modified model card

README.md

@@ -16,71 +16,90 @@ tags:
 ![Python version](https://img.shields.io/badge/python-3.8+-blue.svg)
 ![License](https://img.shields.io/badge/license-MIT-green)
 
-This repository provides the code for a LoRa-finetuned version of the Stable Diffusion 2.1 model specifically optimized 
-for generating face images. The package includes both training and inference capabilities, along with a pretrained model 
-and the synthetic annotations used for finetuning.
-
-## Features
-- **Finetuning Script:** `finetune.py` applies LoRa adjustments to both the UNet denoiser and the text encoder of Stable Diffusion.
-- **Inference Script:** `generate.py` Ready-to-use script for generating images using the pretrained model.
-- **Pretrained Model:** `download.py` downloads our pretrained model from Hugging Face.
-
-# Environment Setup
-Set up a conda environment to run the model using the following commands:
-```bash
-conda create -n text2face
-conda activate text2face
-
-# Install requirements
-pip install -r requirements.txt
-```
+This is a LoRa-finetuned version of the Stable Diffusion 2.1 model specifically optimized 
+for generating face images. The model was trained with [FFHQ](https://github.com/NVlabs/ffhq-dataset) and [easyportrait](https://github.com/hukenovs/easyportrait) 
+using synthetic text captions for both datasets. 
+Details on the dataset format and preparation will be available soon. 
+
 
-# Checkpoints
-You can download the pretrained LoRa weights for the diffusion model and text encoder using our provided Python script `download.py`
+## Checkpoints
+You can download the pretrained LoRa weights for the diffusion model and text encoder using 
 
 ```python
 from huggingface_hub import hf_hub_download
 
-hf_hub_download(repo_id="michaeltrs/text2face", filename="checkpoints/lora30k/pytorch_lora_weights.safetensors", local_dir="./test")
+hf_hub_download(repo_id="michaeltrs/text2face",
+                    filename="checkpoints/lora30k/pytorch_lora_weights.safetensors",
+                    local_dir="checkpoints")
 ```
 
-# Inference
+## Inference
 Generate images using the `generate.py` script, which loads the SD2.1 foundation model from Hugging Face and applies the LoRa weights. 
 Generation is driven by defining a prompt and optionally a negative prompt.
+```python
+from diffusers import StableDiffusionPipeline
+import torch
+
+
+class Model:
+    def __init__(self, checkpoint="checkpoints/lora30k", weight_name="pytorch_lora_weights.safetensors", device="cuda"):
+        self.checkpoint = checkpoint
+        state_dict, network_alphas = StableDiffusionPipeline.lora_state_dict(
+            # Path to my trained lora output_dir
+            checkpoint,
+            weight_name=weight_name
+        )
+        self.pipe = StableDiffusionPipeline.from_pretrained(
+            "stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16).to(device)
+        self.pipe.load_lora_into_unet(state_dict, network_alphas, self.pipe.unet, adapter_name='test_lora')
+        self.pipe.load_lora_into_text_encoder(state_dict, network_alphas, self.pipe.text_encoder, adapter_name='test_lora')
+        self.pipe.set_adapters(["test_lora"], adapter_weights=[1.0])
+
+
+    def generate(self, prompt, negprompt='', steps=50, savedir=None, seed=1):
+        lora_scale = 1.0
+        image = self.pipe(prompt,
+                     negative_prompt=negprompt,
+                     num_inference_steps=steps,
+                     cross_attention_kwargs={"scale": lora_scale},
+                     generator=torch.manual_seed(seed)).images[0]
+        if savedir is None:
+            image.save(f"{self.checkpoint}/{'_'.join(prompt.replace('.', ' ').split(' '))}.png")
+        else:
+            image.save(f"{savedir}/{'_'.join(prompt.replace('.', ' ').split(' '))}.png")
+        return image
+
+
+if __name__ == "__main__":
 
-# Finetuning
-Use `finetune.py` to finetune a stable diffusion model using LoRAs for the UNet denoiser and the text encoder. 
-Example command for training:
-
-```bash
-accelerate config
-accelerate config default
-
-export MODEL_NAME="stabilityai/stable-diffusion-2-1"
-export TRAIN_DIR="<root dir for training data>"
-
-accelerate launch  finetune_lora.py --pretrained_model_name_or_path=$MODEL_NAME   \
---train_data_dir=$TRAIN_DIR   \
---train_text_encoder   \
---checkpointing_steps 5000   \
---resolution=768   \
---center_crop   \
---train_batch_size=4  \
---num_train_epochs 20   \
---gradient_accumulation_steps=1  \
---gradient_checkpointing  \
---num_validation_images 5  \
---learning_rate=1e-05  \
---learning_rate_text_encoder=1e-05 \
---max_grad_norm=1  \
---rank 8  \
---text_encoder_rank 8 \
---lr_scheduler="constant" \
---lr_warmup_steps=0  \
---output_dir="<output directory for trained model>"  \
---resume_from_checkpoint "latest" \
---validation_prompts "A young Latina woman, around 27 years old, with long hair and pale skin, expressing a mix of happiness and neutral emotions. She has fully open eyes and arched eyebrows." "The person is a 44-year-old Asian male with gray hair and a receding hairline. He has a big nose, closed mouth and is feeling a mix of anger and sadness." "A Latino Hispanic male, 22 years old, with straight hair, an oval face, and eyes fully open. His emotion is sad and partly neutral." "A white male, 28 years old, with a neutral emotion, sideburns, pale skin, little hair, an attractive appearance, a 5 o'clock shadow, and pointy nose." "A young, black, female individual with an oval face and big eyes, with a happy and partly surprised expression."
+    model = Model()
+
+    prompt = 'A happy 55 year old male with blond hair and a goatee smiles with visible teeth.'
+    negprompt = ''
+
+    image = model.generate(prompt, negprompt=negprompt, steps=50, seed=42)
 ```
 
-# Datasets
-Details on the dataset format and preparation will be available soon.
+## Limitations
+
+This model, Text2Face-LoRa, is finetuned from Stable Diffusion 2.1 and as such, inherits all the limitations and biases 
+associated with the base model. These biases may manifest in skewed representations across different ethnicities and 
+genders due to the nature of the training data originally used for Stable Diffusion 2.1.
+
+### Specific Limitations Include:
+
+- **Ethnic and Gender Biases**: The model may generate images that do not equally represent the diversity of human 
+features in different ethnic and gender groups, potentially reinforcing or exacerbating existing stereotypes.
+
+- **Selection Bias in Finetuning Datasets**: The datasets used for finetuning this model were selected with specific 
+criteria in mind, which may not encompass a wide enough variety of data points to correct for the inherited biases of the base model.
+
+- **Caption Generation Bias**: The synthetic annotations used to finetune this model were generated by automated 
+face analysis models, which themselves may be biased. This could lead to inaccuracies in facial feature interpretation 
+and representation, particularly for less-represented demographics in the training data.
+
+### Ethical Considerations:
+
+Users are encouraged to consider these limitations when deploying the model in real-world applications, especially 
+those involving diverse human subjects. It is advisable to perform additional validations and seek ways to mitigate 
+these biases in practical use cases.

download.py

@@ -1,7 +0,0 @@
-from huggingface_hub import hf_hub_download
-
-if __name__ == "__main__":
-    hf_hub_download(repo_id="michaeltrs/text2face",
-                    filename="checkpoints/lora30k/pytorch_lora_weights.safetensors",
-                    local_dir="checkpoints")
-

finetune_lora.py

@@ -1,975 +0,0 @@
-# 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.
-
-Modified from: https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/train_text_to_image_lora.py
-"""
-
-import argparse
-import logging
-import math
-import os
-import random
-import shutil
-from pathlib import Path
-
-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 peft import LoraConfig
-from peft.utils import get_peft_model_state_dict
-from torchvision import transforms
-from tqdm.auto import tqdm
-from transformers import CLIPTextModel, CLIPTokenizer
-
-import diffusers
-from diffusers import AutoencoderKL, DDPMScheduler, DiffusionPipeline, StableDiffusionPipeline, UNet2DConditionModel
-from diffusers.optimization import get_scheduler
-from diffusers.training_utils import compute_snr
-from diffusers.utils import check_min_version, is_wandb_available, make_image_grid
-from diffusers.utils.import_utils import is_xformers_available
-
-# import traceback
-
-
-# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
-check_min_version("0.25.0.dev0")
-
-logger = get_logger(__name__, log_level="INFO")
-
-
-# TODO: This function should be removed once training scripts are rewritten in PEFT
-def text_encoder_lora_state_dict(text_encoder):
-    state_dict = {}
-
-    def text_encoder_attn_modules(text_encoder):
-        from transformers import CLIPTextModel, CLIPTextModelWithProjection
-
-        attn_modules = []
-
-        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
-            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
-                name = f"text_model.encoder.layers.{i}.self_attn"
-                mod = layer.self_attn
-                attn_modules.append((name, mod))
-
-        return attn_modules
-
-    for name, module in text_encoder_attn_modules(text_encoder):
-        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
-
-        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
-            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
-
-    return state_dict
-
-
-def save_model_card(args, repo_id: str, images=None, base_model=str, dataset_name=str, repo_folder=None):
-    img_str = ""
-    if len(images) > 0:
-        image_grid = make_image_grid(images, 1, len(args.validation_prompts))
-        image_grid.save(os.path.join(repo_folder, "val_imgs_grid.png"))
-        img_str += "![val_imgs_grid](./val_imgs_grid.png)\n"
-
-    yaml = f"""
----
-license: creativeml-openrail-m
-base_model: {base_model}
-tags:
-- stable-diffusion
-- stable-diffusion-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}
-"""
-    with open(os.path.join(repo_folder, "README.md"), "w") as f:
-        f.write(yaml + model_card)
-
-
-def parse_args():
-    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(
-        "--revision",
-        type=str,
-        default=None,
-        required=False,
-        help="Revision of pretrained model identifier from huggingface.co/models.",
-    )
-    parser.add_argument(
-        "--variant",
-        type=str,
-        default=None,
-        help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
-    )
-    parser.add_argument(
-        "--dataset_name",
-        type=str,
-        default=None,
-        help=(
-            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
-            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
-            " or to a folder containing files that �� Datasets can understand."
-        ),
-    )
-    parser.add_argument(
-        "--dataset_config_name",
-        type=str,
-        default=None,
-        help="The config of the Dataset, leave as None if there's only one config.",
-    )
-    parser.add_argument(
-        "--train_data_dir",
-        type=str,
-        default=None,
-        help=(
-            "A folder containing the training data. Folder contents must follow the structure described in"
-            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
-            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
-        ),
-    )
-    parser.add_argument(
-        "--image_column", type=str, default="image", help="The column of the dataset containing an image."
-    )
-    parser.add_argument(
-        "--caption_column",
-        type=str,
-        default="text",
-        help="The column of the dataset containing a caption or a list of captions.",
-    )
-    # parser.add_argument(
-    #     "--validation_prompt", type=str, default=None, help="A prompt that is sampled during training for inference."
-    # )
-    parser.add_argument(
-        "--validation_prompts",
-        type=str,
-        default=None,
-        nargs="+",
-        help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."),
-    )
-    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=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(
-        "--random_flip",
-        action="store_true",
-        help="whether to randomly flip images horizontally",
-    )
-    parser.add_argument(
-        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
-    )
-    parser.add_argument("--num_train_epochs", type=int, default=100)
-    parser.add_argument(
-        "--max_train_steps",
-        type=int,
-        default=None,
-        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
-    )
-    parser.add_argument(
-        "--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(
-        "--learning_rate_text_encoder",
-        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(
-        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
-    )
-    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("--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(
-        "--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(
-        "--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("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
-    parser.add_argument(
-        "--checkpointing_steps",
-        type=int,
-        default=500,
-        help=(
-            "Save a checkpoint of the training state every X updates. These checkpoints are only 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(
-        "--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."),
-    )
-    parser.add_argument(
-        "--text_encoder_rank",
-        type=int,
-        default=4,
-        help=("The dimension of the LoRA update matrices for the text encoder."),
-    )
-    parser.add_argument("--train_text_encoder",
-                        action="store_true", help="Whether to train the text encoder")
-
-    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 main():
-    args = parse_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(
-        # filename=f'{args.output_dir}/error_logs.txt',
-        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
-        datefmt="%m/%d/%Y %H:%M:%S",
-        level=logging.ERROR,
-    )
-    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 scheduler, tokenizer and models.
-    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
-    tokenizer = CLIPTokenizer.from_pretrained(
-        args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
-    )
-    text_encoder = CLIPTextModel.from_pretrained(
-        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision
-    )
-    vae = AutoencoderKL.from_pretrained(
-        args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
-    )
-    unet = UNet2DConditionModel.from_pretrained(
-        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
-    )
-    # freeze parameters of models to save more memory
-    unet.requires_grad_(False)
-    vae.requires_grad_(False)
-    text_encoder.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
-
-    # Freeze the unet parameters before adding adapters
-    for param in unet.parameters():
-        param.requires_grad_(False)
-
-    unet_lora_config = LoraConfig(
-        r=args.rank,
-        init_lora_weights="gaussian",
-        target_modules=["to_k", "to_q", "to_v", "to_out.0"]
-    )
-
-    # Freeze the text encoder parameters before adding adapters
-    if args.train_text_encoder:
-        for param in text_encoder.parameters():
-            param.requires_grad_(False)
-
-        text_encoder_lora_config = LoraConfig(
-            r=args.text_encoder_rank,
-            init_lora_weights="gaussian",
-            target_modules=["q_proj", "v_proj", "k_proj", "out_proj"]
-        )
-
-    # Move unet, vae and text_encoder to device and cast to weight_dtype
-    unet.to(accelerator.device, dtype=weight_dtype)
-    vae.to(accelerator.device, dtype=weight_dtype)
-    text_encoder.to(accelerator.device, dtype=weight_dtype)
-
-    unet.add_adapter(unet_lora_config)
-    if args.train_text_encoder:
-        text_encoder.add_adapter(text_encoder_lora_config)
-
-    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()
-            text_encoder.enable_xformers_memory_efficient_attention()
-
-        else:
-            raise ValueError("xformers is not available. Make sure it is installed correctly")
-
-    lora_layers = list(filter(lambda p: p.requires_grad, unet.parameters()))
-    lora_layers_text_encoder = list(filter(lambda p: p.requires_grad, text_encoder.parameters()))
-
-    # 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
-        )
-
-    # Initialize the optimizer
-    if args.use_8bit_adam:
-        try:
-            import bitsandbytes as bnb
-        except ImportError:
-            raise ImportError(
-                "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
-            )
-
-        optimizer_cls = bnb.optim.AdamW8bit
-    else:
-        optimizer_cls = torch.optim.AdamW
-
-    optimizer = optimizer_cls(
-        lora_layers + lora_layers_text_encoder,
-        lr=args.learning_rate,
-        betas=(args.adam_beta1, args.adam_beta2),
-        weight_decay=args.adam_weight_decay,
-        eps=args.adam_epsilon,
-    )
-
-    # Get the datasets: you can either provide your own training and evaluation files (see below)
-    # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
-
-    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
-    # download the dataset.
-    if args.dataset_name is not None:
-        # Downloading and loading a dataset from the hub.
-        dataset = load_dataset(
-            args.dataset_name,
-            args.dataset_config_name,
-            cache_dir=args.cache_dir,
-            data_dir=args.train_data_dir,
-        )
-    else:
-        data_files = {}
-        if args.train_data_dir is not None:
-            data_files["train"] = os.path.join(args.train_data_dir, "**")
-        dataset = load_dataset(
-            "imagefolder",
-            data_files=data_files,
-            cache_dir=args.cache_dir,
-        )
-        # See more about loading custom images at
-        # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder
-
-    # Preprocessing the datasets.
-    # We need to tokenize inputs and targets.
-    column_names = dataset["train"].column_names
-
-    # 6. Get the column names for input/target.
-    dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None)
-    if args.image_column is None:
-        image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
-    else:
-        image_column = args.image_column
-        if image_column not in column_names:
-            raise ValueError(
-                f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
-            )
-    if args.caption_column is None:
-        caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
-    else:
-        caption_column = args.caption_column
-        if caption_column not in column_names:
-            raise ValueError(
-                f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
-            )
-
-    # Preprocessing the datasets.
-    # 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."
-                )
-        inputs = tokenizer(
-            captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
-        )
-        return inputs.input_ids
-
-    # Preprocessing the datasets.
-    train_transforms = transforms.Compose(
-        [
-            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
-            transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
-            transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
-            transforms.ToTensor(),
-            transforms.Normalize([0.5], [0.5]),
-        ]
-    )
-
-    def preprocess_train(examples):
-        images = [image.convert("RGB") for image in examples[image_column]]
-        examples["pixel_values"] = [train_transforms(image) for image in images]
-        examples["input_ids"] = tokenize_captions(examples)
-        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()
-        input_ids = torch.stack([example["input_ids"] for example in examples])
-        return {"pixel_values": pixel_values, "input_ids": input_ids}
-
-    def save_validation_images(global_step):
-        # Run a round of inference.
-        # Create the pipeline using the trained modules and save it.
-        # accelerator.wait_for_everyone()
-        if accelerator.is_main_process:
-
-            pipeline = StableDiffusionPipeline.from_pretrained(
-                args.pretrained_model_name_or_path,
-                text_encoder=accelerator.unwrap_model(text_encoder),
-                vae=vae,  #
-                unet=accelerator.unwrap_model(unet),
-                revision=args.revision,
-                variant=args.variant,
-                torch_dtype=weight_dtype,
-
-            )
-            pipeline.save_pretrained(args.output_dir)
-
-            # Run a final round of inference.
-
-            images = []
-            for rep in range(args.num_validation_images):
-                images.append([])
-
-                logger.info("Running inference for collecting generated images...")
-                pipeline = pipeline.to(accelerator.device)
-                pipeline.torch_dtype = weight_dtype
-                pipeline.set_progress_bar_config(disable=True)
-
-                if args.enable_xformers_memory_efficient_attention:
-                    pipeline.enable_xformers_memory_efficient_attention()
-
-                if args.seed is None:
-                    generator = None
-                else:
-                    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)
-
-                for i in range(len(args.validation_prompts)):
-                    with torch.autocast("cuda"):
-                        image = pipeline(args.validation_prompts[i], num_inference_steps=20,
-                                         generator=generator).images[0]
-                    images[rep].append(image)
-
-            images = [i for im_ in images for i in im_]
-            image_grid = make_image_grid(images, args.num_validation_images, len(args.validation_prompts))
-            if global_step == 0:
-                img_savename = os.path.join(f'{args.output_dir}', f"val_imgs_grid_init.png")
-                txt_savename = os.path.join(f'{args.output_dir}', f"val_prompts_init.txt")
-            else:
-                img_savename = os.path.join(f'{args.output_dir}/checkpoint-{global_step}', f"val_imgs_grid_{global_step}.png")
-                txt_savename = os.path.join(f'{args.output_dir}/checkpoint-{global_step}', f"val_prompts_{global_step}.txt")
-
-            image_grid.save(img_savename)
-
-            with open(txt_savename, 'w') as f:
-                for line in args.validation_prompts:
-                    f.write(f"{line}\n")
-
-            del pipeline
-
-    # 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 * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
-    )
-
-    # Prepare everything with our `accelerator`.
-    unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-        unet, text_encoder, 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 = dict(vars(args))
-        tracker_config.pop("validation_prompts")
-        accelerator.init_trackers(args.output_dir, 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 Epochs = {args.num_train_epochs}")
-    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
-    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
-    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
-    logger.info(f"  Total optimization steps = {args.max_train_steps}")
-    global_step = 0
-    first_epoch = 0
-
-    # Potentially load in the weights and states from a previous save
-    if args.resume_from_checkpoint:
-        if args.resume_from_checkpoint != "latest":
-            path = os.path.basename(args.resume_from_checkpoint)
-        else:
-            # Get the most recent checkpoint
-            dirs = os.listdir(args.output_dir)
-            dirs = [d for d in dirs if d.startswith("checkpoint")]
-            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
-            path = dirs[-1] if len(dirs) > 0 else None
-
-        if path is None:
-            accelerator.print(
-                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
-            )
-            args.resume_from_checkpoint = None
-            initial_global_step = 0
-        else:
-            accelerator.print(f"Resuming from checkpoint {path}")
-            accelerator.load_state(os.path.join(args.output_dir, path))
-            global_step = int(path.split("-")[1])
-
-            initial_global_step = global_step
-            first_epoch = global_step // num_update_steps_per_epoch
-    else:
-        initial_global_step = 0
-
-    progress_bar = tqdm(
-        range(0, args.max_train_steps),
-        initial=initial_global_step,
-        desc="Steps",
-        # Only show the progress bar once on each machine.
-        disable=not accelerator.is_local_main_process,
-    )
-
-    if args.validation_prompts is not None:
-        save_validation_images(0)
-
-    for epoch in range(first_epoch, args.num_train_epochs):
-        unet.train()
-        if args.train_text_encoder:
-            text_encoder.train()
-        train_loss = 0.0
-        for step, batch in enumerate(train_dataloader):
-            with accelerator.accumulate(unet):
-                # Convert images to latent space
-                latents = vae.encode(batch["pixel_values"].to(dtype=weight_dtype)).latent_dist.sample()
-                latents = latents * vae.config.scaling_factor
-
-                # Sample noise that we'll add to the latents
-                noise = torch.randn_like(latents)
-                if args.noise_offset:
-                    # https://www.crosslabs.org//blog/diffusion-with-offset-noise
-                    noise += args.noise_offset * torch.randn(
-                        (latents.shape[0], latents.shape[1], 1, 1), device=latents.device
-                    )
-
-                bsz = latents.shape[0]
-                # Sample a random timestep for each image
-                timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
-                timesteps = timesteps.long()
-
-                # Add noise to the latents according to the noise magnitude at each timestep
-                # (this is the forward diffusion process)
-                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
-
-                # Get the text embedding for conditioning
-                encoder_hidden_states = text_encoder(batch["input_ids"])[0]
-
-                # Get the target for loss depending on the prediction type
-                if args.prediction_type is not None:
-                    # set prediction_type of scheduler if defined
-                    noise_scheduler.register_to_config(prediction_type=args.prediction_type)
-
-                if noise_scheduler.config.prediction_type == "epsilon":
-                    target = noise
-                elif noise_scheduler.config.prediction_type == "v_prediction":
-                    target = noise_scheduler.get_velocity(latents, noise, timesteps)
-                else:
-                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")
-
-                # Predict the noise residual and compute loss
-                model_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
-
-                if args.snr_gamma is None:
-                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
-                else:
-                    # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
-                    # Since we predict the noise instead of x_0, the original formulation is slightly changed.
-                    # This is discussed in Section 4.2 of the same paper.
-                    snr = compute_snr(noise_scheduler, timesteps)
-                    if noise_scheduler.config.prediction_type == "v_prediction":
-                        # Velocity objective requires that we add one to SNR values before we divide by them.
-                        snr = snr + 1
-                    mse_loss_weights = (
-                        torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
-                    )
-
-                    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:
-                    if args.train_text_encoder:
-                        params_to_clip = lora_layers + lora_layers_text_encoder
-                    else:
-                        params_to_clip = lora_layers
-                    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 global_step % args.checkpointing_steps == 0:
-                    if accelerator.is_main_process:
-                        # _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)
-
-                        unet_lora_state_dict = get_peft_model_state_dict(accelerator.unwrap_model(unet))
-                        text_encoder_lora_state_dict = get_peft_model_state_dict(accelerator.unwrap_model(text_encoder))
-
-                        StableDiffusionPipeline.save_lora_weights(
-                            save_directory=save_path,
-                            unet_lora_layers=unet_lora_state_dict,
-                            text_encoder_lora_layers=text_encoder_lora_state_dict,
-                            safe_serialization=True,
-                        )
-
-                        logger.info(f"Saved state to {save_path}")
-
-                        if args.validation_prompts is not None:
-                            save_validation_images(global_step)
-
-            logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
-            progress_bar.set_postfix(**logs)
-
-    accelerator.end_training()
-
-
-if __name__ == "__main__":
-    main()