheduler( args.lr_scheduler, optimizer=optimizer, num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, ) # Prepare everything with our `accelerator`. unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( unet, optimizer, train_dataloader, lr_scheduler ) if args.use_ema: ema_unet.to(accelerator.device) # We need to recalculate our total training steps as the size of the training dataloader may have changed. num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) if overrode_max_train_steps: args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch # Afterwards we recalculate our number of training epochs args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) # We need to initialize the trackers we use, and also store our configuration. # The trackers initializes automatically on the main process. if accelerator.is_main_process: accelerator.init_trackers("text2image-fine-tune-sdxl", config=vars(args)) if args.use_lpl: lpl_fn = LatentPerceptualLoss( vae=vae, loss_type=args.lpl_loss_type, grad_ckpt=args.gradient_checkpointing, pow_law=args.lpl_pow_law, norm_type=args.lpl_norm_type, num_mid_blocks=args.lpl_num_blocks, feature_type="feature", remove_outliers=args.lpl_remove_outliers, ) lpl_fn.to(accelerator.device) else: lpl_fn = None # Function for unwrapping if torch.compile() was used in accelerate. def unwrap_model(model): model = accelerator.unwrap_model(model) model = model._orig_mod if is_compiled_module(model) else model return model if torch.backends.mps.is_available() or "playground" in args.pretrained_model_name_or_path: autocast_ctx = nullcontext() else: autocast_ctx = torch.autocast(accelerator.device.type) # Train! total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps logger.info("***** Running training *****") logger.info(f" Num examples = {len(precomputed_dataset)}") logger.info(f" Num Epochs = {args.num_train_epochs}") logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") logger.info(f" Total optimization steps = {args.max_train_steps}") global_step = 0 first_epoch = 0 # Potentially load in the weights and states from a previous save if args.resume_from_checkpoint: if args.resume_from_checkpoint != "latest": path = os.path.basename(args.resume_from_checkpoint) else: # Get the most recent checkpoint dirs = os.listdir(args.output_dir) dirs = [d for d in dirs if d.startswith("checkpoint")] dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) path = dirs[-1] if len(dirs) > 0 else None if path is None: accelerator.print( f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." ) args.resume_from_checkpoint = None initial_global_step = 0 else: accelerator.print(f"Resuming from checkpoint {path}") accelerator.load_state(os.path.join(args.output_dir, path)) global_step = int(path.split("-")[1]) initial_global_step = global_step first_epoch = global_step // num_update_steps_per_epoch else: initial_global_step = 0 progress_bar = tqdm( range(0, args.max_train_steps), initial=initial_global_step, desc="Steps", # Only show the progress bar once on each machine. disable=not accelerator.is_local_main_process, ) # Get scheduler alphas and sigmas for LPL z0_hat calculation alphas_cumprod = noise_scheduler.alphas_cumprod.to(accelerator.device) for epoch in range(first_epoch, args.num_train_epochs): train_loss = 0.0 for step, batch in enumerate(train_dataloader): with accelerator.accumulate(unet): # Sample noise that we'll add to the latents model_input = batch["model_input"].to(accelerator.device) noise = torch.randn_like(model_input) if args.noise_offset: # https://www.crosslabs.org//blog/diffusion-with-offset-noise noise += args.noise_offset * torch.randn( (model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device ) bsz = model_input.shape[0] if args.timestep_bias_strategy == "none": # Sample a random timestep for each image without bias. timesteps = torch.randint( 0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device ) else: # Sample a random timestep for each image, potentially biased by the timestep weights. # Biasing the timestep weights allows us to spend less time training irrelevant timesteps. weights = generate_timestep_weights(args, noise_scheduler.config.num_train_timesteps).to( model_input.device ) timesteps = torch.multinomial(weights, bsz, replacement=True).long() # Add noise to the model input according to the noise magnitude at each timestep # (this is the forward diffusion process) noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps).to(dtype=weight_dtype) # time ids def compute_time_ids(original_size, crops_coords_top_left): # Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids target_size = (args.resolution, args.resolution) add_time_ids = list(original_size + crops_coords_top_left + target_size) add_time_ids = torch.tensor([add_time_ids], device=accelerator.device, dtype=weight_dtype) return add_time_ids add_time_ids = torch.cat( [compute_time_ids(s, c) for s, c in zip(batch["original_sizes"], batch["crop_top_lefts"])] ) # Predict the noise residual unet_added_conditions = {"time_ids": add_time_ids} prompt_embeds = batch["prompt_embeds"].to(accelerator.device, dtype=weight_dtype) pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(accelerator.device) unet_added_conditions.update({"text_embeds": pooled_prompt_embeds}) model_pred = unet( noisy_model_input, timesteps, prompt_embeds, added_cond_kwargs=unet_added_conditions, return_dict=False, )[0] # Get the target for loss depending on the prediction type if args.prediction_type is not None: # set prediction_type of scheduler if defined noise_scheduler.register_to_config(prediction_type=args.prediction_type) if noise_scheduler.config.prediction_type == "epsilon": target = noise elif noise_scheduler.config.prediction_type == "v_prediction": target = noise_scheduler.get_velocity(model_input, noise, timesteps) elif noise_scheduler.config.prediction_type == "sample": # We set the target to latents here, but the model_pred will return the noise sample prediction. target = model_input # We will have to subtract the noise residual from the prediction to get the target sample. model_pred = model_pred - noise else: raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") if args.snr_gamma is None: loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") else: # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556. # Since we predict the noise instead of x_0, the original formulation is slightly changed. # This is discussed in Section 4.2 of the same paper. snr = compute_snr(noise_scheduler, timesteps) mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min( dim=1 )[0] if noise_scheduler.config.prediction_type == "epsilon": mse_loss_weights = mse_loss_weights / snr elif noise_scheduler.config.prediction_type == "v_prediction": mse_loss_weights = mse_loss_weights / (snr + 1) loss = F.mse_loss(model_pred.float(), target.float(), reduction="none") loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights loss = loss.mean() lpl_loss_value = torch.tensor(0.0, device=accelerator.device) if args.use_lpl and lpl_fn is not None and global_step >= args.lpl_start: # Apply LPL only below the timestep threshold lpl_mask = timesteps < args.lpl_t_threshold if lpl_mask.any(): # Select samples that meet the threshold masked_indices = torch.where(lpl_mask)[0] z0_masked = model_input[masked_indices] zt_masked = noisy_model_input[masked_indices] t_masked = timesteps[masked_indices] model_pred_masked = model_pred[masked_indices] # Calculate z0_hat for the masked samples alpha_t = alphas_cumprod[t_masked].sqrt().to(torch.float32) sigma_t = (1 - alphas_cumprod[t_masked]).sqrt().to(torch.float32) alpha_t = alpha_t.view(-1, 1, 1, 1) sigma_t = sigma_t.view(-1, 1, 1, 1) if noise_scheduler.config.prediction_type == "epsilon": z0_hat_masked = (zt_masked.float() - sigma_t * model_pred_masked.float()) / alpha_t elif noise_scheduler.config.prediction_type == "v_prediction": z0_hat_masked = alpha_t * zt_masked.float() - sigma_t * model_pred_masked.float() else: # sample prediction z0_hat_masked = model_pred_masked.float() with accelerator.autocast(): lpl_loss_value = lpl_fn.get_loss(z0_hat_masked, z0_masked) if args.lpl_scale: if args.snr_gamma is not None: # Use SNR-based weights if available snr = compute_snr(noise_scheduler, t_masked) snr_weights = torch.stack( [snr, args.snr_gamma * torch.ones_like(t_masked)], dim=1 ).min(dim=1)[0] if noise_scheduler.config.prediction_type == "epsilon": snr_weights = snr_weights / snr elif noise_scheduler.config.prediction_type == "v_prediction": snr_weights = snr_weights / (snr + 1) lpl_loss_value = (lpl_loss_value * snr_weights).mean() else: # If no SNR weighting, just use mean lpl_loss_value = lpl_loss_value.mean() else: lpl_loss_value = lpl_loss_value.mean() # Combine losses total_loss = loss + args.lpl_weight * lpl_loss_value # Gather the losses across all processes for logging avg_loss = accelerator.gather(total_loss.repeat(args.train_batch_size)).mean() train_loss += avg_loss.item() / args.gradient_accumulation_steps # Backpropagate accelerator.backward(total_loss) if accelerator.sync_gradients: params_to_clip = unet.parameters() accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm) optimizer.step() lr_scheduler.step() optimizer.zero_grad() # Checks if the accelerator has performed an optimization step behind the scenes if accelerator.sync_gradients: if args.use_ema: ema_unet.step(unet.parameters()) progress_bar.update(1) global_step += 1 # Enhanced logging for LPL metrics log_data = { "train_loss": train_loss, "diffusion_loss": loss.item(), "learning_rate": lr_scheduler.get_last_lr()[0], } if args.use_lpl and lpl_fn is not None and global_step >= args.lpl_start: if lpl_mask.any(): # LPL application statistics log_data.update( { "lpl/loss": lpl_loss_value.item(), "lpl/num_samples": lpl_mask.sum().item(), "lpl/application_ratio": lpl_mask.float().mean().item(), "lpl/weight": args.lpl_weight, "lpl/weighted_loss": (args.lpl_weight * lpl_loss_value).item(), } ) # SNR statistics for LPL-applied samples if args.snr_gamma is not None: snr_values = snr[masked_indices] log_data.update( { "lpl/snr_mean": snr_values.mean().item(), "lpl/snr_std": snr_values.std().item(), "lpl/snr_min": snr_values.min().item(), "lpl/snr_max": snr_values.max().item(), } ) # Feature statistics if available if hasattr(lpl_fn, "last_feature_stats"): for layer_idx, stats in enumerate(lpl_fn.last_feature_stats): log_data.update( { f"lpl/features/layer_{layer_idx}/mean": stats["mean"], f"lpl/features/layer_{layer_idx}/std": stats["std"], f"lpl/features/layer_{layer_idx}/outlier_ratio": stats.get( "outlier_ratio", 0.0 ), } ) # Memory usage if available if torch.cuda.is_available(): log_data.update( { "lpl/memory/allocated": torch.cuda.memory_allocated() / 1024**2, # MB "lpl/memory/reserved": torch.cuda.memory_reserved() / 1024**2, # MB } ) # Log to accelerator accelerator.log(log_data, step=global_step) # Update progress bar with more metrics progress_bar_logs = { "loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0], } if args.use_lpl and lpl_loss_value.item() > 0: progress_bar_logs.update( { "lpl": lpl_loss_value.item(), "lpl_ratio": lpl_mask.float().mean().item() if lpl_mask.any() else 0.0, } ) progress_bar.set_postfix(**progress_bar_logs) # DeepSpeed requires saving weights on every device; saving weights only on the main process would cause issues. if accelerator.distributed_type == DistributedType.DEEPSPEED or accelerator.is_main_process: if global_step % args.checkpointing_steps == 0: # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` if args.checkpoints_total_limit is not None: checkpoints = os.listdir(args.output_dir) checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints if len(checkpoints) >= args.checkpoints_total_limit: num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 removing_checkpoints = checkpoints[0:num_to_remove] logger.info( f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" ) logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") for removing_checkpoint in removing_checkpoints: removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) shutil.rmtree(removing_checkpoint) save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") accelerator.save_state(save_path) logger.info(f"Saved state to {save_path}") logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} progress_bar.set_postfix(**logs) if global_step >= args.max_train_steps: break if accelerator.is_main_process: if args.validation_prompt is not None and epoch % args.validation_epochs == 0: logger.info( f"Running validation... \n Generating {args.num_validation_images} images with prompt:" f" {args.validation_prompt}." ) if args.use_ema: # Store the UNet parameters temporarily and load the EMA parameters to perform inference. ema_unet.store(unet.parameters()) ema_unet.copy_to(unet.parameters()) # create pipeline vae = AutoencoderKL.from_pretrained( vae_path, subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None, revision=args.revision, variant=args.variant, ) pipeline = StableDiffusionXLPipeline.from_pretrained( args.pretrained_model_name_or_path, vae=vae, unet=accelerator.unwrap_model(unet), revision=args.revision, variant=args.variant, torch_dtype=weight_dtype, ) if args.prediction_type is not None: scheduler_args = {"prediction_type": args.prediction_type} pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args) pipeline = pipeline.to(accelerator.device) pipeline.set_progress_bar_config(disable=True) # run inference generator = ( torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None ) pipeline_args = {"prompt": args.validation_prompt} with autocast_ctx: images = [ pipeline(**pipeline_args, generator=generator, num_inference_steps=25).images[0] for _ in range(args.num_validation_images) ] for tracker in accelerator.trackers: if tracker.name == "tensorboard": np_images = np.stack([np.asarray(img) for img in images]) tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC") if tracker.name == "wandb": tracker.log( { "validation": [ wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images) ] } ) del pipeline if is_torch_npu_available(): torch_npu.npu.empty_cache() elif torch.cuda.is_available(): torch.cuda.empty_cache() if args.use_ema: # Switch back to the original UNet parameters. ema_unet.restore(unet.parameters()) accelerator.wait_for_everyone() if accelerator.is_main_process: unet = unwrap_model(unet) if args.use_ema: ema_unet.copy_to(unet.parameters()) # Serialize pipeline. vae = AutoencoderKL.from_pretrained( vae_path, subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None, revision=args.revision, variant=args.variant, torch_dtype=weight_dtype, ) pipeline = StableDiffusionXLPipeline.from_pretrained( args.pretrained_model_name_or_path, unet=unet, vae=vae, revision=args.revision, variant=args.variant, torch_dtype=weight_dtype, ) if args.prediction_type is not None: scheduler_args = {"prediction_type": args.prediction_type} pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args) pipeline.save_pretrained(args.output_dir) # run inference images = [] if args.validation_prompt and args.num_validation_images > 0: pipeline = pipeline.to(accelerator.device) generator = ( torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None ) with autocast_ctx: images = [ pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0] for _ in range(args.num_validation_images) ] for tracker in accelerator.trackers: if tracker.name == "tensorboard": np_images = np.stack([np.asarray(img) for img in images]) tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC") if tracker.name == "wandb": tracker.log( { "test": [ wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images) ] } ) if args.push_to_hub: save_model_card( repo_id=repo_id, images=images, validation_prompt=args.validation_prompt, base_model=args.pretrained_model_name_or_path, dataset_name=args.dataset_name, repo_folder=args.output_dir, vae_path=args.pretrained_vae_model_name_or_path, ) upload_folder( repo_id=repo_id, folder_path=args.output_dir, commit_message="End of training", ignore_patterns=["step_*", "epoch_*"], ) accelerator.end_training() if __name__ == "__main__": args = parse_args() main(args)