#!/usr/bin/env python # coding=utf-8 # Copyright 2023 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 import argparse import copy import gc import hashlib import itertools import logging import math import os import shutil import warnings from pathlib import Path import numpy as np import torch import torch.nn as nn import torch.nn.functional as F import torch.utils.checkpoint import transformers import deepspeed from accelerate import Accelerator from accelerate.logging import get_logger from accelerate.utils import ProjectConfiguration, set_seed from huggingface_hub import create_repo, upload_folder from packaging import version from PIL import Image from PIL.ImageOps import exif_transpose from torch.utils.data import Dataset from torchvision import transforms from tqdm.auto import tqdm import diffusers from diffusers.loaders import ( LoraLoaderMixin, ) from diffusers.optimization import get_scheduler from diffusers.utils.import_utils import is_xformers_available from diffusers.utils.torch_utils import is_compiled_module from diffusers.utils import ( check_min_version, convert_all_state_dict_to_peft, convert_state_dict_to_diffusers, convert_state_dict_to_kohya, convert_unet_state_dict_to_peft, is_wandb_available, ) from accelerate.utils import DistributedDataParallelKwargs, DistributedType, ProjectConfiguration, set_seed from peft import LoraConfig, set_peft_model_state_dict, get_peft_model from peft.utils import get_peft_model_state_dict from safetensors.torch import save_file from kolors.pipelines.pipeline_stable_diffusion_xl_chatglm_256 import StableDiffusionXLPipeline from kolors.models.modeling_chatglm import ChatGLMModel from kolors.models.tokenization_chatglm import ChatGLMTokenizer from diffusers import ( AutoencoderKL, DDPMScheduler, UNet2DConditionModel, EulerDiscreteScheduler, ) # Will error if the minimal version of diffusers is not installed. Remove at your own risks. check_min_version("0.22.0.dev0") logger = get_logger(__name__) def parse_args(input_args=None): parser = argparse.ArgumentParser(description="Simple example of a training script.") 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( "--unet_id", type=str, default=None, required=False, help="Path to pretrained model or model identifier from huggingface.co/models.", ) parser.add_argument( "--revision", type=str, default=None, required=False, help="Revision of pretrained model identifier from huggingface.co/models.", ) parser.add_argument( "--tokenizer_name", type=str, default=None, help="Pretrained tokenizer name or path if not the same as model_name", ) parser.add_argument( "--instance_data_dir", type=str, default=None, required=True, help="A folder containing the training data of instance images.", ) parser.add_argument( "--class_data_dir", type=str, default=None, required=False, help="A folder containing the training data of class images.", ) parser.add_argument( "--instance_prompt", type=str, default=None, required=True, help="The prompt with identifier specifying the instance", ) parser.add_argument( "--class_prompt", type=str, default=None, help="The prompt to specify images in the same class as provided instance images.", ) 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=50, help=( "Run dreambooth validation every X epochs. Dreambooth validation consists of running the prompt" " `args.validation_prompt` multiple times: `args.num_validation_images`." ), ) parser.add_argument( "--with_prior_preservation", default=False, action="store_true", help="Flag to add prior preservation loss.", ) parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.") parser.add_argument( "--num_class_images", type=int, default=100, help=( "Minimal class images for prior preservation loss. If there are not enough images already present in" " class_data_dir, additional images will be sampled with class_prompt." ), ) parser.add_argument( "--output_dir", type=str, default="lora-dreambooth-model", help="The output directory where the model predictions and checkpoints will be written.", ) parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") parser.add_argument( "--resolution", type=int, default=512, 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( "--train_text_encoder", action="store_true", help="Whether to train the text encoder. If set, the text encoder should be float32 precision.", ) parser.add_argument( "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader." ) parser.add_argument( "--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images." ) parser.add_argument("--num_train_epochs", type=int, default=1) 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=5e-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( "--lr_num_cycles", type=int, default=1, help="Number of hard resets of the lr in cosine_with_restarts scheduler.", ) parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.") 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_weight_decay_text_encoder", type=float, default=1e-03, help="Weight decay to use for text_encoder" ) 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( "--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( "--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( "--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( "--prior_generation_precision", type=str, default=None, choices=["no", "fp32", "fp16", "bf16"], help=( "Choose prior generation precision between fp32, fp16 and bf16 (bfloat16). Bf16 requires PyTorch >=" " 1.10.and an Nvidia Ampere GPU. Default to fp16 if a GPU is available else fp32." ), ) parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") parser.add_argument( "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers." ) parser.add_argument( "--pre_compute_text_embeddings", action="store_true", help="Whether or not to pre-compute text embeddings. If text embeddings are pre-computed, the text encoder will not be kept in memory during training and will leave more GPU memory available for training the rest of the model. This is not compatible with `--train_text_encoder`.", ) parser.add_argument( "--tokenizer_max_length", type=int, default=None, required=False, help="The maximum length of the tokenizer. If not set, will default to the tokenizer's max length.", ) parser.add_argument( "--text_encoder_use_attention_mask", action="store_true", required=False, help="Whether to use attention mask for the text encoder", ) parser.add_argument( "--validation_images", required=False, default=None, nargs="+", help="Optional set of images to use for validation. Used when the target pipeline takes an initial image as input such as when training image variation or superresolution.", ) parser.add_argument( "--class_labels_conditioning", required=False, default=None, help="The optional `class_label` conditioning to pass to the unet, available values are `timesteps`.", ) parser.add_argument( "--rank", type=int, default=4, help=("The dimension of the LoRA update matrices."), ) parser.add_argument( "--img_repeat_nums", type=int, default=1, help=("Repeat training images."), ) parser.add_argument( "--noise_offset", type=float, default=0, help=("The scale of noise offset.") ) parser.add_argument( "--use_preffix_prompt", action="store_true", help="Whether or not to use lora preffix prompt!!!" ) parser.add_argument( "--text_encoder_lr", type=float, default=5e-6, help="Text encoder learning rate to use.", ) 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 if args.with_prior_preservation: if args.class_data_dir is None: raise ValueError("You must specify a data directory for class images.") if args.class_prompt is None: raise ValueError("You must specify prompt for class images.") else: # logger is not available yet if args.class_data_dir is not None: warnings.warn("You need not use --class_data_dir without --with_prior_preservation.") if args.class_prompt is not None: warnings.warn("You need not use --class_prompt without --with_prior_preservation.") if args.train_text_encoder and args.pre_compute_text_embeddings: raise ValueError("`--train_text_encoder` cannot be used with `--pre_compute_text_embeddings`") return args def get_pipe(ckpt_dir): text_encoder = ChatGLMModel.from_pretrained( f'{ckpt_dir}/text_encoder', torch_dtype=torch.float16).half() tokenizer = ChatGLMTokenizer.from_pretrained(f'{ckpt_dir}/text_encoder') vae = AutoencoderKL.from_pretrained(f"{ckpt_dir}/vae", revision=None).half() scheduler = EulerDiscreteScheduler.from_pretrained(f"{ckpt_dir}/scheduler") unet = UNet2DConditionModel.from_pretrained(f"{ckpt_dir}/unet", revision=None).half() pipe = StableDiffusionXLPipeline( vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, force_zeros_for_empty_prompt=False) return pipe class DreamBoothDataset(Dataset): """ A dataset to prepare the instance and class images with the prompts for fine-tuning the model. It pre-processes the images and the tokenizes prompts. """ def __init__( self, instance_data_root, instance_prompt, tokenizer, class_data_root=None, class_prompt=None, class_num=None, size=512, center_crop=False, encoder_hidden_states=None, class_prompt_encoder_hidden_states=None, tokenizer_max_length=None, img_repeat_nums = 1, use_preffix_prompt=False ): self.size = size self.center_crop = center_crop self.tokenizer = tokenizer self.encoder_hidden_states = encoder_hidden_states self.class_prompt_encoder_hidden_states = class_prompt_encoder_hidden_states self.tokenizer_max_length = tokenizer_max_length self.instance_data_root = Path(instance_data_root) if not self.instance_data_root.exists(): raise ValueError("Instance images root doesn't exists.") file_list = list(Path(instance_data_root).iterdir()) self.instance_images_path = [file for file in Path(instance_data_root).iterdir() if file.is_file() and not file.name.endswith('.txt') and not file.name.endswith('.npy')] self.instance_prompt_path = [file for file in Path(instance_data_root).iterdir() if file.is_file() and file.name.endswith('.txt')] # self.instance_images_path = list(Path(instance_data_root).iterdir()) self.instance_images_path *= img_repeat_nums self.instance_prompt_path *= img_repeat_nums self.num_instance_images = len(self.instance_images_path) self.instance_prompt = instance_prompt self._length = self.num_instance_images self.use_preffix_prompt = use_preffix_prompt if class_data_root is not None: self.class_data_root = Path(class_data_root) self.class_data_root.mkdir(parents=True, exist_ok=True) self.class_images_path = list(self.class_data_root.iterdir()) if class_num is not None: self.num_class_images = min(len(self.class_images_path), class_num) else: self.num_class_images = len(self.class_images_path) self._length = max(self.num_class_images, self.num_instance_images) self.class_prompt = class_prompt else: self.class_data_root = None self.image_transforms = transforms.Compose( [ transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR), transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size), transforms.ToTensor(), transforms.Normalize([0.5], [0.5]), ] ) def __len__(self): return self._length def __getitem__(self, index): example = {} img_path = str(self.instance_images_path[index % self.num_instance_images]) instance_image = Image.open(img_path).convert("RGB") instance_image = exif_transpose(instance_image) matching_prompt_files = img_path.rsplit('.',1)[0] + '.txt' if not os.path.exists( matching_prompt_files ): training_prompt = self.instance_prompt else: with open(matching_prompt_files, 'r', encoding="utf-8") as file: training_prompt = file.read() if self.use_preffix_prompt: training_prompt = self.instance_prompt + '，' + training_prompt print(training_prompt) # print(f'using prompt {training_prompt}!!!') origin_width, origin_height = instance_image.size example["instance_images"] = self.image_transforms(instance_image) if self.encoder_hidden_states is not None: example["instance_prompt_ids"] = self.encoder_hidden_states else: text_inputs = self.tokenizer( training_prompt, padding='max_length', max_length=256, truncation=True, return_tensors="pt" ) example["instance_prompt_ids"] = text_inputs.input_ids.squeeze(0) example["instance_attention_mask"] = text_inputs.attention_mask.squeeze(0) example["instance_position_ids"] = text_inputs.position_ids.squeeze(0) resize_ratio = max(self.size / origin_width, self.size / origin_height) resize_size = (round(origin_height * resize_ratio), round(origin_width * resize_ratio)) offsite_width = (resize_size[1] - self.size) // 2 offsite_height = (resize_size[0] - self.size) // 2 example["instance_add_ids"] = torch.tensor([resize_size[0], resize_size[1], offsite_height, offsite_width, self.size, self.size]) if self.class_data_root: class_image = Image.open(self.class_images_path[index % self.num_class_images]) class_image = exif_transpose(class_image) if not class_image.mode == "RGB": class_image = class_image.convert("RGB") origin_width, origin_height = class_image.size example["class_images"] = self.image_transforms(class_image) if self.class_prompt_encoder_hidden_states is not None: example["class_prompt_ids"] = self.class_prompt_encoder_hidden_states else: class_text_inputs = self.tokenizer( self.class_prompt, padding='max_length', max_length=256, truncation=True, return_tensors="pt" ) example["class_prompt_ids"] = class_text_inputs.input_ids.squeeze(0) example["class_attention_mask"] = class_text_inputs.attention_mask.squeeze(0) example["class_position_ids"] = class_text_inputs.position_ids.squeeze(0) resize_ratio = max(self.size / origin_width, self.size / origin_height) resize_size = (round(origin_height * resize_ratio), round(origin_width * resize_ratio)) offsite_width = (resize_size[1] - self.size) // 2 offsite_height = (resize_size[0] - self.size) // 2 example["class_add_ids"] = torch.tensor([resize_size[0], resize_size[1], offsite_height, offsite_width, self.size, self.size]) return example def collate_fn(examples, with_prior_preservation=False): has_attention_mask = "instance_attention_mask" in examples[0] input_ids = [example["instance_prompt_ids"] for example in examples] pixel_values = [example["instance_images"] for example in examples] add_time_ids = [example["instance_add_ids"] for example in examples] attention_mask = [example["instance_attention_mask"] for example in examples] position_ids = [example["instance_position_ids"] for example in examples] # Concat class and instance examples for prior preservation. # We do this to avoid doing two forward passes. if with_prior_preservation: input_ids += [example["class_prompt_ids"] for example in examples] pixel_values += [example["class_images"] for example in examples] add_time_ids += [example["class_add_ids"] for example in examples] attention_mask += [example["class_attention_mask"] for example in examples] position_ids += [example["class_position_ids"] for example in examples] pixel_values = torch.stack(pixel_values) pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float() input_ids = torch.stack([input_id for input_id in input_ids]) attention_mask = torch.stack([msk for msk in attention_mask]) add_time_ids = torch.stack([add_id for add_id in add_time_ids]) position_ids = torch.stack([pid for pid in position_ids]) batch = { "input_ids": input_ids, "pixel_values": pixel_values, "add_time_ids":add_time_ids, "attention_mask":attention_mask, "position_ids": position_ids } return batch class PromptDataset(Dataset): "A simple dataset to prepare the prompts to generate class images on multiple GPUs." def __init__(self, prompt, num_samples): self.prompt = prompt self.num_samples = num_samples def __len__(self): return self.num_samples def __getitem__(self, index): example = {} example["prompt"] = self.prompt example["index"] = index return example def tokenize_prompt(tokenizer, prompt, tokenizer_max_length=None): if tokenizer_max_length is not None: max_length = tokenizer_max_length else: max_length = tokenizer.model_max_length text_inputs = tokenizer( prompt, truncation=True, padding="max_length", max_length=max_length, return_tensors="pt", ) return text_inputs def encode_prompt(text_encoder, input_ids, attention_mask, text_encoder_use_attention_mask=None): text_input_ids = input_ids.to(text_encoder.device) if text_encoder_use_attention_mask: attention_mask = attention_mask.to(text_encoder.device) else: attention_mask = None prompt_embeds = text_encoder( text_input_ids, attention_mask=attention_mask, ) prompt_embeds = prompt_embeds[0] return prompt_embeds class SuperModel(nn.Module): def __init__(self, unet, text_encoder): super(SuperModel, self).__init__() self.unet = unet self.text_encoder = text_encoder def main(args): logging_dir = Path(args.output_dir, args.logging_dir) deepspeed.init_distributed() accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir) accelerator = Accelerator( gradient_accumulation_steps=args.gradient_accumulation_steps, mixed_precision=args.mixed_precision, log_with=args.report_to, project_config=accelerator_project_config, ) 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 # Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate # This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models. # TODO (sayakpaul): Remove this check when gradient accumulation with two models is enabled in accelerate. if args.train_text_encoder and args.gradient_accumulation_steps > 1 and accelerator.num_processes > 1: raise ValueError( "Gradient accumulation is not supported when training the text encoder in distributed training. " "Please set gradient_accumulation_steps to 1. This feature will be supported in the future." ) # 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: transformers.utils.logging.set_verbosity_warning() diffusers.utils.logging.set_verbosity_info() else: 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) # Generate class images if prior preservation is enabled. if args.with_prior_preservation: class_images_dir = Path(args.class_data_dir) if not class_images_dir.exists(): class_images_dir.mkdir(parents=True) cur_class_images = len(list(class_images_dir.iterdir())) if cur_class_images < args.num_class_images: torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32 if args.prior_generation_precision == "fp32": torch_dtype = torch.float32 elif args.prior_generation_precision == "mixed_precision": torch_dtype = torch.float16 elif args.prior_generation_precision == "bf16": torch_dtype = torch.bfloat16 pipeline = get_pipe(args.pretrained_model_name_or_path) pipeline.set_progress_bar_config(disable=True) num_new_images = args.num_class_images - cur_class_images logger.info(f"Number of class images to sample: {num_new_images}.") sample_dataset = PromptDataset(args.class_prompt, num_new_images) sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size) sample_dataloader = accelerator.prepare(sample_dataloader) pipeline.to(accelerator.device) for example in tqdm( sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process ): images = pipeline(example["prompt"]).images for i, image in enumerate(images): hash_image = hashlib.sha1(image.tobytes()).hexdigest() image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg" image.save(image_filename) del pipeline if torch.cuda.is_available(): torch.cuda.empty_cache() # 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 tokenizer tokenizer = ChatGLMTokenizer.from_pretrained( os.path.join(args.pretrained_model_name_or_path, "text_encoder"), revision=args.revision, trust_remote_code=True) # Load scheduler and models noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler") text_encoder = ChatGLMModel.from_pretrained( args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, trust_remote_code=True) vae = AutoencoderKL.from_pretrained( args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision ) if args.unet_id is None: unet = UNet2DConditionModel.from_pretrained( f'{args.pretrained_model_name_or_path}/unet', low_cpu_mem_usage=False, ignore_mismatched_sizes=True, revision=args.revision ) print(f'init from unet_id {args.pretrained_model_name_or_path}/unet!!!') else: unet = UNet2DConditionModel.from_pretrained( args.unet_id, low_cpu_mem_usage=False, ignore_mismatched_sizes=True, revision=args.revision ) print(f'init from unet_id {args.unet_id}!!!') # We only train the additional adapter LoRA layers if vae is not None: vae.requires_grad_(False) text_encoder.requires_grad_(False) unet.requires_grad_(False) # For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) 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 unet.to(accelerator.device, dtype=weight_dtype) if vae is not None: vae.to(accelerator.device, dtype=torch.float32) text_encoder.to(accelerator.device, dtype=weight_dtype) 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.warn( "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") if args.gradient_checkpointing: unet.enable_gradient_checkpointing() if args.train_text_encoder: text_encoder.gradient_checkpointing_enable() unet_lora_config = LoraConfig( r=args.rank, lora_alpha=args.rank, init_lora_weights="gaussian", target_modules=["to_k", "to_q", "to_v", "to_out.0"], ) unet.add_adapter(unet_lora_config) if args.train_text_encoder: text_lora_config = LoraConfig( r=args.rank, lora_alpha=args.rank, init_lora_weights="gaussian", target_modules=["query_key_value"], ) text_encoder = get_peft_model(text_encoder, text_lora_config) text_encoder.print_trainable_parameters() def unwrap_model(model): model = accelerator.unwrap_model(model) model = model._orig_mod if is_compiled_module(model) else model return model def save_model_hook(models, weights, output_dir): #Save unet & text encoder LoRA weights if accelerator.is_main_process: # there are only two options here. Either are just the unet attn processor layers # or there are the unet and text encoder attn layers unet_lora_layers_to_save = None unet_lora_layers_to_save = convert_state_dict_to_diffusers(get_peft_model_state_dict(unwrap_model(models[0].unet))) unwrap_model(models[0].text_encoder).save_pretrained(output_dir) StableDiffusionXLPipeline.save_lora_weights(output_dir,unet_lora_layers=unet_lora_layers_to_save) def save_model_hook_no_text(models, weights, output_dir): #Save only unet LoRA weights if accelerator.is_main_process: # there are only two options here. Either are just the unet attn processor layers # or there are the unet and text encoder attn layers unet_lora_layers_to_save = None unet_lora_layers_to_save = convert_state_dict_to_diffusers(get_peft_model_state_dict(unwrap_model(models[0]))) StableDiffusionXLPipeline.save_lora_weights(output_dir,unet_lora_layers=unet_lora_layers_to_save) def load_model_hook(models, input_dir): unet_ = accelerator.unwrap_model(unet) lora_state_dict, _ = LoraLoaderMixin.lora_state_dict(input_dir) unet_state_dict = {f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")} unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict) incompatible_keys = set_peft_model_state_dict(unet_, unet_state_dict, adapter_name="default") if incompatible_keys is not None: # check only for unexpected keys unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None) if unexpected_keys: logger.warning( f"Loading adapter weights from state_dict led to unexpected keys not found in the model: " f" {unexpected_keys}. " ) if args.train_text_encoder: text_encoder = PeftModel.from_pretrained(text_encoder, input_dir) if args.train_text_encoder: accelerator.register_save_state_pre_hook(save_model_hook) else: accelerator.register_save_state_pre_hook(save_model_hook_no_text) accelerator.register_load_state_pre_hook(load_model_hook) # 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 super_model = SuperModel(unet=unet, text_encoder=text_encoder) unet_lora_parameters = list(filter(lambda p: p.requires_grad, super_model.unet.parameters())) if args.train_text_encoder: text_lora_parameters = list(filter(lambda p: p.requires_grad, super_model.text_encoder.parameters())) # Optimization parameters unet_lora_parameters_with_lr = {"params": unet_lora_parameters, "lr": args.learning_rate} if args.train_text_encoder: # different learning rate for text encoder and unet text_lora_parameters_with_lr = { "params": text_lora_parameters, "weight_decay": args.adam_weight_decay_text_encoder, "lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate, } params_to_optimize = [ unet_lora_parameters_with_lr, text_lora_parameters_with_lr, ] else: params_to_optimize = [unet_lora_parameters_with_lr] 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, ) if args.pre_compute_text_embeddings: def compute_text_embeddings(prompt): with torch.no_grad(): text_inputs = tokenize_prompt(tokenizer, prompt, tokenizer_max_length=args.tokenizer_max_length) prompt_embeds = encode_prompt( text_encoder, text_inputs.input_ids, text_inputs.attention_mask, text_encoder_use_attention_mask=args.text_encoder_use_attention_mask, ) return prompt_embeds pre_computed_encoder_hidden_states = compute_text_embeddings(args.instance_prompt) validation_prompt_negative_prompt_embeds = compute_text_embeddings("") if args.validation_prompt is not None: validation_prompt_encoder_hidden_states = compute_text_embeddings(args.validation_prompt) else: validation_prompt_encoder_hidden_states = None if args.class_prompt is not None: pre_computed_class_prompt_encoder_hidden_states = compute_text_embeddings(args.class_prompt) else: pre_computed_class_prompt_encoder_hidden_states = None text_encoder = None tokenizer = None gc.collect() torch.cuda.empty_cache() else: pre_computed_encoder_hidden_states = None validation_prompt_encoder_hidden_states = None validation_prompt_negative_prompt_embeds = None pre_computed_class_prompt_encoder_hidden_states = None # Dataset and DataLoaders creation: train_dataset = DreamBoothDataset( instance_data_root=args.instance_data_dir, instance_prompt=args.instance_prompt, class_data_root=args.class_data_dir if args.with_prior_preservation else None, class_prompt=args.class_prompt, class_num=args.num_class_images, tokenizer=tokenizer, size=args.resolution, center_crop=args.center_crop, encoder_hidden_states=pre_computed_encoder_hidden_states, class_prompt_encoder_hidden_states=pre_computed_class_prompt_encoder_hidden_states, tokenizer_max_length=args.tokenizer_max_length, img_repeat_nums= args.img_repeat_nums, use_preffix_prompt= args.use_preffix_prompt ) train_dataloader = torch.utils.data.DataLoader( train_dataset, batch_size=args.train_batch_size, shuffle=True, collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation), 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 * accelerator.num_processes, num_training_steps=args.max_train_steps * accelerator.num_processes, num_cycles=args.lr_num_cycles, power=args.lr_power, ) # Prepare everything with our `accelerator`. if args.train_text_encoder: super_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( super_model, optimizer, train_dataloader, lr_scheduler ) else: super_model.unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( super_model.unet, optimizer, train_dataloader, lr_scheduler ) # 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: tracker_config = vars(copy.deepcopy(args)) tracker_config.pop("validation_images") accelerator.init_trackers("dreambooth-sdxl-single-clip", config=tracker_config) # 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(train_dataset)}") logger.info(f" Num batches each epoch = {len(train_dataloader)}") 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 mos 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, ) for epoch in range(first_epoch, args.num_train_epochs): super_model.unet.train() if args.train_text_encoder: super_model.text_encoder.train() for step, batch in enumerate(train_dataloader): with accelerator.accumulate(super_model.unet): pixel_values = batch["pixel_values"].to(dtype=weight_dtype) text_ids = batch['input_ids'] attention_mask, position_ids = batch['attention_mask'], batch['position_ids'] if vae is not None: # Convert images to latent space model_input = vae.encode(pixel_values.to(torch.float32)).latent_dist.sample() model_input = model_input.to(dtype=(weight_dtype)) model_input = model_input * vae.config.scaling_factor else: model_input = pixel_values # Sample noise that we'll add to the latents noise = torch.randn_like(model_input) noise += args.noise_offset * torch.randn( (model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device ) bsz, channels, height, width = model_input.shape # Sample a random timestep for each image timesteps = torch.randint( 0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device ) timesteps = timesteps.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) elems_to_repeat = bsz // 2 if args.with_prior_preservation else bsz output = super_model.text_encoder( input_ids=text_ids, attention_mask=attention_mask, position_ids=position_ids, output_hidden_states=True) prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone() text_proj = output.hidden_states[-1][-1, :, :].clone() encoder_hidden_states = prompt_embeds add_time_ids_cond = batch['add_time_ids'].to(model_input.device) added_cond_kwargs = {"text_embeds": text_proj, "time_ids": add_time_ids_cond} if accelerator.unwrap_model(super_model.unet).config.in_channels == channels * 2: noisy_model_input = torch.cat([noisy_model_input, noisy_model_input], dim=1) if args.class_labels_conditioning == "timesteps": class_labels = timesteps else: class_labels = None # Predict the noise residual model_pred = super_model.unet(noisy_model_input, timesteps, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs).sample # if model predicts variance, throw away the prediction. we will only train on the # simplified training objective. This means that all schedulers using the fine tuned # model must be configured to use one of the fixed variance variance types. if model_pred.shape[1] == 6: model_pred, _ = torch.chunk(model_pred, 2, dim=1) # Get the target for loss depending on the 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) else: raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") if args.with_prior_preservation: # Chunk the noise and model_pred into two parts and compute the loss on each part separately. model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0) target, target_prior = torch.chunk(target, 2, dim=0) # Compute instance loss loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") # Compute prior loss prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean") # Add the prior loss to the instance loss. loss = loss + args.prior_loss_weight * prior_loss else: loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") accelerator.backward(loss) if accelerator.sync_gradients: params_to_clip = ( itertools.chain(unet_lora_parameters, text_lora_parameters) if args.train_text_encoder else unet_lora_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: progress_bar.update(1) global_step += 1 if accelerator.distributed_type == DistributedType.DEEPSPEED or accelerator.is_main_process: if global_step % args.checkpointing_steps == 0 or global_step == 1: # _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 = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} progress_bar.set_postfix(**logs) accelerator.log(logs, step=global_step) if global_step >= args.max_train_steps: break accelerator.end_training() exit() if __name__ == "__main__": args = parse_args() main(args)