diff --git a/docs/source/en/training/lora.md b/docs/source/en/training/lora.md index 478fa84a23..e1c93d831d 100644 --- a/docs/source/en/training/lora.md +++ b/docs/source/en/training/lora.md @@ -401,4 +401,8 @@ Thanks to [@isidentical](https://github.com/isidentical) for helping us on integ ### Known limitations specific to the Kohya-styled LoRAs -* When images don't looks similar to other UIs such ComfyUI, it can be beacause of multiple reasons as explained [here](https://github.com/huggingface/diffusers/pull/4287/#issuecomment-1655110736). \ No newline at end of file +* When images don't looks similar to other UIs such ComfyUI, it can be beacause of multiple reasons as explained [here](https://github.com/huggingface/diffusers/pull/4287/#issuecomment-1655110736). + +## Stable Diffusion XL + +We support fine-tuning of the UNet shipped in [Stable Diffusion XL](https://huggingface.co/papers/2307.01952) with DreamBooth and LoRA via the `train_text_to_image_lora_sdxl.py` script. Please refer to the docs [here](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/README_sdxl.md). diff --git a/examples/test_examples.py b/examples/test_examples.py index cc57024c35..3fc26b38a7 100644 --- a/examples/test_examples.py +++ b/examples/test_examples.py @@ -1420,3 +1420,64 @@ class ExamplesTestsAccelerate(unittest.TestCase): {x for x in os.listdir(tmpdir) if "checkpoint" in x}, {"checkpoint-6", "checkpoint-8", "checkpoint-10"}, ) + + def test_text_to_image_lora_sdxl(self): + with tempfile.TemporaryDirectory() as tmpdir: + test_args = f""" + examples/text_to_image/train_text_to_image_lora_sdxl.py + --pretrained_model_name_or_path hf-internal-testing/tiny-stable-diffusion-xl-pipe + --dataset_name hf-internal-testing/dummy_image_text_data + --resolution 64 + --train_batch_size 1 + --gradient_accumulation_steps 1 + --max_train_steps 2 + --learning_rate 5.0e-04 + --scale_lr + --lr_scheduler constant + --lr_warmup_steps 0 + --output_dir {tmpdir} + """.split() + + run_command(self._launch_args + test_args) + # save_pretrained smoke test + self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.bin"))) + + # make sure the state_dict has the correct naming in the parameters. + lora_state_dict = torch.load(os.path.join(tmpdir, "pytorch_lora_weights.bin")) + is_lora = all("lora" in k for k in lora_state_dict.keys()) + self.assertTrue(is_lora) + + def test_text_to_image_lora_sdxl_with_text_encoder(self): + with tempfile.TemporaryDirectory() as tmpdir: + test_args = f""" + examples/text_to_image/train_text_to_image_lora_sdxl.py + --pretrained_model_name_or_path hf-internal-testing/tiny-stable-diffusion-xl-pipe + --dataset_name hf-internal-testing/dummy_image_text_data + --resolution 64 + --train_batch_size 1 + --gradient_accumulation_steps 1 + --max_train_steps 2 + --learning_rate 5.0e-04 + --scale_lr + --lr_scheduler constant + --lr_warmup_steps 0 + --output_dir {tmpdir} + --train_text_encoder + """.split() + + run_command(self._launch_args + test_args) + # save_pretrained smoke test + self.assertTrue(os.path.isfile(os.path.join(tmpdir, "pytorch_lora_weights.bin"))) + + # make sure the state_dict has the correct naming in the parameters. + lora_state_dict = torch.load(os.path.join(tmpdir, "pytorch_lora_weights.bin")) + is_lora = all("lora" in k for k in lora_state_dict.keys()) + self.assertTrue(is_lora) + + # when not training the text encoder, all the parameters in the state dict should start + # with `"unet"` or `"text_encoder"` or `"text_encoder_2"` in their names. + keys = lora_state_dict.keys() + starts_with_unet = all( + k.startswith("unet") or k.startswith("text_encoder") or k.startswith("text_encoder_2") for k in keys + ) + self.assertTrue(starts_with_unet) diff --git a/examples/text_to_image/README.md b/examples/text_to_image/README.md index 62dd776170..a0bbb08275 100644 --- a/examples/text_to_image/README.md +++ b/examples/text_to_image/README.md @@ -316,3 +316,7 @@ xFormers training is not available for Flax/JAX. **Note**: According to [this issue](https://github.com/huggingface/diffusers/issues/2234#issuecomment-1416931212), xFormers `v0.0.16` cannot be used for training in some GPUs. If you observe that problem, please install a development version as indicated in that comment. + +## Stable Diffusion XL + +We support fine-tuning of the UNet and Text Encoder shipped in [Stable Diffusion XL](https://huggingface.co/papers/2307.01952) with LoRA via the `train_text_to_image_lora_xl.py` script. Please refer to the docs [here](./README_sdxl.md). diff --git a/examples/text_to_image/README_sdxl.md b/examples/text_to_image/README_sdxl.md new file mode 100644 index 0000000000..15237c9565 --- /dev/null +++ b/examples/text_to_image/README_sdxl.md @@ -0,0 +1,134 @@ +# LoRA training example for Stable Diffusion XL (SDXL) + +Low-Rank Adaption of Large Language Models was first introduced by Microsoft in [LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685) by *Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen*. + +In a nutshell, LoRA allows adapting pretrained models by adding pairs of rank-decomposition matrices to existing weights and **only** training those newly added weights. This has a couple of advantages: + +- Previous pretrained weights are kept frozen so that model is not prone to [catastrophic forgetting](https://www.pnas.org/doi/10.1073/pnas.1611835114). +- Rank-decomposition matrices have significantly fewer parameters than original model, which means that trained LoRA weights are easily portable. +- LoRA attention layers allow to control to which extent the model is adapted toward new training images via a `scale` parameter. + +[cloneofsimo](https://github.com/cloneofsimo) was the first to try out LoRA training for Stable Diffusion in the popular [lora](https://github.com/cloneofsimo/lora) GitHub repository. + +With LoRA, it's possible to fine-tune Stable Diffusion on a custom image-caption pair dataset +on consumer GPUs like Tesla T4, Tesla V100. + +## Running locally with PyTorch + +### Installing the dependencies + +Before running the scripts, make sure to install the library's training dependencies: + +**Important** + +To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment: + +```bash +git clone https://github.com/huggingface/diffusers +cd diffusers +pip install -e . +``` + +Then cd in the `examples/text_to_image` folder and run +```bash +pip install -r requirements_sdxl.txt +``` + +And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with: + +```bash +accelerate config +``` + +Or for a default accelerate configuration without answering questions about your environment + +```bash +accelerate config default +``` + +Or if your environment doesn't support an interactive shell (e.g., a notebook) + +```python +from accelerate.utils import write_basic_config +write_basic_config() +``` + +When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups. + +### Training + +First, you need to set up your development environment as is explained in the [installation section](#installing-the-dependencies). Make sure to set the `MODEL_NAME` and `DATASET_NAME` environment variables. Here, we will use [Stable Diffusion XL 1.0-base](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) and the [Pokemons dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions). + +**___Note: It is quite useful to monitor the training progress by regularly generating sample images during training. [Weights and Biases](https://docs.wandb.ai/quickstart) is a nice solution to easily see generating images during training. All you need to do is to run `pip install wandb` before training to automatically log images.___** + +```bash +export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0" +export DATASET_NAME="lambdalabs/pokemon-blip-captions" +``` + +For this example we want to directly store the trained LoRA embeddings on the Hub, so +we need to be logged in and add the `--push_to_hub` flag. + +```bash +huggingface-cli login +``` + +Now we can start training! + +```bash +accelerate launch train_text_to_image_lora_sdxl.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --dataset_name=$DATASET_NAME --caption_column="text" \ + --resolution=1024 --random_flip \ + --train_batch_size=1 \ + --num_train_epochs=2 --checkpointing_steps=500 \ + --learning_rate=1e-04 --lr_scheduler="constant" --lr_warmup_steps=0 \ + --seed=42 \ + --output_dir="sd-pokemon-model-lora-sdxl" \ + --validation_prompt="cute dragon creature" --report_to="wandb" \ + --push_to_hub +``` + +The above command will also run inference as fine-tuning progresses and log the results to Weights and Biases. + +### Finetuning the text encoder and UNet + +The script also allows you to finetune the `text_encoder` along with the `unet`. + +🚨 Training the text encoder requires additional memory. + +Pass the `--train_text_encoder` argument to the training script to enable finetuning the `text_encoder` and `unet`: + +```bash +accelerate launch train_text_to_image_lora_sdxl.py \ + --pretrained_model_name_or_path=$MODEL_NAME \ + --dataset_name=$DATASET_NAME --caption_column="text" \ + --resolution=1024 --random_flip \ + --train_batch_size=1 \ + --num_train_epochs=2 --checkpointing_steps=500 \ + --learning_rate=1e-04 --lr_scheduler="constant" --lr_warmup_steps=0 \ + --seed=42 \ + --output_dir="sd-pokemon-model-lora-sdxl-txt" \ + --train_text_encoder \ + --validation_prompt="cute dragon creature" --report_to="wandb" \ + --push_to_hub +``` + +### Inference + +Once you have trained a model using above command, the inference can be done simply using the `DiffusionPipeline` after loading the trained LoRA weights. You +need to pass the `output_dir` for loading the LoRA weights which, in this case, is `sd-pokemon-model-lora-sdxl`. + +```python +from diffusers import DiffusionPipeline +import torch + +model_path = "takuoko/sd-pokemon-model-lora-sdxl" +pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16) +pipe.to("cuda") +pipe.load_lora_weights(model_path) + +prompt = "A pokemon with green eyes and red legs." +image = pipe(prompt, num_inference_steps=30, guidance_scale=7.5).images[0] +image.save("pokemon.png") +``` diff --git a/examples/text_to_image/requirements_sdxl.txt b/examples/text_to_image/requirements_sdxl.txt new file mode 100644 index 0000000000..7a612982f4 --- /dev/null +++ b/examples/text_to_image/requirements_sdxl.txt @@ -0,0 +1,6 @@ +accelerate>=0.16.0 +torchvision +transformers>=4.25.1 +ftfy +tensorboard +Jinja2 diff --git a/examples/text_to_image/train_text_to_image_lora_sdxl.py b/examples/text_to_image/train_text_to_image_lora_sdxl.py new file mode 100644 index 0000000000..b5e7ee3e7d --- /dev/null +++ b/examples/text_to_image/train_text_to_image_lora_sdxl.py @@ -0,0 +1,1282 @@ +#!/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 +# limitations under the License. +"""Fine-tuning script for Stable Diffusion for text2image with support for LoRA.""" + +import argparse +import itertools +import logging +import math +import os +import random +import shutil +from pathlib import Path +from typing import Dict + +import datasets +import numpy as np +import torch +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 ProjectConfiguration, set_seed +from datasets import load_dataset +from huggingface_hub import create_repo, upload_folder +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.loaders import LoraLoaderMixin, text_encoder_lora_state_dict +from diffusers.models.attention_processor import LoRAAttnProcessor, LoRAAttnProcessor2_0 +from diffusers.optimization import get_scheduler +from diffusers.utils import check_min_version, is_wandb_available +from diffusers.utils.import_utils import is_xformers_available + + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.19.0.dev0") + +logger = get_logger(__name__) + + +def save_model_card( + repo_id: str, + images=None, + base_model=str, + dataset_name=str, + train_text_encoder=False, + repo_folder=None, + vae_path=None, +): + img_str = "" + 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" + + yaml = f""" +--- +license: creativeml-openrail-m +base_model: {base_model} +dataset: {dataset_name} +tags: +- stable-diffusion-xl +- stable-diffusion-xl-diffusers +- text-to-image +- diffusers +- lora +inference: true +--- + """ + model_card = f""" +# LoRA text2image fine-tuning - {repo_id} + +These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n +{img_str} + +LoRA for the text encoder was enabled: {train_text_encoder}. + +Special VAE used for training: {vae_path}. +""" + with open(os.path.join(repo_folder, "README.md"), "w") as f: + f.write(yaml + model_card) + + +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="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( + "--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( + "--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( + "--output_dir", + type=str, + default="sd-model-finetuned-lora", + 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_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=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( + "--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( + "--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.prediciton_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( + "--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("--noise_offset", type=float, default=0, help="The scale of noise offset.") + parser.add_argument( + "--rank", + type=int, + default=4, + help=("The dimension of the LoRA update matrices."), + ) + + 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.") + + return args + + +DATASET_NAME_MAPPING = { + "lambdalabs/pokemon-blip-captions": ("image", "text"), +} + + +def unet_attn_processors_state_dict(unet) -> Dict[str, torch.tensor]: + """ + Returns: + a state dict containing just the attention processor parameters. + """ + attn_processors = unet.attn_processors + + attn_processors_state_dict = {} + + for attn_processor_key, attn_processor in attn_processors.items(): + for parameter_key, parameter in attn_processor.state_dict().items(): + attn_processors_state_dict[f"{attn_processor_key}.{parameter_key}"] = parameter + + return attn_processors_state_dict + + +def tokenize_prompt(tokenizer, prompt): + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + return text_input_ids + + +# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt +def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None): + prompt_embeds_list = [] + + for i, text_encoder in enumerate(text_encoders): + if tokenizers is not None: + tokenizer = tokenizers[i] + text_input_ids = tokenize_prompt(tokenizer, prompt) + else: + assert text_input_ids_list is not None + text_input_ids = text_input_ids_list[i] + + prompt_embeds = text_encoder( + text_input_ids.to(text_encoder.device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + prompt_embeds = prompt_embeds.hidden_states[-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, pooled_prompt_embeds + + +def main(args): + logging_dir = Path(args.output_dir, args.logging_dir) + + 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 + + # 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") + text_encoder_one = text_encoder_cls_one.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision + ) + text_encoder_two = text_encoder_cls_two.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision + ) + 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 + ) + unet = UNet2DConditionModel.from_pretrained( + args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision + ) + + # We only train the additional adapter LoRA layers + vae.requires_grad_(False) + text_encoder_one.requires_grad_(False) + text_encoder_two.requires_grad_(False) + unet.requires_grad_(False) + + # For mixed precision training we cast all non-trainable weigths (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 + # The VAE is in float32 to avoid NaN losses. + unet.to(accelerator.device, dtype=weight_dtype) + if args.pretrained_vae_model_name_or_path is None: + vae.to(accelerator.device, dtype=torch.float32) + else: + vae.to(accelerator.device, dtype=weight_dtype) + text_encoder_one.to(accelerator.device, dtype=weight_dtype) + text_encoder_two.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") + + # now we will add new LoRA weights to the attention layers + # Set correct lora layers + unet_lora_attn_procs = {} + unet_lora_parameters = [] + for name, attn_processor in unet.attn_processors.items(): + cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim + if name.startswith("mid_block"): + hidden_size = unet.config.block_out_channels[-1] + elif name.startswith("up_blocks"): + block_id = int(name[len("up_blocks.")]) + hidden_size = list(reversed(unet.config.block_out_channels))[block_id] + elif name.startswith("down_blocks"): + block_id = int(name[len("down_blocks.")]) + hidden_size = unet.config.block_out_channels[block_id] + + lora_attn_processor_class = ( + LoRAAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else LoRAAttnProcessor + ) + module = lora_attn_processor_class( + hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, rank=args.rank + ) + unet_lora_attn_procs[name] = module + unet_lora_parameters.extend(module.parameters()) + + unet.set_attn_processor(unet_lora_attn_procs) + + def compute_snr(timesteps): + """ + Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849 + """ + alphas_cumprod = noise_scheduler.alphas_cumprod + sqrt_alphas_cumprod = alphas_cumprod**0.5 + sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5 + + # Expand the tensors. + # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026 + sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None] + alpha = sqrt_alphas_cumprod.expand(timesteps.shape) + + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None] + sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape) + + # Compute SNR. + snr = (alpha / sigma) ** 2 + return snr + + # The text encoder comes from 🤗 transformers, so we cannot directly modify it. + # So, instead, we monkey-patch the forward calls of its attention-blocks. + if args.train_text_encoder: + # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16 + text_lora_parameters_one = LoraLoaderMixin._modify_text_encoder( + text_encoder_one, dtype=torch.float32, rank=args.rank + ) + text_lora_parameters_two = LoraLoaderMixin._modify_text_encoder( + text_encoder_two, dtype=torch.float32, rank=args.rank + ) + + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + def save_model_hook(models, weights, output_dir): + # there are only two options here. Either are just the unet attn processor layers + # or there are the unet and text encoder atten layers + unet_lora_layers_to_save = None + text_encoder_one_lora_layers_to_save = None + text_encoder_two_lora_layers_to_save = None + + for model in models: + if isinstance(model, type(accelerator.unwrap_model(unet))): + unet_lora_layers_to_save = unet_attn_processors_state_dict(model) + elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))): + text_encoder_one_lora_layers_to_save = text_encoder_lora_state_dict(model) + elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))): + text_encoder_two_lora_layers_to_save = text_encoder_lora_state_dict(model) + else: + raise ValueError(f"unexpected save model: {model.__class__}") + + # make sure to pop weight so that corresponding model is not saved again + weights.pop() + + StableDiffusionXLPipeline.save_lora_weights( + output_dir, + unet_lora_layers=unet_lora_layers_to_save, + text_encoder_lora_layers=text_encoder_one_lora_layers_to_save, + text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save, + ) + + def load_model_hook(models, input_dir): + unet_ = None + text_encoder_one_ = None + text_encoder_two_ = None + + while len(models) > 0: + model = models.pop() + + if isinstance(model, type(accelerator.unwrap_model(unet))): + unet_ = model + elif isinstance(model, type(accelerator.unwrap_model(text_encoder_one))): + text_encoder_one_ = model + elif isinstance(model, type(accelerator.unwrap_model(text_encoder_two))): + text_encoder_two_ = model + else: + raise ValueError(f"unexpected save model: {model.__class__}") + + lora_state_dict, network_alpha = LoraLoaderMixin.lora_state_dict(input_dir) + LoraLoaderMixin.load_lora_into_unet(lora_state_dict, network_alpha=network_alpha, unet=unet_) + LoraLoaderMixin.load_lora_into_text_encoder( + lora_state_dict, network_alpha=network_alpha, text_encoder=text_encoder_one_ + ) + LoraLoaderMixin.load_lora_into_text_encoder( + lora_state_dict, network_alpha=network_alpha, text_encoder=text_encoder_two_ + ) + + accelerator.register_save_state_pre_hook(save_model_hook) + 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 + + # Optimizer creation + params_to_optimize = ( + itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two) + if args.train_text_encoder + else unet_lora_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, + ) + 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. + # We need to tokenize input captions and transform the images. + def tokenize_captions(examples, is_train=True): + captions = [] + for caption in examples[caption_column]: + if 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]) + else: + raise ValueError( + f"Caption column `{caption_column}` should contain either strings or lists of strings." + ) + tokens_one = tokenize_prompt(tokenizer_one, captions) + tokens_two = tokenize_prompt(tokenizer_two, captions) + return tokens_one, tokens_two + + # 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.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) + if args.random_flip and random.random() < 0.5: + # flip + x1 = image.width - x1 + image = train_flip(image) + 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 + tokens_one, tokens_two = tokenize_captions(examples) + examples["input_ids_one"] = tokens_one + examples["input_ids_two"] = tokens_two + 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) + + def collate_fn(examples): + pixel_values = torch.stack([example["pixel_values"] for example in examples]) + pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float() + original_sizes = [example["original_sizes"] for example in examples] + crop_top_lefts = [example["crop_top_lefts"] for example in examples] + input_ids_one = torch.stack([example["input_ids_one"] for example in examples]) + input_ids_two = torch.stack([example["input_ids_two"] for example in examples]) + return { + "pixel_values": pixel_values, + "input_ids_one": input_ids_one, + "input_ids_two": input_ids_two, + "original_sizes": original_sizes, + "crop_top_lefts": crop_top_lefts, + } + + # DataLoaders creation: + train_dataloader = torch.utils.data.DataLoader( + train_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`. + if args.train_text_encoder: + unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet, text_encoder_one, text_encoder_two, optimizer, train_dataloader, lr_scheduler + ) + else: + unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + 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: + accelerator.init_trackers("text2image-fine-tune", config=vars(args)) + + # 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 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 + else: + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + global_step = int(path.split("-")[1]) + + resume_global_step = global_step * args.gradient_accumulation_steps + first_epoch = global_step // num_update_steps_per_epoch + resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps) + + # Only show the progress bar once on each machine. + progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process) + progress_bar.set_description("Steps") + + for epoch in range(first_epoch, args.num_train_epochs): + unet.train() + if args.train_text_encoder: + text_encoder_one.train() + text_encoder_two.train() + train_loss = 0.0 + for step, batch in enumerate(train_dataloader): + # Skip steps until we reach the resumed step + if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step: + if step % args.gradient_accumulation_steps == 0: + progress_bar.update(1) + continue + + with accelerator.accumulate(unet): + pixel_values = batch["pixel_values"].to(dtype=weight_dtype) + + # Convert images to latent space + model_input = vae.encode(pixel_values).latent_dist.sample() + model_input = model_input * vae.config.scaling_factor + if args.pretrained_vae_model_name_or_path is None: + model_input = model_input.to(weight_dtype) + + # Sample noise that we'll add to the latents + 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] + # 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) + + # 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]) + add_time_ids = add_time_ids.to(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, pooled_prompt_embeds = encode_prompt( + text_encoders=[text_encoder_one, text_encoder_two], + tokenizers=None, + prompt=None, + text_input_ids_list=[batch["input_ids_one"], batch["input_ids_two"]], + ) + unet_added_conditions.update({"text_embeds": pooled_prompt_embeds}) + prompt_embeds = prompt_embeds + model_pred = unet( + noisy_model_input, timesteps, prompt_embeds, added_cond_kwargs=unet_added_conditions + ).sample + + # 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) + 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(timesteps) + mse_loss_weights = ( + torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr + ) + # We first calculate the original loss. Then we mean over the non-batch dimensions and + # rebalance the sample-wise losses with their respective loss weights. + # Finally, we take the mean of the rebalanced loss. + 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() + + # Gather the losses across all processes for logging (if we use distributed training). + avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean() + train_loss += avg_loss.item() / args.gradient_accumulation_steps + + # Backpropagate + accelerator.backward(loss) + if accelerator.sync_gradients: + params_to_clip = ( + itertools.chain(unet_lora_parameters, text_lora_parameters_one, text_lora_parameters_two) + 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 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + + if 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}." + ) + # create pipeline + if not args.train_text_encoder: + text_encoder_one = text_encoder_cls_one.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision + ) + text_encoder_two = text_encoder_cls_two.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision + ) + pipeline = StableDiffusionXLPipeline.from_pretrained( + args.pretrained_model_name_or_path, + vae=vae, + text_encoder=accelerator.unwrap_model(text_encoder_one), + text_encoder_2=accelerator.unwrap_model(text_encoder_two), + unet=accelerator.unwrap_model(unet), + revision=args.revision, + torch_dtype=weight_dtype, + ) + + 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 else None + pipeline_args = {"prompt": args.validation_prompt} + + with torch.cuda.amp.autocast(): + images = [ + pipeline(**pipeline_args, 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("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 + torch.cuda.empty_cache() + + # Save the lora layers + accelerator.wait_for_everyone() + if accelerator.is_main_process: + unet = accelerator.unwrap_model(unet) + unet = unet.to(torch.float32) + unet_lora_layers = unet_attn_processors_state_dict(unet) + + if args.train_text_encoder: + text_encoder_one = accelerator.unwrap_model(text_encoder_one) + text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder_one.to(torch.float32)) + text_encoder_two = accelerator.unwrap_model(text_encoder_two) + text_encoder_2_lora_layers = text_encoder_lora_state_dict(text_encoder_two.to(torch.float32)) + else: + text_encoder_lora_layers = None + text_encoder_2_lora_layers = None + + StableDiffusionXLPipeline.save_lora_weights( + save_directory=args.output_dir, + unet_lora_layers=unet_lora_layers, + text_encoder_lora_layers=text_encoder_lora_layers, + text_encoder_2_lora_layers=text_encoder_2_lora_layers, + ) + + # Final inference + # Load previous pipeline + vae = AutoencoderKL.from_pretrained( + vae_path, + subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None, + revision=args.revision, + torch_dtype=weight_dtype, + ) + pipeline = StableDiffusionXLPipeline.from_pretrained( + args.pretrained_model_name_or_path, vae=vae, revision=args.revision, torch_dtype=weight_dtype + ) + pipeline = pipeline.to(accelerator.device) + + # load attention processors + pipeline.load_lora_weights(args.output_dir) + + # run inference + images = [] + if args.validation_prompt and args.num_validation_images > 0: + generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None + 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, + images=images, + base_model=args.pretrained_model_name_or_path, + train_text_encoder=args.train_text_encoder, + prompt=args.instance_prompt, + 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)