From d488567a0321d271ed29801766cf6a7bd401d223 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=9D=8E=E4=B8=80=E5=8A=9B?= Date: Thu, 18 Jul 2024 10:15:41 +0800 Subject: [PATCH 1/7] add sd2.1, sd1.5, sdxl MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit link #IA9UH2 add sd2.1, sd1.5, sdxl based diffusers Signed-off-by: 李一力 --- .../stable-diffusion/diffusers/README.md | 64 + .../diffusers/configs/default_config.yaml | 20 + .../diffusers/configs/single_config.yaml | 20 + .../diffusers/configs/zero2_config.yaml | 23 + .../diffusers/requirements.txt | 5 + .../diffusers/run_sd_1.5_multi.sh | 31 + .../diffusers/run_sd_1.5_single.sh | 30 + .../diffusers/run_sd_2.1_multi.sh | 34 + .../diffusers/run_sd_2.1_single.sh | 30 + .../stable-diffusion/diffusers/run_sd_xl.sh | 35 + .../diffusers/train_text_to_image.py | 1129 ++++++++++++++ .../diffusers/train_text_to_image_sdxl.py | 1371 +++++++++++++++++ 12 files changed, 2792 insertions(+) create mode 100644 multimodal/diffusion/stable-diffusion/diffusers/README.md create mode 100644 multimodal/diffusion/stable-diffusion/diffusers/configs/default_config.yaml create mode 100644 multimodal/diffusion/stable-diffusion/diffusers/configs/single_config.yaml create mode 100644 multimodal/diffusion/stable-diffusion/diffusers/configs/zero2_config.yaml create mode 100644 multimodal/diffusion/stable-diffusion/diffusers/requirements.txt create mode 100755 multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh create mode 100755 multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh create mode 100755 multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh create mode 100755 multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh create mode 100755 multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh create mode 100644 multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py create mode 100755 multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py diff --git a/multimodal/diffusion/stable-diffusion/diffusers/README.md b/multimodal/diffusion/stable-diffusion/diffusers/README.md new file mode 100644 index 000000000..c51b106f8 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/README.md @@ -0,0 +1,64 @@ +table Diffusion + +## Model description + +Stable Diffusion is a latent text-to-image diffusion model. + +## Step 1: Installation + +- Install + +```bash +pip3 install -r requirements.txt +pip3 install pillow --upgrade +``` + + +## Step 2: Preparing datasets +You just need to run the script, and it will automatically download the required data and weights. Or you can manually download the weights and data locally. + +dataset: download the lambdalabs/pokemon-blip-captions from [huggingface page](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions). + +weights: download the stabilityai/stable-diffusion-2-1-base from [huggingface page](https://huggingface.co/stabilityai/stable-diffusion-2-1-base). + +weights: download the runwayml/stable-diffusion-v1-5 from [huggingface page](https://huggingface.co/runwayml/stable-diffusion-v1-5). + + +weights: download the stabilityai/stable-diffusion-xl-base-1.0 from [huggingface page](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0). + +weights: download the madebyollin/sdxl-vae-fp16-fix from [huggingface page](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix). + +## Step 3: Training + +If you have downloaded the weights and data, please import the environment variables like below. +```bash +export MODEL_PATH=/path/to/sd_weights +export DATA_PATH=/path/to/data +export VAE_PATH=/path/to/vae_weights # only for sdxl +``` +### sd2.1 or sd1.5 +#### One GPU +```bash +bash run_sd_single.sh +``` +#### 8 GPUs +``` +bash run_sd_multi.sh +``` +### sdxl +#### 8 GPUs +```bash +bash run_sd_xl.sh +``` + +## Results +### sd2.1 + +GPUs | FPS +---- | --- +BI-V100 x8 | ips per gpu=16 +``` +## Reference + +- [diffusers](https://github.com/huggingface/diffusers) + diff --git a/multimodal/diffusion/stable-diffusion/diffusers/configs/default_config.yaml b/multimodal/diffusion/stable-diffusion/diffusers/configs/default_config.yaml new file mode 100644 index 000000000..829e0b662 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/configs/default_config.yaml @@ -0,0 +1,20 @@ +compute_environment: LOCAL_MACHINE +debug: false +deepspeed_config: + gradient_accumulation_steps: 1 + steps_per_print: 1 + zero3_init_flag: true + zero_stage: 0 +distributed_type: DEEPSPEED +downcast_bf16: 'no' +machine_rank: 0 +main_training_function: main +mixed_precision: fp16 +num_machines: 1 +num_processes: 16 +rdzv_backend: static +same_network: true +tpu_env: [] +tpu_use_cluster: false +tpu_use_sudo: false +use_cpu: false diff --git a/multimodal/diffusion/stable-diffusion/diffusers/configs/single_config.yaml b/multimodal/diffusion/stable-diffusion/diffusers/configs/single_config.yaml new file mode 100644 index 000000000..a20a8a46b --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/configs/single_config.yaml @@ -0,0 +1,20 @@ +compute_environment: LOCAL_MACHINE +debug: false +deepspeed_config: + gradient_accumulation_steps: 1 + steps_per_print: 1 + zero3_init_flag: true + zero_stage: 0 +distributed_type: DEEPSPEED +downcast_bf16: 'no' +machine_rank: 0 +main_training_function: main +mixed_precision: fp16 +num_machines: 1 +num_processes: 1 +rdzv_backend: static +same_network: true +tpu_env: [] +tpu_use_cluster: false +tpu_use_sudo: false +use_cpu: false diff --git a/multimodal/diffusion/stable-diffusion/diffusers/configs/zero2_config.yaml b/multimodal/diffusion/stable-diffusion/diffusers/configs/zero2_config.yaml new file mode 100644 index 000000000..8ffdbbd1e --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/configs/zero2_config.yaml @@ -0,0 +1,23 @@ +compute_environment: LOCAL_MACHINE +debug: false +deepspeed_config: + gradient_accumulation_steps: 1 + gradient_clipping: 1.0 + offload_optimizer_device: none + offload_param_device: none + zero3_init_flag: false + zero_stage: 2 +distributed_type: DEEPSPEED +downcast_bf16: 'no' +enable_cpu_affinity: false +machine_rank: 0 +main_training_function: main +mixed_precision: fp16 +num_machines: 1 +num_processes: 16 +rdzv_backend: static +same_network: true +tpu_env: [] +tpu_use_cluster: false +tpu_use_sudo: false +use_cpu: false diff --git a/multimodal/diffusion/stable-diffusion/diffusers/requirements.txt b/multimodal/diffusion/stable-diffusion/diffusers/requirements.txt new file mode 100644 index 000000000..7b23b4ce9 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/requirements.txt @@ -0,0 +1,5 @@ +accelerate +datasets +ftfy +Jinja2 +peft==0.7.0 diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh new file mode 100755 index 000000000..d95cc8e60 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh @@ -0,0 +1,31 @@ +export CLIP_FLASH_ATTN=1 +export USE_NHWC_GN=1 +export USE_IXFORMER_GEGLU=1 +export USE_APEX_LN=1 +export ENABLE_FLASH_ATTENTION_WITH_IXDNN=1 +echo $ENABLE_FLASH_ATTENTION_WITH_IXDNN + +export MODEL_PATH=${MODEL_PATH:-runwayml/stable-diffusion-v1-5} +export DATASET_PATH=${DATASET_PATH:-lambdalabs/pokemon-blip-captions} + + +accelerate launch --config_file configs/default_config.yaml --mixed_precision="fp16" train_text_to_image.py \ + --pretrained_model_name_or_path=$MODEL_PATH \ + --dataset_name=$DATASET_PATH \ + --resolution=512 \ + --seed 42 \ + --center_crop \ + --random_flip \ + --train_batch_size=32 \ + --gradient_accumulation_steps=1 \ + --gradient_checkpointing \ + --learning_rate=1e-05 \ + --max_grad_norm=1 \ + --lr_scheduler="constant" \ + --lr_warmup_steps=0 \ + --output_dir="sd-pokemon-model-3" \ + --max_train_steps=100 \ + --NHWC \ + --dataloader_num_workers=32 \ + --apex_fused_adam + # --use_ema diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh new file mode 100755 index 000000000..f21d71687 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh @@ -0,0 +1,30 @@ +export CLIP_FLASH_ATTN=1 +export USE_NHWC_GN=1 +export USE_IXFORMER_GEGLU=1 +export USE_APEX_LN=1 +export ENABLE_FLASH_ATTENTION_WITH_IXDNN=1 + +export MODEL_PATH=${MODEL_PATH:-runwayml/stable-diffusion-v1-5} +export DATASET_PATH=${DATASET_PATH:-lambdalabs/pokemon-blip-captions} + + +accelerate launch --config_file configs/single_config.yaml --mixed_precision="fp16" train_text_to_image.py \ + --pretrained_model_name_or_path=$MODEL_PATH \ + --dataset_name=$DATASET_PATH \ + --resolution=512 \ + --seed 42 \ + --center_crop \ + --random_flip \ + --train_batch_size=32 \ + --gradient_accumulation_steps=1 \ + --gradient_checkpointing \ + --learning_rate=1e-05 \ + --max_grad_norm=1 \ + --lr_scheduler="constant" \ + --lr_warmup_steps=0 \ + --output_dir="sd-pokemon-model-3" \ + --max_train_steps=100 \ + --NHWC \ + --dataloader_num_workers=32 \ + --apex_fused_adam + # --use_ema diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh new file mode 100755 index 000000000..3ec9edbc7 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh @@ -0,0 +1,34 @@ +export CLIP_FLASH_ATTN=1 +export USE_NHWC_GN=1 +export USE_IXFORMER_GEGLU=1 +export USE_APEX_LN=1 +export ENABLE_FLASH_ATTENTION_WITH_IXDNN=1 +echo $ENABLE_FLASH_ATTENTION_WITH_IXDNN + +export MODEL_PATH=${MODEL_PATH:-stabilityai/stable-diffusion-2-1-base} +export DATASET_PATH=${DATASET_PATH:-lambdalabs/pokemon-blip-captions} +echo $MODEL_PATH +echo $DATASET_PATH + +# cd /data/yili.li/Diffusers/diffusers/examples/text_to_image + +accelerate launch --config_file configs/default_config.yaml --mixed_precision="fp16" train_text_to_image.py \ + --pretrained_model_name_or_path=$MODEL_PATH \ + --dataset_name=$DATASET_PATH \ + --resolution=512 \ + --seed 42 \ + --center_crop \ + --random_flip \ + --train_batch_size=32 \ + --gradient_accumulation_steps=1 \ + --gradient_checkpointing \ + --learning_rate=1e-05 \ + --max_grad_norm=1 \ + --lr_scheduler="constant" \ + --lr_warmup_steps=0 \ + --output_dir="sd-pokemon-model-3" \ + --max_train_steps=100 \ + --NHWC \ + --dataloader_num_workers=32 \ + --apex_fused_adam + # --use_ema diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh new file mode 100755 index 000000000..a865a5f05 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh @@ -0,0 +1,30 @@ +export CLIP_FLASH_ATTN=1 +export USE_NHWC_GN=1 +export USE_IXFORMER_GEGLU=1 +export USE_APEX_LN=1 +export ENABLE_FLASH_ATTENTION_WITH_IXDNN=1 + +export MODEL_PATH=${MODEL_PATH:-stabilityai/stable-diffusion-2-1-base} +export DATASET_PATH=${DATASET_PATH:-lambdalabs/pokemon-blip-captions} + + +accelerate launch --config_file configs/single_config.yaml --mixed_precision="fp16" train_text_to_image.py \ + --pretrained_model_name_or_path=$MODEL_PATH \ + --dataset_name=$DATASET_PATH \ + --resolution=512 \ + --seed 42 \ + --center_crop \ + --random_flip \ + --train_batch_size=32 \ + --gradient_accumulation_steps=1 \ + --gradient_checkpointing \ + --learning_rate=1e-05 \ + --max_grad_norm=1 \ + --lr_scheduler="constant" \ + --lr_warmup_steps=0 \ + --output_dir="sd-pokemon-model-3" \ + --max_train_steps=100 \ + --NHWC \ + --dataloader_num_workers=32 \ + --apex_fused_adam + # --use_ema diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh new file mode 100755 index 000000000..6d532eb59 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh @@ -0,0 +1,35 @@ +export CLIP_FLASH_ATTN=1 +export USE_NHWC_GN=1 +export USE_IXFORMER_GEGLU=1 +export USE_APEX_LN=1 +export USE_NATIVE_ATTN=0 +export ENABLE_FLASH_ATTENTION_WITH_IXDNN=1 + +export MODEL_PATH=${MODEL_PATH:-stabilityai/stable-diffusion-xl-base-1.0} +export DATASET_PATH=${DATASET_PATH:-lambdalabs/pokemon-blip-captions} +export VAE_PATH=${VAE_PATH:-madebyollin/sdxl-vae-fp16-fix} + + +accelerate launch --config_file configs/zero2_config.yaml --mixed_precision="fp16" train_text_to_image_sdxl.py \ + --pretrained_model_name_or_path=$MODEL_PATH \ + --pretrained_vae_model_name_or_path=$VAE_PATH \ + --dataset_name=$DATASET_PATH \ + --resolution=512 \ + --seed 42 \ + --gradient_checkpointing \ + --center_crop \ + --random_flip \ + --train_batch_size=32 \ + --gradient_accumulation_steps=1 \ + --gradient_checkpointing \ + --learning_rate=1e-05 \ + --max_grad_norm=1 \ + --lr_scheduler="constant" \ + --lr_warmup_steps=0 \ + --output_dir="sd-pokemon-model-3" \ + --max_train_steps=100 \ + --dataloader_num_workers=32 \ + --NHWC \ + --apex_fused_adam + # --use_ema + diff --git a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py new file mode 100644 index 000000000..ada797894 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py @@ -0,0 +1,1129 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 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. + +import argparse +import logging +import math +import os +import random +import shutil +from pathlib import Path + +import accelerate +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.state import AcceleratorState +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 tqdm.auto import tqdm +from transformers import CLIPTextModel, CLIPTokenizer +from transformers.utils import ContextManagers + +import diffusers +from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel +from diffusers.optimization import get_scheduler +from diffusers.training_utils import EMAModel, compute_snr +from diffusers.utils import check_min_version, deprecate, is_wandb_available, make_image_grid +from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card +from diffusers.utils.import_utils import is_xformers_available +from diffusers.utils.torch_utils import is_compiled_module + + +if is_wandb_available(): + import wandb + + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.27.0") + +logger = get_logger(__name__, log_level="INFO") + +DATASET_NAME_MAPPING = { + "lambdalabs/pokemon-blip-captions": ("image", "text"), +} + + +def save_model_card( + args, + repo_id: str, + images: list = None, + repo_folder: str = 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" + + model_description = f""" +# Text-to-image finetuning - {repo_id} + +This pipeline was finetuned from **{args.pretrained_model_name_or_path}** on the **{args.dataset_name}** dataset. Below are some example images generated with the finetuned pipeline using the following prompts: {args.validation_prompts}: \n +{img_str} + +## Pipeline usage + +You can use the pipeline like so: + +```python +from diffusers import DiffusionPipeline +import torch + +pipeline = DiffusionPipeline.from_pretrained("{repo_id}", torch_dtype=torch.float16) +prompt = "{args.validation_prompts[0]}" +image = pipeline(prompt).images[0] +image.save("my_image.png") +``` + +## Training info + +These are the key hyperparameters used during training: + +* Epochs: {args.num_train_epochs} +* Learning rate: {args.learning_rate} +* Batch size: {args.train_batch_size} +* Gradient accumulation steps: {args.gradient_accumulation_steps} +* Image resolution: {args.resolution} +* Mixed-precision: {args.mixed_precision} + +""" + wandb_info = "" + if is_wandb_available(): + wandb_run_url = None + if wandb.run is not None: + wandb_run_url = wandb.run.url + + if wandb_run_url is not None: + wandb_info = f""" +More information on all the CLI arguments and the environment are available on your [`wandb` run page]({wandb_run_url}). +""" + + model_description += wandb_info + + model_card = load_or_create_model_card( + repo_id_or_path=repo_id, + from_training=True, + license="creativeml-openrail-m", + base_model=args.pretrained_model_name_or_path, + model_description=model_description, + inference=True, + ) + + tags = ["stable-diffusion", "stable-diffusion-diffusers", "text-to-image", "diffusers", "diffusers-training"] + model_card = populate_model_card(model_card, tags=tags) + + model_card.save(os.path.join(repo_folder, "README.md")) + + +def log_validation(vae, text_encoder, tokenizer, unet, args, accelerator, weight_dtype, epoch): + logger.info("Running validation... ") + + pipeline = StableDiffusionPipeline.from_pretrained( + args.pretrained_model_name_or_path, + vae=accelerator.unwrap_model(vae), + text_encoder=accelerator.unwrap_model(text_encoder), + tokenizer=tokenizer, + unet=accelerator.unwrap_model(unet), + safety_checker=None, + revision=args.revision, + variant=args.variant, + torch_dtype=weight_dtype, + ) + pipeline = pipeline.to(accelerator.device) + 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) + + images = [] + 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.append(image) + + 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") + elif tracker.name == "wandb": + tracker.log( + { + "validation": [ + wandb.Image(image, caption=f"{i}: {args.validation_prompts[i]}") + for i, image in enumerate(images) + ] + } + ) + else: + logger.warning(f"image logging not implemented for {tracker.name}") + + del pipeline + torch.cuda.empty_cache() + + return images + + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of a training script.") + parser.add_argument( + "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1." + ) + 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( + "--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( + "--validation_prompts", + type=str, + default=None, + nargs="+", + help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."), + ) + parser.add_argument( + "--output_dir", + type=str, + default="sd-model-finetuned", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument( + "--cache_dir", + type=str, + default=None, + help="The directory where the downloaded models and datasets will be stored.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--resolution", + type=int, + default=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( + "--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("--use_ema", action="store_true", help="Whether to use EMA model.") + parser.add_argument( + "--non_ema_revision", + type=str, + default=None, + required=False, + help=( + "Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or" + " remote repository specified with --pretrained_model_name_or_path." + ), + ) + parser.add_argument( + "--dataloader_num_workers", + type=int, + default=16, + 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.prediction_type` is chosen.", + ) + parser.add_argument( + "--hub_model_id", + type=str, + default=None, + help="The name of the repository to keep in sync with the local `output_dir`.", + ) + parser.add_argument( + "--logging_dir", + type=str, + default="logs", + help=( + "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" + " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." + ), + ) + parser.add_argument( + "--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( + "--validation_epochs", + type=int, + default=5, + help="Run validation every X epochs.", + ) + parser.add_argument( + "--tracker_project_name", + type=str, + default="text2image-fine-tune", + help=( + "The `project_name` argument passed to Accelerator.init_trackers for" + " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator" + ), + ) + parser.add_argument( + "--NHWC", + action="store_true", + help="Whether or not using NHWC for training", + ) + parser.add_argument( + "--apex_fused_adam", + action="store_true", + help="Whether or not using fused_adam optimizer", + ) + + 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.") + + # default to using the same revision for the non-ema model if not specified + if args.non_ema_revision is None: + args.non_ema_revision = args.revision + + return args + + +def main(): + args = parse_args() + + if args.report_to == "wandb" and args.hub_token is not None: + raise ValueError( + "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token." + " Please use `huggingface-cli login` to authenticate with the Hub." + ) + + if args.non_ema_revision is not None: + deprecate( + "non_ema_revision!=None", + "0.15.0", + message=( + "Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to" + " use `--variant=non_ema` instead." + ), + ) + logging_dir = os.path.join(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, + ) + + # 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 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 + ) + + def deepspeed_zero_init_disabled_context_manager(): + """ + returns either a context list that includes one that will disable zero.Init or an empty context list + """ + deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None + if deepspeed_plugin is None: + return [] + + return [deepspeed_plugin.zero3_init_context_manager(enable=False)] + + # Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3. + # For this to work properly all models must be run through `accelerate.prepare`. But accelerate + # will try to assign the same optimizer with the same weights to all models during + # `deepspeed.initialize`, which of course doesn't work. + # + # For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2 + # frozen models from being partitioned during `zero.Init` which gets called during + # `from_pretrained` So CLIPTextModel and AutoencoderKL will not enjoy the parameter sharding + # across multiple gpus and only UNet2DConditionModel will get ZeRO sharded. + with ContextManagers(deepspeed_zero_init_disabled_context_manager()): + text_encoder = CLIPTextModel.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant + ) + 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.non_ema_revision + ) + # Freeze vae and text_encoder and set unet to trainable + vae.requires_grad_(False) + text_encoder.requires_grad_(False) + unet.train() + + # Create EMA for the unet. + if args.use_ema: + ema_unet = UNet2DConditionModel.from_pretrained( + args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant + ) + ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config) + + if args.enable_xformers_memory_efficient_attention: + if is_xformers_available(): + import xformers + + xformers_version = version.parse(xformers.__version__) + if xformers_version == version.parse("0.0.16"): + logger.warning( + "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details." + ) + unet.enable_xformers_memory_efficient_attention() + else: + raise ValueError("xformers is not available. Make sure it is installed correctly") + + # `accelerate` 0.16.0 will have better support for customized saving + if version.parse(accelerate.__version__) >= version.parse("0.16.0"): + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + def save_model_hook(models, weights, output_dir): + if accelerator.is_main_process: + if args.use_ema: + ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema")) + + for i, model in enumerate(models): + model.save_pretrained(os.path.join(output_dir, "unet")) + + # make sure to pop weight so that corresponding model is not saved again + # weights.pop() + + def load_model_hook(models, input_dir): + if args.use_ema: + load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel) + ema_unet.load_state_dict(load_model.state_dict()) + ema_unet.to(accelerator.device) + del load_model + + for _ in range(len(models)): + # pop models so that they are not loaded again + model = models.pop() + + # load diffusers style into model + load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet") + model.register_to_config(**load_model.config) + + model.load_state_dict(load_model.state_dict()) + del load_model + + accelerator.register_save_state_pre_hook(save_model_hook) + accelerator.register_load_state_pre_hook(load_model_hook) + + if args.gradient_checkpointing: + unet.enable_gradient_checkpointing() + + # Enable TF32 for faster training on Ampere GPUs, + # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices + if args.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.scale_lr: + args.learning_rate = ( + args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes + ) + + # 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 + elif args.apex_fused_adam: + import apex + optimizer_cls = apex.optimizers.FusedAdam + + else: + optimizer_cls = torch.optim.AdamW + + optimizer = optimizer_cls( + unet.parameters(), + 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) + + # for testing ips + from datasets import concatenate_datasets + train_dataset = concatenate_datasets([train_dataset for i in range(10)]) + + + 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} + + # 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, + pin_memory=True, + prefetch_factor = 2 + ) + + # 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, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet, optimizer, train_dataloader, lr_scheduler + ) + + if args.use_ema: + ema_unet.to(accelerator.device) + + # For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision + # as these weights are only used for inference, keeping weights in full precision is not required. + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + args.mixed_precision = accelerator.mixed_precision + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + args.mixed_precision = accelerator.mixed_precision + + # Move text_encode and vae to gpu and cast to weight_dtype + text_encoder.to(accelerator.device, dtype=weight_dtype) + vae.to(accelerator.device, dtype=weight_dtype) + + # 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.tracker_project_name, tracker_config) + + # Function for unwrapping if model was compiled with `torch.compile`. + def unwrap_model(model): + model = accelerator.unwrap_model(model) + model = model._orig_mod if is_compiled_module(model) else model + return model + + # 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.NHWC: + unet = unet.to(memory_format=torch.channels_last) + vae = vae.to(memory_format=torch.channels_last) + + import time + for epoch in range(first_epoch, args.num_train_epochs): + train_loss = 0.0 + iter_start = time.time() + for step, batch in enumerate(train_dataloader): + with accelerator.accumulate(unet): + # Convert images to latent space + if args.NHWC: + batch["pixel_values"] = batch["pixel_values"].to(memory_format=torch.channels_last) + latents = vae.encode(batch["pixel_values"].to(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 + ) + if args.input_perturbation: + new_noise = noise + args.input_perturbation * torch.randn_like(noise) + 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) + if args.input_perturbation: + noisy_latents = noise_scheduler.add_noise(latents, new_noise, timesteps) + else: + noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) + + # Get the text embedding for conditioning + encoder_hidden_states = text_encoder(batch["input_ids"], return_dict=False)[0] + + # Get the target for loss depending on the prediction type + if args.prediction_type is not None: + # set prediction_type of scheduler if defined + noise_scheduler.register_to_config(prediction_type=args.prediction_type) + + if noise_scheduler.config.prediction_type == "epsilon": + target = noise + elif noise_scheduler.config.prediction_type == "v_prediction": + target = noise_scheduler.get_velocity(latents, noise, timesteps) + else: + raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") + + # Predict the noise residual and compute loss + if args.NHWC: + noisy_latents = noisy_latents.to(memory_format=torch.channels_last) + # timesteps = timesteps.to(memory_format=torch.channels_last) + # encoder_hidden_states = encoder_hidden_states.to(memory_format=torch.channels_last) + model_pred = unet(noisy_latents, timesteps, encoder_hidden_states, return_dict=False)[0] + + if args.snr_gamma is None: + loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") + else: + # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556. + # Since we predict the noise instead of x_0, the original formulation is slightly changed. + # This is discussed in Section 4.2 of the same paper. + snr = compute_snr(noise_scheduler, timesteps) + mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min( + dim=1 + )[0] + if noise_scheduler.config.prediction_type == "epsilon": + mse_loss_weights = mse_loss_weights / snr + elif noise_scheduler.config.prediction_type == "v_prediction": + mse_loss_weights = mse_loss_weights / (snr + 1) + + loss = F.mse_loss(model_pred.float(), target.float(), reduction="none") + loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights + loss = loss.mean() + + # 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: + accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm) + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + if args.use_ema: + ema_unet.step(unet.parameters()) + progress_bar.update(1) + global_step += 1 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + iter_elapse = time.time() - iter_start + iter_start = time.time() + ips_per_device = total_batch_size / iter_elapse / accelerator.num_processes + ips_per_gpu = ips_per_device * 2 + + 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}") + # 这段代码是为了解决NHWC时保存模型出错 + if args.NHWC: + origin_model = accelerator._models[0] + model = origin_model.to(memory_format=torch.contiguous_format) + accelerator._models[0] = model + 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], + "ips_per_device": ips_per_device, "ips_per_gpu": ips_per_gpu} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_prompts is not None and epoch % args.validation_epochs == 0: + if args.use_ema: + # Store the UNet parameters temporarily and load the EMA parameters to perform inference. + ema_unet.store(unet.parameters()) + ema_unet.copy_to(unet.parameters()) + log_validation( + vae, + text_encoder, + tokenizer, + unet, + args, + accelerator, + weight_dtype, + global_step, + ) + if args.use_ema: + # Switch back to the original UNet parameters. + ema_unet.restore(unet.parameters()) + + # Create the pipeline using the trained modules and save it. + accelerator.wait_for_everyone() + if accelerator.is_main_process: + unet = unwrap_model(unet) + if args.use_ema: + ema_unet.copy_to(unet.parameters()) + + pipeline = StableDiffusionPipeline.from_pretrained( + args.pretrained_model_name_or_path, + text_encoder=text_encoder, + vae=vae.to(memory_format=torch.contiguous_format) if args.NHWC else vae, + unet=unet.to(memory_format=torch.contiguous_format) if args.NHWC else unet, + revision=args.revision, + variant=args.variant, + ) + pipeline.save_pretrained(args.output_dir) + + # Run a final round of inference. + images = [] + if args.validation_prompts is not None: + 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.append(image) + + if args.push_to_hub: + save_model_card(args, repo_id, images, repo_folder=args.output_dir) + 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__": + main() diff --git a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py new file mode 100755 index 000000000..ae4cfa520 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py @@ -0,0 +1,1371 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2024 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 XL for text2image.""" + +import argparse +import functools +import gc +import logging +import math +import os +import random +import shutil +from pathlib import Path +import time + +import accelerate +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.state import AcceleratorState +from accelerate.logging import get_logger +from accelerate.utils import ProjectConfiguration, set_seed +from datasets import concatenate_datasets, 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 +from transformers.utils import ContextManagers + +import diffusers +from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionXLPipeline, UNet2DConditionModel +from diffusers.optimization import get_scheduler +from diffusers.training_utils import EMAModel, compute_snr +from diffusers.utils import check_min_version, is_wandb_available +from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card +from diffusers.utils.import_utils import is_xformers_available +from diffusers.utils.torch_utils import is_compiled_module + + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.27.0") + +logger = get_logger(__name__) + + +DATASET_NAME_MAPPING = { + "lambdalabs/pokemon-blip-captions": ("image", "text"), +} + + +def save_model_card( + repo_id: str, + images: list = None, + validation_prompt: str = None, + base_model: str = None, + dataset_name: str = None, + repo_folder: str = None, + vae_path: str = None, +): + img_str = "" + if images is not None: + for i, image in enumerate(images): + image.save(os.path.join(repo_folder, f"image_{i}.png")) + img_str += f"![img_{i}](./image_{i}.png)\n" + + model_description = f""" +# Text-to-image finetuning - {repo_id} + +This pipeline was finetuned from **{base_model}** on the **{dataset_name}** dataset. Below are some example images generated with the finetuned pipeline using the following prompt: {validation_prompt}: \n +{img_str} + +Special VAE used for training: {vae_path}. +""" + + model_card = load_or_create_model_card( + repo_id_or_path=repo_id, + from_training=True, + license="creativeml-openrail-m", + base_model=base_model, + model_description=model_description, + inference=True, + ) + + tags = [ + "stable-diffusion-xl", + "stable-diffusion-xl-diffusers", + "text-to-image", + "diffusers-training", + "diffusers", + ] + model_card = populate_model_card(model_card, tags=tags) + + model_card.save(os.path.join(repo_folder, "README.md")) + + +def import_model_class_from_model_name_or_path( + pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder" +): + text_encoder_config = PretrainedConfig.from_pretrained( + pretrained_model_name_or_path, subfolder=subfolder, revision=revision + ) + model_class = text_encoder_config.architectures[0] + + if model_class == "CLIPTextModel": + from transformers import CLIPTextModel + + return CLIPTextModel + elif model_class == "CLIPTextModelWithProjection": + from transformers import CLIPTextModelWithProjection + + return CLIPTextModelWithProjection + else: + raise ValueError(f"{model_class} is not supported.") + + +def parse_args(input_args=None): + parser = argparse.ArgumentParser(description="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( + "--variant", + type=str, + default=None, + help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16", + ) + parser.add_argument( + "--dataset_name", + type=str, + default=None, + help=( + "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private," + " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem," + " or to a folder containing files that 🤗 Datasets can understand." + ), + ) + parser.add_argument( + "--dataset_config_name", + type=str, + default=None, + help="The config of the Dataset, leave as None if there's only one config.", + ) + parser.add_argument( + "--train_data_dir", + type=str, + default=None, + help=( + "A folder containing the training data. Folder contents must follow the structure described in" + " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file" + " must exist to provide the captions for the images. Ignored if `dataset_name` is specified." + ), + ) + parser.add_argument( + "--image_column", type=str, default="image", help="The column of the dataset containing an image." + ) + parser.add_argument( + "--caption_column", + type=str, + default="text", + help="The column of the dataset containing a caption or a list of captions.", + ) + parser.add_argument( + "--validation_prompt", + type=str, + default=None, + help="A prompt that is used during validation to verify that the model is learning.", + ) + parser.add_argument( + "--num_validation_images", + type=int, + default=4, + help="Number of images that should be generated during validation with `validation_prompt`.", + ) + parser.add_argument( + "--validation_epochs", + type=int, + default=1, + help=( + "Run fine-tuning validation every X epochs. The validation process consists of running the prompt" + " `args.validation_prompt` multiple times: `args.num_validation_images`." + ), + ) + parser.add_argument( + "--max_train_samples", + type=int, + default=None, + help=( + "For debugging purposes or quicker training, truncate the number of training examples to this " + "value if set." + ), + ) + parser.add_argument( + "--proportion_empty_prompts", + type=float, + default=0, + help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).", + ) + parser.add_argument( + "--output_dir", + type=str, + default="sdxl-model-finetuned", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument( + "--cache_dir", + type=str, + default=None, + help="The directory where the downloaded models and datasets will be stored.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--resolution", + type=int, + default=1024, + help=( + "The resolution for input images, all the images in the train/validation dataset will be resized to this" + " resolution" + ), + ) + parser.add_argument( + "--center_crop", + default=False, + action="store_true", + help=( + "Whether to center crop the input images to the resolution. If not set, the images will be randomly" + " cropped. The images will be resized to the resolution first before cropping." + ), + ) + parser.add_argument( + "--random_flip", + action="store_true", + help="whether to randomly flip images horizontally", + ) + parser.add_argument( + "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader." + ) + parser.add_argument("--num_train_epochs", type=int, default=100) + parser.add_argument( + "--max_train_steps", + type=int, + default=None, + help="Total number of training steps to perform. If provided, overrides num_train_epochs.", + ) + parser.add_argument( + "--checkpointing_steps", + type=int, + default=500, + help=( + "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final" + " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming" + " training using `--resume_from_checkpoint`." + ), + ) + parser.add_argument( + "--checkpoints_total_limit", + type=int, + default=None, + help=("Max number of checkpoints to store."), + ) + parser.add_argument( + "--resume_from_checkpoint", + type=str, + default=None, + help=( + "Whether training should be resumed from a previous checkpoint. Use a path saved by" + ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.' + ), + ) + parser.add_argument( + "--gradient_accumulation_steps", + type=int, + default=1, + help="Number of updates steps to accumulate before performing a backward/update pass.", + ) + parser.add_argument( + "--gradient_checkpointing", + action="store_true", + help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.", + ) + parser.add_argument( + "--learning_rate", + type=float, + default=1e-4, + help="Initial learning rate (after the potential warmup period) to use.", + ) + parser.add_argument( + "--scale_lr", + action="store_true", + default=False, + help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.", + ) + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--timestep_bias_strategy", + type=str, + default="none", + choices=["earlier", "later", "range", "none"], + help=( + "The timestep bias strategy, which may help direct the model toward learning low or high frequency details." + " Choices: ['earlier', 'later', 'range', 'none']." + " The default is 'none', which means no bias is applied, and training proceeds normally." + " The value of 'later' will increase the frequency of the model's final training timesteps." + ), + ) + parser.add_argument( + "--timestep_bias_multiplier", + type=float, + default=1.0, + help=( + "The multiplier for the bias. Defaults to 1.0, which means no bias is applied." + " A value of 2.0 will double the weight of the bias, and a value of 0.5 will halve it." + ), + ) + parser.add_argument( + "--timestep_bias_begin", + type=int, + default=0, + help=( + "When using `--timestep_bias_strategy=range`, the beginning (inclusive) timestep to bias." + " Defaults to zero, which equates to having no specific bias." + ), + ) + parser.add_argument( + "--timestep_bias_end", + type=int, + default=1000, + help=( + "When using `--timestep_bias_strategy=range`, the final timestep (inclusive) to bias." + " Defaults to 1000, which is the number of timesteps that Stable Diffusion is trained on." + ), + ) + parser.add_argument( + "--timestep_bias_portion", + type=float, + default=0.25, + help=( + "The portion of timesteps to bias. Defaults to 0.25, which 25% of timesteps will be biased." + " A value of 0.5 will bias one half of the timesteps. The value provided for `--timestep_bias_strategy` determines" + " whether the biased portions are in the earlier or later timesteps." + ), + ) + parser.add_argument( + "--snr_gamma", + type=float, + default=None, + help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. " + "More details here: https://arxiv.org/abs/2303.09556.", + ) + parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.") + parser.add_argument( + "--allow_tf32", + action="store_true", + help=( + "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see" + " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices" + ), + ) + parser.add_argument( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument( + "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.") + parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer") + parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.") + parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.") + parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.") + parser.add_argument( + "--prediction_type", + type=str, + default=None, + help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediction_type` is chosen.", + ) + parser.add_argument( + "--hub_model_id", + type=str, + default=None, + help="The name of the repository to keep in sync with the local `output_dir`.", + ) + parser.add_argument( + "--logging_dir", + type=str, + default="logs", + help=( + "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to" + " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***." + ), + ) + parser.add_argument( + "--report_to", + type=str, + default="tensorboard", + help=( + 'The integration to report the results and logs to. Supported platforms are `"tensorboard"`' + ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.' + ), + ) + parser.add_argument( + "--mixed_precision", + type=str, + default=None, + choices=["no", "fp16", "bf16"], + help=( + "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >=" + " 1.10.and an Nvidia Ampere GPU. Default to the value of accelerate config of the current system or the" + " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config." + ), + ) + parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank") + parser.add_argument( + "--enable_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( + "--NHWC", + action="store_true", + help="Whether or not using NHWC for training", + ) + parser.add_argument( + "--apex_fused_adam", + action="store_true", + help="Whether or not using fused_adam optimizer", + ) + + if input_args is not None: + args = parser.parse_args(input_args) + else: + args = parser.parse_args() + + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + # Sanity checks + if args.dataset_name is None and args.train_data_dir is None: + raise ValueError("Need either a dataset name or a training folder.") + + if args.proportion_empty_prompts < 0 or args.proportion_empty_prompts > 1: + raise ValueError("`--proportion_empty_prompts` must be in the range [0, 1].") + + return args + + +# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt +def encode_prompt(batch, text_encoders, tokenizers, proportion_empty_prompts, caption_column, is_train=True): + prompt_embeds_list = [] + prompt_batch = batch[caption_column] + + captions = [] + for caption in prompt_batch: + if random.random() < proportion_empty_prompts: + captions.append("") + elif isinstance(caption, str): + captions.append(caption) + elif isinstance(caption, (list, np.ndarray)): + # take a random caption if there are multiple + captions.append(random.choice(caption) if is_train else caption[0]) + + with torch.no_grad(): + for tokenizer, text_encoder in zip(tokenizers, text_encoders): + text_inputs = tokenizer( + captions, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + prompt_embeds = text_encoder( + text_input_ids.to(text_encoder.device), + output_hidden_states=True, + return_dict=False, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + prompt_embeds = prompt_embeds[-1][-2] + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1) + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1) + return {"prompt_embeds": prompt_embeds.cpu(), "pooled_prompt_embeds": pooled_prompt_embeds.cpu()} + + +def compute_vae_encodings(batch, vae): + memory_format = torch.channels_last if int(os.environ["USE_NHWC_GN"]) else torch.contiguous_format + images = batch.pop("pixel_values") + pixel_values = torch.stack(list(images)) + pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float() + pixel_values = pixel_values.to(vae.device, dtype=vae.dtype, memory_format=memory_format) + + with torch.no_grad(): + model_input = vae.encode(pixel_values).latent_dist.sample() + model_input = model_input.to(memory_format=torch.contiguous_format) * vae.config.scaling_factor + return {"model_input": model_input.cpu()} + + +def generate_timestep_weights(args, num_timesteps): + weights = torch.ones(num_timesteps) + + # Determine the indices to bias + num_to_bias = int(args.timestep_bias_portion * num_timesteps) + + if args.timestep_bias_strategy == "later": + bias_indices = slice(-num_to_bias, None) + elif args.timestep_bias_strategy == "earlier": + bias_indices = slice(0, num_to_bias) + elif args.timestep_bias_strategy == "range": + # Out of the possible 1000 timesteps, we might want to focus on eg. 200-500. + range_begin = args.timestep_bias_begin + range_end = args.timestep_bias_end + if range_begin < 0: + raise ValueError( + "When using the range strategy for timestep bias, you must provide a beginning timestep greater or equal to zero." + ) + if range_end > num_timesteps: + raise ValueError( + "When using the range strategy for timestep bias, you must provide an ending timestep smaller than the number of timesteps." + ) + bias_indices = slice(range_begin, range_end) + else: # 'none' or any other string + return weights + if args.timestep_bias_multiplier <= 0: + return ValueError( + "The parameter --timestep_bias_multiplier is not intended to be used to disable the training of specific timesteps." + " If it was intended to disable timestep bias, use `--timestep_bias_strategy none` instead." + " A timestep bias multiplier less than or equal to 0 is not allowed." + ) + + # Apply the bias + weights[bias_indices] *= args.timestep_bias_multiplier + + # Normalize + weights /= weights.sum() + + return weights + + +def main(args): + if int(os.environ.get("USE_NHWC_GN", 0)): + assert args.NHWC, "USE_NHWC_GN requires NHWC to be true" + assert int(os.supports_bytes_environ) + if args.report_to == "wandb" and args.hub_token is not None: + raise ValueError( + "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token." + " Please use `huggingface-cli login` to authenticate with the Hub." + ) + + logging_dir = Path(args.output_dir, args.logging_dir) + + accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir) + + 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") + + def deepspeed_zero_init_disabled_context_manager(): + """ + returns either a context list that includes one that will disable zero.Init or an empty context list + """ + deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None + if deepspeed_plugin is None: + return [] + + return [deepspeed_plugin.zero3_init_context_manager(enable=False)] + + 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 + ) + + with ContextManagers(deepspeed_zero_init_disabled_context_manager()): + # Check for terminal SNR in combination with SNR Gamma + text_encoder_one = text_encoder_cls_one.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant + ) + text_encoder_two = text_encoder_cls_two.from_pretrained( + args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant + ) + # vae 因为需要用float32的缘故，其中的attn部分需要用native实现，因为flash-attn 不支持float32 + origin_attn = os.environ.get("USE_NATIVE_ATTN", 0) + os.environ["USE_NATIVE_ATTN"] = "1" + vae = AutoencoderKL.from_pretrained( + vae_path, + subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None, + revision=args.revision, + variant=args.variant, + ) + os.environ["USE_NATIVE_ATTN"] = origin_attn + unet = UNet2DConditionModel.from_pretrained( + args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant + ) + + # Freeze vae and text encoders. + vae.requires_grad_(False) + text_encoder_one.requires_grad_(False) + text_encoder_two.requires_grad_(False) + # Set unet as trainable. + unet.train() + + # For mixed precision training we cast all non-trainable weights to half-precision + # as these weights are only used for inference, keeping weights in full precision is not required. + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + + # Move unet, vae and text_encoder to device and cast to weight_dtype + # The VAE is in float32 to avoid NaN losses. + vae.to(accelerator.device, dtype=torch.float32) + text_encoder_one.to(accelerator.device, dtype=weight_dtype) + text_encoder_two.to(accelerator.device, dtype=weight_dtype) + + # Create EMA for the unet. + if args.use_ema: + ema_unet = UNet2DConditionModel.from_pretrained( + args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant + ) + ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config) + + if args.enable_xformers_memory_efficient_attention: + if is_xformers_available(): + import xformers + + xformers_version = version.parse(xformers.__version__) + if xformers_version == version.parse("0.0.16"): + logger.warning( + "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details." + ) + unet.enable_xformers_memory_efficient_attention() + else: + raise ValueError("xformers is not available. Make sure it is installed correctly") + + # `accelerate` 0.16.0 will have better support for customized saving + if version.parse(accelerate.__version__) >= version.parse("0.16.0"): + # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format + def save_model_hook(models, weights, output_dir): + if accelerator.is_main_process: + if args.use_ema: + ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema")) + + for i, model in enumerate(models): + model.save_pretrained(os.path.join(output_dir, "unet")) + + # make sure to pop weight so that corresponding model is not saved again + weights.pop() + + def load_model_hook(models, input_dir): + if args.use_ema: + load_model = EMAModel.from_pretrained(os.path.join(input_dir, "unet_ema"), UNet2DConditionModel) + ema_unet.load_state_dict(load_model.state_dict()) + ema_unet.to(accelerator.device) + del load_model + + for _ in range(len(models)): + # pop models so that they are not loaded again + model = models.pop() + + # load diffusers style into model + load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet") + model.register_to_config(**load_model.config) + + model.load_state_dict(load_model.state_dict()) + del load_model + + accelerator.register_save_state_pre_hook(save_model_hook) + accelerator.register_load_state_pre_hook(load_model_hook) + + if args.gradient_checkpointing: + unet.enable_gradient_checkpointing() + + # Enable TF32 for faster training on Ampere GPUs, + # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices + if args.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.scale_lr: + args.learning_rate = ( + args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes + ) + + # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`." + ) + + optimizer_class = bnb.optim.AdamW8bit + elif args.apex_fused_adam: + import apex + optimizer_class = apex.optimizers.FusedAdam + else: + optimizer_class = torch.optim.AdamW + + # Optimizer creation + params_to_optimize = unet.parameters() + optimizer = optimizer_class( + params_to_optimize, + lr=args.learning_rate, + betas=(args.adam_beta1, args.adam_beta2), + weight_decay=args.adam_weight_decay, + eps=args.adam_epsilon, + ) + + # Get the datasets: you can either provide your own training and evaluation files (see below) + # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub). + + # In distributed training, the load_dataset function guarantees that only one local process can concurrently + # download the dataset. + if args.dataset_name is not None: + # Downloading and loading a dataset from the hub. + dataset = load_dataset( + args.dataset_name, + args.dataset_config_name, + cache_dir=args.cache_dir, + ) + else: + data_files = {} + if args.train_data_dir is not None: + data_files["train"] = os.path.join(args.train_data_dir, "**") + dataset = load_dataset( + "imagefolder", + data_files=data_files, + cache_dir=args.cache_dir, + ) + # See more about loading custom images at + # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder + + # Preprocessing the datasets. + # We need to tokenize inputs and targets. + column_names = dataset["train"].column_names + + # 6. Get the column names for input/target. + dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None) + if args.image_column is None: + image_column = dataset_columns[0] if dataset_columns is not None else column_names[0] + else: + image_column = args.image_column + if image_column not in column_names: + raise ValueError( + f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}" + ) + if args.caption_column is None: + caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1] + else: + caption_column = args.caption_column + if caption_column not in column_names: + raise ValueError( + f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}" + ) + + # Preprocessing the datasets. + train_resize = transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR) + train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution) + train_flip = transforms.RandomHorizontalFlip(p=1.0) + train_transforms = transforms.Compose([transforms.ToTensor(), transforms.Normalize([0.5], [0.5])]) + + def preprocess_train(examples): + images = [image.convert("RGB") for image in examples[image_column]] + # image aug + original_sizes = [] + all_images = [] + crop_top_lefts = [] + for image in images: + original_sizes.append((image.height, image.width)) + image = train_resize(image) + if args.random_flip and random.random() < 0.5: + # flip + image = train_flip(image) + if args.center_crop: + y1 = max(0, int(round((image.height - args.resolution) / 2.0))) + x1 = max(0, int(round((image.width - args.resolution) / 2.0))) + image = train_crop(image) + else: + y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution)) + image = crop(image, y1, x1, h, w) + crop_top_left = (y1, x1) + crop_top_lefts.append(crop_top_left) + image = train_transforms(image) + all_images.append(image) + + examples["original_sizes"] = original_sizes + examples["crop_top_lefts"] = crop_top_lefts + examples["pixel_values"] = all_images + return examples + + with accelerator.main_process_first(): + if args.max_train_samples is not None: + dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples)) + # Set the training transforms + train_dataset = dataset["train"].with_transform(preprocess_train) + + # Let's first compute all the embeddings so that we can free up the text encoders + # from memory. We will pre-compute the VAE encodings too. + text_encoders = [text_encoder_one, text_encoder_two] + tokenizers = [tokenizer_one, tokenizer_two] + compute_embeddings_fn = functools.partial( + encode_prompt, + text_encoders=text_encoders, + tokenizers=tokenizers, + proportion_empty_prompts=args.proportion_empty_prompts, + caption_column=args.caption_column, + ) + compute_vae_encodings_fn = functools.partial(compute_vae_encodings, vae=vae) + with accelerator.main_process_first(): + from datasets.fingerprint import Hasher + + # fingerprint used by the cache for the other processes to load the result + # details: https://github.com/huggingface/diffusers/pull/4038#discussion_r1266078401 + new_fingerprint = Hasher.hash(args) + new_fingerprint_for_vae = Hasher.hash(vae_path) + train_dataset_with_embeddings = train_dataset.map( + compute_embeddings_fn, batched=True, new_fingerprint=new_fingerprint + ) + train_dataset_with_vae = train_dataset.map( + compute_vae_encodings_fn, + batched=True, + # batch_size=args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps,\ + batch_size=args.train_batch_size, + new_fingerprint=new_fingerprint_for_vae, + ) + precomputed_dataset = concatenate_datasets( + [train_dataset_with_embeddings, train_dataset_with_vae.remove_columns(["image", "text"])], axis=1 + ) + precomputed_dataset = precomputed_dataset.with_transform(preprocess_train) + + del text_encoders, tokenizers, vae + gc.collect() + torch.cuda.empty_cache() + + def collate_fn(examples): + model_input = torch.stack([torch.tensor(example["model_input"]) for example in examples]) + original_sizes = [example["original_sizes"] for example in examples] + crop_top_lefts = [example["crop_top_lefts"] for example in examples] + prompt_embeds = torch.stack([torch.tensor(example["prompt_embeds"]) for example in examples]) + pooled_prompt_embeds = torch.stack([torch.tensor(example["pooled_prompt_embeds"]) for example in examples]) + + return { + "model_input": model_input, + "prompt_embeds": prompt_embeds, + "pooled_prompt_embeds": pooled_prompt_embeds, + "original_sizes": original_sizes, + "crop_top_lefts": crop_top_lefts, + } + + # DataLoaders creation: + # for testing ips + precomputed_dataset = concatenate_datasets([precomputed_dataset for i in range(10)]) + + train_dataloader = torch.utils.data.DataLoader( + precomputed_dataset, + shuffle=True, + collate_fn=collate_fn, + batch_size=args.train_batch_size, + num_workers=args.dataloader_num_workers, + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, + num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, + ) + + # Prepare everything with our `accelerator`. + unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet, optimizer, train_dataloader, lr_scheduler + ) + + if args.use_ema: + ema_unet.to(accelerator.device) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + accelerator.init_trackers("text2image-fine-tune-sdxl", config=vars(args)) + + # Function for unwrapping if torch.compile() was used in accelerate. + def unwrap_model(model): + model = accelerator.unwrap_model(model) + model = model._orig_mod if is_compiled_module(model) else model + return model + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(precomputed_dataset)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + # Potentially load in the weights and states from a previous save + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + initial_global_step = 0 + else: + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + global_step = int(path.split("-")[1]) + + initial_global_step = global_step + first_epoch = global_step // num_update_steps_per_epoch + + else: + initial_global_step = 0 + + progress_bar = tqdm( + range(0, args.max_train_steps), + initial=initial_global_step, + desc="Steps", + # Only show the progress bar once on each machine. + disable=not accelerator.is_local_main_process, + ) + + if args.NHWC: + unet = unet.to(memory_format=torch.channels_last) + + for epoch in range(first_epoch, args.num_train_epochs): + train_loss = 0.0 + iter_start = time.time() + for step, batch in enumerate(train_dataloader): + with accelerator.accumulate(unet): + # Sample noise that we'll add to the latents + model_input = batch["model_input"].to(accelerator.device) + noise = torch.randn_like(model_input) + if args.noise_offset: + # https://www.crosslabs.org//blog/diffusion-with-offset-noise + noise += args.noise_offset * torch.randn( + (model_input.shape[0], model_input.shape[1], 1, 1), device=model_input.device + ) + + bsz = model_input.shape[0] + if args.timestep_bias_strategy == "none": + # Sample a random timestep for each image without bias. + timesteps = torch.randint( + 0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device + ) + else: + # Sample a random timestep for each image, potentially biased by the timestep weights. + # Biasing the timestep weights allows us to spend less time training irrelevant timesteps. + weights = generate_timestep_weights(args, noise_scheduler.config.num_train_timesteps).to( + model_input.device + ) + timesteps = torch.multinomial(weights, bsz, replacement=True).long() + + # Add noise to the model input according to the noise magnitude at each timestep + # (this is the forward diffusion process) + noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps) + + # 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 = batch["prompt_embeds"].to(accelerator.device) + pooled_prompt_embeds = batch["pooled_prompt_embeds"].to(accelerator.device) + unet_added_conditions.update({"text_embeds": pooled_prompt_embeds}) + + if args.NHWC: + noisy_model_input = noisy_model_input.to(memory_format=torch.channels_last) + model_pred = unet( + noisy_model_input, + timesteps, + prompt_embeds, + added_cond_kwargs=unet_added_conditions, + return_dict=False, + )[0] + + # Get the target for loss depending on the prediction type + if args.prediction_type is not None: + # set prediction_type of scheduler if defined + noise_scheduler.register_to_config(prediction_type=args.prediction_type) + + if noise_scheduler.config.prediction_type == "epsilon": + target = noise + elif noise_scheduler.config.prediction_type == "v_prediction": + target = noise_scheduler.get_velocity(model_input, noise, timesteps) + elif noise_scheduler.config.prediction_type == "sample": + # We set the target to latents here, but the model_pred will return the noise sample prediction. + target = model_input + # We will have to subtract the noise residual from the prediction to get the target sample. + model_pred = model_pred - noise + else: + raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") + + if args.snr_gamma is None: + loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") + else: + # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556. + # Since we predict the noise instead of x_0, the original formulation is slightly changed. + # This is discussed in Section 4.2 of the same paper. + snr = compute_snr(noise_scheduler, timesteps) + mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min( + dim=1 + )[0] + if noise_scheduler.config.prediction_type == "epsilon": + mse_loss_weights = mse_loss_weights / snr + elif noise_scheduler.config.prediction_type == "v_prediction": + mse_loss_weights = mse_loss_weights / (snr + 1) + + loss = F.mse_loss(model_pred.float(), target.float(), reduction="none") + loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights + loss = loss.mean() + + # 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 = unet.parameters() + accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm) + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + if args.use_ema: + ema_unet.step(unet.parameters()) + progress_bar.update(1) + global_step += 1 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + iter_elapse = time.time() - iter_start + iter_start = time.time() + ips_per_device = total_batch_size / iter_elapse / accelerator.num_processes + ips_per_gpu = ips_per_device * 2 + + 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}") + #if args.NHWC: + # origin_model = accelerator._models[0] + # model = origin_model.to(memory_format=torch.contiguous_format) + # accelerator._models[0] = model + + 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], + "ips_per_device": ips_per_device, "ips_per_gpu": ips_per_gpu} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_prompt is not None and epoch % args.validation_epochs == 0: + logger.info( + f"Running validation... \n Generating {args.num_validation_images} images with prompt:" + f" {args.validation_prompt}." + ) + if args.use_ema: + # Store the UNet parameters temporarily and load the EMA parameters to perform inference. + ema_unet.store(unet.parameters()) + ema_unet.copy_to(unet.parameters()) + + # create pipeline + vae = AutoencoderKL.from_pretrained( + vae_path, + subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None, + revision=args.revision, + variant=args.variant, + ) + pipeline = StableDiffusionXLPipeline.from_pretrained( + args.pretrained_model_name_or_path, + vae=vae, + unet=accelerator.unwrap_model(unet), + revision=args.revision, + variant=args.variant, + torch_dtype=weight_dtype, + ) + if args.prediction_type is not None: + scheduler_args = {"prediction_type": args.prediction_type} + pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args) + + pipeline = pipeline.to(accelerator.device) + pipeline.set_progress_bar_config(disable=True) + + # run inference + generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None + pipeline_args = {"prompt": args.validation_prompt} + + with torch.cuda.amp.autocast(): + images = [ + pipeline(**pipeline_args, generator=generator, num_inference_steps=25).images[0] + for _ in range(args.num_validation_images) + ] + + for tracker in accelerator.trackers: + if tracker.name == "tensorboard": + np_images = np.stack([np.asarray(img) for img in images]) + tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC") + if tracker.name == "wandb": + tracker.log( + { + "validation": [ + wandb.Image(image, caption=f"{i}: {args.validation_prompt}") + for i, image in enumerate(images) + ] + } + ) + + del pipeline + torch.cuda.empty_cache() + + accelerator.wait_for_everyone() + if accelerator.is_main_process: + unet = unwrap_model(unet) + if args.use_ema: + ema_unet.copy_to(unet.parameters()) + + # Serialize pipeline. + vae = AutoencoderKL.from_pretrained( + vae_path, + subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None, + revision=args.revision, + variant=args.variant, + torch_dtype=weight_dtype, + ) + pipeline = StableDiffusionXLPipeline.from_pretrained( + args.pretrained_model_name_or_path, + unet=unet, + vae=vae, + revision=args.revision, + variant=args.variant, + torch_dtype=weight_dtype, + ) + if args.prediction_type is not None: + scheduler_args = {"prediction_type": args.prediction_type} + pipeline.scheduler = pipeline.scheduler.from_config(pipeline.scheduler.config, **scheduler_args) + pipeline.save_pretrained(args.output_dir) + + # run inference + images = [] + if args.validation_prompt and args.num_validation_images > 0: + pipeline = pipeline.to(accelerator.device) + generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None + with torch.cuda.amp.autocast(): + images = [ + pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0] + for _ in range(args.num_validation_images) + ] + + for tracker in accelerator.trackers: + if tracker.name == "tensorboard": + np_images = np.stack([np.asarray(img) for img in images]) + tracker.writer.add_images("test", np_images, epoch, dataformats="NHWC") + if tracker.name == "wandb": + tracker.log( + { + "test": [ + wandb.Image(image, caption=f"{i}: {args.validation_prompt}") + for i, image in enumerate(images) + ] + } + ) + + if args.push_to_hub: + save_model_card( + repo_id=repo_id, + images=images, + validation_prompt=args.validation_prompt, + base_model=args.pretrained_model_name_or_path, + dataset_name=args.dataset_name, + repo_folder=args.output_dir, + vae_path=args.pretrained_vae_model_name_or_path, + ) + upload_folder( + repo_id=repo_id, + folder_path=args.output_dir, + commit_message="End of training", + ignore_patterns=["step_*", "epoch_*"], + ) + + accelerator.end_training() + + +if __name__ == "__main__": + args = parse_args() + main(args) -- Gitee From 42486ca8b96f87f31dc1efea6a41cfa12bfcbf86 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=9D=8E=E4=B8=80=E5=8A=9B?= Date: Thu, 18 Jul 2024 10:23:05 +0800 Subject: [PATCH 2/7] fix README --- .../diffusion/stable-diffusion/diffusers/README.md | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/multimodal/diffusion/stable-diffusion/diffusers/README.md b/multimodal/diffusion/stable-diffusion/diffusers/README.md index c51b106f8..56c70a0ab 100644 --- a/multimodal/diffusion/stable-diffusion/diffusers/README.md +++ b/multimodal/diffusion/stable-diffusion/diffusers/README.md @@ -1,4 +1,4 @@ -table Diffusion +# Stable Diffusion ## Model description @@ -7,8 +7,15 @@ Stable Diffusion is a latent text-to-image diffusion model. ## Step 1: Installation - Install - ```bash +wget http://files.deepspark.org.cn:880/deepspark/add-ons/deepspeed-0.14.3+corex.20240718020249-cp310-cp310-linux_x86_64.whl +wget http://files.deepspark.org.cn:880/deepspark/add-ons/diffusers-0.22.0-py3-none-any.whl +wget http://files.deepspark.org.cn:880/deepspark/add-ons/transformers-4.38.1-py3-none-any.whl +``` +```bash +pip3 install deepspeed-0.14.3+corex.20240718020249-cp310-cp310-linux_x86_64.whl +pip3 install diffusers-0.22.0-py3-none-any.whl +pip3 install transformers-4.38.1-py3-none-any.whl pip3 install -r requirements.txt pip3 install pillow --upgrade ``` -- Gitee From c4752af231cb92319d301bb7206730c0076ee377 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=9D=8E=E4=B8=80=E5=8A=9B?= Date: Thu, 18 Jul 2024 10:28:14 +0800 Subject: [PATCH 3/7] add copyright --- .../diffusers/run_sd_1.5_multi.sh | 16 ++++++++++++++++ .../diffusers/run_sd_1.5_single.sh | 16 ++++++++++++++++ .../diffusers/run_sd_2.1_multi.sh | 16 ++++++++++++++++ .../diffusers/run_sd_2.1_single.sh | 16 ++++++++++++++++ .../stable-diffusion/diffusers/run_sd_xl.sh | 16 ++++++++++++++++ .../diffusers/train_text_to_image.py | 16 ++++++++++++++++ .../diffusers/train_text_to_image_sdxl.py | 16 ++++++++++++++++ 7 files changed, 112 insertions(+) diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh index d95cc8e60..6f53ca558 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh @@ -1,3 +1,19 @@ +# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# 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. + + export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh index f21d71687..b5e328c36 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh @@ -1,3 +1,19 @@ +# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# 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. + + export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh index 3ec9edbc7..830101960 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh @@ -1,3 +1,19 @@ +# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# 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. + + export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh index a865a5f05..7bc92db1c 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh @@ -1,3 +1,19 @@ +# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# 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. + + export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh index 6d532eb59..7f9cad5f2 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh @@ -1,3 +1,19 @@ +# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# 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. + + export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 diff --git a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py index ada797894..2efd5c4d1 100644 --- a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py +++ b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py @@ -1,3 +1,19 @@ +# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# 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. + + #!/usr/bin/env python # coding=utf-8 # Copyright 2024 The HuggingFace Inc. team. All rights reserved. diff --git a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py index ae4cfa520..27cdd493b 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py +++ b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py @@ -1,3 +1,19 @@ +# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# 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. + + #!/usr/bin/env python # coding=utf-8 # Copyright 2024 The HuggingFace Inc. team. All rights reserved. -- Gitee From a4d6b1617e1e38a3f4603969237ad96ef3798b9a Mon Sep 17 00:00:00 2001 From: majorli Date: Fri, 26 Jul 2024 11:01:43 +0800 Subject: [PATCH 4/7] add stable diffusion 1.5 model Signed-off-by: majorli --- .../diffusers/run_sd_1.5_multi.sh | 4 +- .../diffusers/run_sd_1.5_single.sh | 4 +- .../stable-diffusion/sd_1.5/README.md | 53 +++++++++++++++++++ 3 files changed, 57 insertions(+), 4 deletions(-) create mode 100644 multimodal/diffusion/stable-diffusion/sd_1.5/README.md diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh index 6f53ca558..a3acee13f 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh @@ -39,8 +39,8 @@ accelerate launch --config_file configs/default_config.yaml --mixed_precision="f --max_grad_norm=1 \ --lr_scheduler="constant" \ --lr_warmup_steps=0 \ - --output_dir="sd-pokemon-model-3" \ - --max_train_steps=100 \ + --output_dir="sd-pokemon-model-multi" \ + --max_train_steps=900 \ --NHWC \ --dataloader_num_workers=32 \ --apex_fused_adam diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh index b5e328c36..9ca1b0e4d 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh @@ -38,8 +38,8 @@ accelerate launch --config_file configs/single_config.yaml --mixed_precision="fp --max_grad_norm=1 \ --lr_scheduler="constant" \ --lr_warmup_steps=0 \ - --output_dir="sd-pokemon-model-3" \ - --max_train_steps=100 \ + --output_dir="sd-pokemon-model-single" \ + --max_train_steps=15000 \ --NHWC \ --dataloader_num_workers=32 \ --apex_fused_adam diff --git a/multimodal/diffusion/stable-diffusion/sd_1.5/README.md b/multimodal/diffusion/stable-diffusion/sd_1.5/README.md new file mode 100644 index 000000000..5177fb91b --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/sd_1.5/README.md @@ -0,0 +1,53 @@ +# Stable Diffusion 1.5 + +## Model description + +Stable Diffusion is a latent text-to-image diffusion model capable of generating photo-realistic images given any text input. + +## Step 1: Preparation + +You just need to run the script, and it will automatically download the required data and weights. Or you can manually download the weights and data locally. + +dataset: download the lambdalabs/pokemon-blip-captions from [huggingface page](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions). + +weights: download the runwayml/stable-diffusion-v1-5 from [huggingface page](https://huggingface.co/runwayml/stable-diffusion-v1-5). + +## Step 2: Installation + +```bash +pip3 install http://files.deepspark.org.cn:880/deepspark/add-ons/deepspeed-0.14.3+corex.20240718020249-cp310-cp310-linux_x86_64.whl +pip3 install http://files.deepspark.org.cn:880/deepspark/add-ons/diffusers-0.22.0-py3-none-any.whl +pip3 install http://files.deepspark.org.cn:880/deepspark/add-ons/transformers-4.38.1-py3-none-any.whl +pip3 install -r requirements.txt +pip3 install pillow --upgrade +``` + +## Step 3: Training + +If you have downloaded the weights and dataset, please export the environment variables like below. + +```bash +export MODEL_PATH=/path/to/sd_weights +export DATASET_PATH=/path/to/data +``` + +```bash +# Go to diffusers path +cd ${PROJ_ROOT}/multimodal/diffusion/stable-diffusion/diffusers + +# Single GPU +bash run_sd_1.5_single.sh + +# Multi GPUs +bash run_sd_1.5_multi.sh +``` + +## Results + +| Model | GPUs | ips_per_device | ips_per_gpu | +| ------ | ------- | -------------- | ----------- | +| SD 1.5 | BI-V150 | 6.76 | 13.5 | + +## Reference + +- [diffusers](https://github.com/huggingface/diffusers) -- Gitee From 17806949c95532b33adf16c1d555fa9f99cd40d4 Mon Sep 17 00:00:00 2001 From: majorli Date: Fri, 26 Jul 2024 11:09:36 +0800 Subject: [PATCH 5/7] add stable diffusion 2.1 model Signed-off-by: majorli --- .../diffusers/run_sd_1.5_multi.sh | 4 +- .../diffusers/run_sd_1.5_single.sh | 4 +- .../diffusers/run_sd_2.1_multi.sh | 7 +-- .../diffusers/run_sd_2.1_single.sh | 6 +- .../stable-diffusion/sd_1.5/README.md | 8 ++- .../stable-diffusion/sd_2.1/README.md | 57 +++++++++++++++++++ 6 files changed, 69 insertions(+), 17 deletions(-) create mode 100644 multimodal/diffusion/stable-diffusion/sd_2.1/README.md diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh index a3acee13f..d1f2c4b5c 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_multi.sh @@ -13,7 +13,6 @@ # License for the specific language governing permissions and limitations # under the License. - export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 @@ -24,7 +23,6 @@ echo $ENABLE_FLASH_ATTENTION_WITH_IXDNN export MODEL_PATH=${MODEL_PATH:-runwayml/stable-diffusion-v1-5} export DATASET_PATH=${DATASET_PATH:-lambdalabs/pokemon-blip-captions} - accelerate launch --config_file configs/default_config.yaml --mixed_precision="fp16" train_text_to_image.py \ --pretrained_model_name_or_path=$MODEL_PATH \ --dataset_name=$DATASET_PATH \ @@ -39,7 +37,7 @@ accelerate launch --config_file configs/default_config.yaml --mixed_precision="f --max_grad_norm=1 \ --lr_scheduler="constant" \ --lr_warmup_steps=0 \ - --output_dir="sd-pokemon-model-multi" \ + --output_dir="sd-pokemon-1.5-multi" \ --max_train_steps=900 \ --NHWC \ --dataloader_num_workers=32 \ diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh index 9ca1b0e4d..2c6a44c8c 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_1.5_single.sh @@ -13,7 +13,6 @@ # License for the specific language governing permissions and limitations # under the License. - export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 @@ -23,7 +22,6 @@ export ENABLE_FLASH_ATTENTION_WITH_IXDNN=1 export MODEL_PATH=${MODEL_PATH:-runwayml/stable-diffusion-v1-5} export DATASET_PATH=${DATASET_PATH:-lambdalabs/pokemon-blip-captions} - accelerate launch --config_file configs/single_config.yaml --mixed_precision="fp16" train_text_to_image.py \ --pretrained_model_name_or_path=$MODEL_PATH \ --dataset_name=$DATASET_PATH \ @@ -38,7 +36,7 @@ accelerate launch --config_file configs/single_config.yaml --mixed_precision="fp --max_grad_norm=1 \ --lr_scheduler="constant" \ --lr_warmup_steps=0 \ - --output_dir="sd-pokemon-model-single" \ + --output_dir="sd-pokemon-1.5-single" \ --max_train_steps=15000 \ --NHWC \ --dataloader_num_workers=32 \ diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh index 830101960..aef721bfe 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_multi.sh @@ -13,7 +13,6 @@ # License for the specific language governing permissions and limitations # under the License. - export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 @@ -26,8 +25,6 @@ export DATASET_PATH=${DATASET_PATH:-lambdalabs/pokemon-blip-captions} echo $MODEL_PATH echo $DATASET_PATH -# cd /data/yili.li/Diffusers/diffusers/examples/text_to_image - accelerate launch --config_file configs/default_config.yaml --mixed_precision="fp16" train_text_to_image.py \ --pretrained_model_name_or_path=$MODEL_PATH \ --dataset_name=$DATASET_PATH \ @@ -42,8 +39,8 @@ accelerate launch --config_file configs/default_config.yaml --mixed_precision="f --max_grad_norm=1 \ --lr_scheduler="constant" \ --lr_warmup_steps=0 \ - --output_dir="sd-pokemon-model-3" \ - --max_train_steps=100 \ + --output_dir="sd-pokemon-2.1-multi" \ + --max_train_steps=900 \ --NHWC \ --dataloader_num_workers=32 \ --apex_fused_adam diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh index 7bc92db1c..3194015f2 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_2.1_single.sh @@ -13,7 +13,6 @@ # License for the specific language governing permissions and limitations # under the License. - export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 @@ -23,7 +22,6 @@ export ENABLE_FLASH_ATTENTION_WITH_IXDNN=1 export MODEL_PATH=${MODEL_PATH:-stabilityai/stable-diffusion-2-1-base} export DATASET_PATH=${DATASET_PATH:-lambdalabs/pokemon-blip-captions} - accelerate launch --config_file configs/single_config.yaml --mixed_precision="fp16" train_text_to_image.py \ --pretrained_model_name_or_path=$MODEL_PATH \ --dataset_name=$DATASET_PATH \ @@ -38,8 +36,8 @@ accelerate launch --config_file configs/single_config.yaml --mixed_precision="fp --max_grad_norm=1 \ --lr_scheduler="constant" \ --lr_warmup_steps=0 \ - --output_dir="sd-pokemon-model-3" \ - --max_train_steps=100 \ + --output_dir="sd-pokemon-2.1-single" \ + --max_train_steps=15000 \ --NHWC \ --dataloader_num_workers=32 \ --apex_fused_adam diff --git a/multimodal/diffusion/stable-diffusion/sd_1.5/README.md b/multimodal/diffusion/stable-diffusion/sd_1.5/README.md index 5177fb91b..1bc2cd4b0 100644 --- a/multimodal/diffusion/stable-diffusion/sd_1.5/README.md +++ b/multimodal/diffusion/stable-diffusion/sd_1.5/README.md @@ -8,9 +8,13 @@ Stable Diffusion is a latent text-to-image diffusion model capable of generating You just need to run the script, and it will automatically download the required data and weights. Or you can manually download the weights and data locally. -dataset: download the lambdalabs/pokemon-blip-captions from [huggingface page](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions). +### Weights + +Download the runwayml/stable-diffusion-v1-5 from [huggingface page](https://huggingface.co/runwayml/stable-diffusion-v1-5). -weights: download the runwayml/stable-diffusion-v1-5 from [huggingface page](https://huggingface.co/runwayml/stable-diffusion-v1-5). +### Datasets + +dataset: download the lambdalabs/pokemon-blip-captions from [huggingface page](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions). ## Step 2: Installation diff --git a/multimodal/diffusion/stable-diffusion/sd_2.1/README.md b/multimodal/diffusion/stable-diffusion/sd_2.1/README.md new file mode 100644 index 000000000..523a36870 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/sd_2.1/README.md @@ -0,0 +1,57 @@ +# Stable Diffusion 2.1 + +## Model description + +Stable Diffusion is a latent text-to-image diffusion model capable of generating photo-realistic images given any text input. + +## Step 1: Preparation + +You just need to run the script, and it will automatically download the required data and weights. Or you can manually download the weights and data locally. + +### Weights + +Download the stabilityai/stable-diffusion-2-1-base from [huggingface page](https://huggingface.co/stabilityai/stable-diffusion-2-1-base). + +### Datasets + +Download the lambdalabs/pokemon-blip-captions from [huggingface page](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions). + +## Step 2: Installation + +```bash +pip3 install http://files.deepspark.org.cn:880/deepspark/add-ons/deepspeed-0.14.3+corex.20240718020249-cp310-cp310-linux_x86_64.whl +pip3 install http://files.deepspark.org.cn:880/deepspark/add-ons/diffusers-0.22.0-py3-none-any.whl +pip3 install http://files.deepspark.org.cn:880/deepspark/add-ons/transformers-4.38.1-py3-none-any.whl +pip3 install -r requirements.txt +pip3 install pillow --upgrade +``` + +## Step 3: Training + +If you have downloaded the weights and dataset, please export the environment variables like below. + +```bash +export MODEL_PATH=/path/to/sd_weights +export DATASET_PATH=/path/to/data +``` + +```bash +# Go to diffusers path +cd ${PROJ_ROOT}/multimodal/diffusion/stable-diffusion/diffusers + +# Single GPU +bash run_sd_2.1_single.sh + +# Multi GPUs +bash run_sd_2.1_multi.sh +``` + +## Results + +| Model | GPUs | ips_per_device | ips_per_gpu | +| ------ | ------- | -------------- | ----------- | +| SD 2.1 | BI-V150 | | | + +## Reference + +- [diffusers](https://github.com/huggingface/diffusers) -- Gitee From 54c0aa599a19af261bd8445f1a6170e24a4664a8 Mon Sep 17 00:00:00 2001 From: majorli Date: Fri, 26 Jul 2024 11:15:45 +0800 Subject: [PATCH 6/7] add stable diffusion xl model Signed-off-by: majorli --- .../stable-diffusion/diffusers/README.md | 71 ------------------- .../stable-diffusion/diffusers/run_sd_xl.sh | 5 +- .../diffusers/train_text_to_image.py | 18 +---- .../diffusers/train_text_to_image_sdxl.py | 18 +---- .../stable-diffusion/sd_2.1/README.md | 2 +- .../stable-diffusion/sd_xl/README.md | 56 +++++++++++++++ 6 files changed, 63 insertions(+), 107 deletions(-) delete mode 100644 multimodal/diffusion/stable-diffusion/diffusers/README.md create mode 100644 multimodal/diffusion/stable-diffusion/sd_xl/README.md diff --git a/multimodal/diffusion/stable-diffusion/diffusers/README.md b/multimodal/diffusion/stable-diffusion/diffusers/README.md deleted file mode 100644 index 56c70a0ab..000000000 --- a/multimodal/diffusion/stable-diffusion/diffusers/README.md +++ /dev/null @@ -1,71 +0,0 @@ -# Stable Diffusion - -## Model description - -Stable Diffusion is a latent text-to-image diffusion model. - -## Step 1: Installation - -- Install -```bash -wget http://files.deepspark.org.cn:880/deepspark/add-ons/deepspeed-0.14.3+corex.20240718020249-cp310-cp310-linux_x86_64.whl -wget http://files.deepspark.org.cn:880/deepspark/add-ons/diffusers-0.22.0-py3-none-any.whl -wget http://files.deepspark.org.cn:880/deepspark/add-ons/transformers-4.38.1-py3-none-any.whl -``` -```bash -pip3 install deepspeed-0.14.3+corex.20240718020249-cp310-cp310-linux_x86_64.whl -pip3 install diffusers-0.22.0-py3-none-any.whl -pip3 install transformers-4.38.1-py3-none-any.whl -pip3 install -r requirements.txt -pip3 install pillow --upgrade -``` - - -## Step 2: Preparing datasets -You just need to run the script, and it will automatically download the required data and weights. Or you can manually download the weights and data locally. - -dataset: download the lambdalabs/pokemon-blip-captions from [huggingface page](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions). - -weights: download the stabilityai/stable-diffusion-2-1-base from [huggingface page](https://huggingface.co/stabilityai/stable-diffusion-2-1-base). - -weights: download the runwayml/stable-diffusion-v1-5 from [huggingface page](https://huggingface.co/runwayml/stable-diffusion-v1-5). - - -weights: download the stabilityai/stable-diffusion-xl-base-1.0 from [huggingface page](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0). - -weights: download the madebyollin/sdxl-vae-fp16-fix from [huggingface page](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix). - -## Step 3: Training - -If you have downloaded the weights and data, please import the environment variables like below. -```bash -export MODEL_PATH=/path/to/sd_weights -export DATA_PATH=/path/to/data -export VAE_PATH=/path/to/vae_weights # only for sdxl -``` -### sd2.1 or sd1.5 -#### One GPU -```bash -bash run_sd_single.sh -``` -#### 8 GPUs -``` -bash run_sd_multi.sh -``` -### sdxl -#### 8 GPUs -```bash -bash run_sd_xl.sh -``` - -## Results -### sd2.1 - -GPUs | FPS ----- | --- -BI-V100 x8 | ips per gpu=16 -``` -## Reference - -- [diffusers](https://github.com/huggingface/diffusers) - diff --git a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh index 7f9cad5f2..467e3ca40 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh +++ b/multimodal/diffusion/stable-diffusion/diffusers/run_sd_xl.sh @@ -13,7 +13,6 @@ # License for the specific language governing permissions and limitations # under the License. - export CLIP_FLASH_ATTN=1 export USE_NHWC_GN=1 export USE_IXFORMER_GEGLU=1 @@ -42,8 +41,8 @@ accelerate launch --config_file configs/zero2_config.yaml --mixed_precision="fp1 --max_grad_norm=1 \ --lr_scheduler="constant" \ --lr_warmup_steps=0 \ - --output_dir="sd-pokemon-model-3" \ - --max_train_steps=100 \ + --output_dir="sd-pokemon-xl" \ + --max_train_steps=900 \ --dataloader_num_workers=32 \ --NHWC \ --apex_fused_adam diff --git a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py index 2efd5c4d1..9b4af3fdf 100644 --- a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py +++ b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image.py @@ -1,22 +1,8 @@ -# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. -# 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. - - #!/usr/bin/env python # coding=utf-8 # Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# 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. diff --git a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py index 27cdd493b..c1043f2fd 100755 --- a/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py +++ b/multimodal/diffusion/stable-diffusion/diffusers/train_text_to_image_sdxl.py @@ -1,22 +1,8 @@ -# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. -# 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. - - #!/usr/bin/env python # coding=utf-8 # Copyright 2024 The HuggingFace Inc. team. All rights reserved. +# Copyright (c) 2024, Shanghai Iluvatar CoreX Semiconductor Co., Ltd. +# 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. diff --git a/multimodal/diffusion/stable-diffusion/sd_2.1/README.md b/multimodal/diffusion/stable-diffusion/sd_2.1/README.md index 523a36870..11237849d 100644 --- a/multimodal/diffusion/stable-diffusion/sd_2.1/README.md +++ b/multimodal/diffusion/stable-diffusion/sd_2.1/README.md @@ -50,7 +50,7 @@ bash run_sd_2.1_multi.sh | Model | GPUs | ips_per_device | ips_per_gpu | | ------ | ------- | -------------- | ----------- | -| SD 2.1 | BI-V150 | | | +| SD 2.1 | BI-V150 | | | ## Reference diff --git a/multimodal/diffusion/stable-diffusion/sd_xl/README.md b/multimodal/diffusion/stable-diffusion/sd_xl/README.md new file mode 100644 index 000000000..72758bf47 --- /dev/null +++ b/multimodal/diffusion/stable-diffusion/sd_xl/README.md @@ -0,0 +1,56 @@ +# Stable Diffusion XL + +## Model description + +Stable Diffusion is a latent text-to-image diffusion model capable of generating photo-realistic images given any text input. + +## Step 1: Preparation + +You just need to run the script, and it will automatically download the required data and weights. Or you can manually download the weights and data locally. + +### Weights + +Download the stabilityai/stable-diffusion-xl-base-1.0 from [huggingface page](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0). + +Download the madebyollin/sdxl-vae-fp16-fix from [huggingface page](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix). + +### Datasets + +dataset: download the lambdalabs/pokemon-blip-captions from [huggingface page](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions). + +## Step 2: Installation + +```bash +pip3 install http://files.deepspark.org.cn:880/deepspark/add-ons/deepspeed-0.14.3+corex.20240718020249-cp310-cp310-linux_x86_64.whl +pip3 install http://files.deepspark.org.cn:880/deepspark/add-ons/diffusers-0.22.0-py3-none-any.whl +pip3 install http://files.deepspark.org.cn:880/deepspark/add-ons/transformers-4.38.1-py3-none-any.whl +pip3 install -r requirements.txt +pip3 install pillow --upgrade +``` + +## Step 3: Training + +If you have downloaded the weights and dataset, please export the environment variables like below. + +```bash +export MODEL_PATH=/path/to/sd_weights +export DATASET_PATH=/path/to/data +export VAE_PATH=/path/to/vae_weights +``` + +```bash +# Go to diffusers path +cd ${PROJ_ROOT}/multimodal/diffusion/stable-diffusion/diffusers + +bash run_sd_xl.sh +``` + +## Results + +| Model | GPUs | ips_per_device | ips_per_gpu | +| ----- | ------- | -------------- | ----------- | +| SD XL | BI-V150 | | | + +## Reference + +- [diffusers](https://github.com/huggingface/diffusers) -- Gitee From 8871a2310eaf215ec31654cf5726cc50324b1c88 Mon Sep 17 00:00:00 2001 From: majorli Date: Fri, 26 Jul 2024 11:18:20 +0800 Subject: [PATCH 7/7] add stable diffusion 1.5/2.1/xl models to full list Signed-off-by: majorli --- README.md | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index 780e6b864..ef2a253b8 100644 --- a/README.md +++ b/README.md @@ -395,7 +395,10 @@ DeepSparkHub甄选上百个应用算法和模型，覆盖AI和通用计算各领 [ControlNet](multimodal/diffusion/ControlNet/README.md) | PyTorch | Fill50K [DDPM](multimodal/diffusion/ddpm/README.md) | PyTorch | CIFAR-10 [L-Verse](multimodal/Language-Image_Pre-Training/L-Verse/pytorch/README.md) | PyTorch | ImageNet -[Stable Diffusion](multimodal/diffusion/stable-diffusion/training/README.md) | PyTorch | pokemon-images +[Stable Diffusion 1.4](multimodal/diffusion/stable-diffusion/training/README.md) | PyTorch | pokemon-images +[Stable Diffusion 1.5](multimodal/diffusion/stable-diffusion/sd_1.5/README.md) | PyTorch | pokemon-images +[Stable Diffusion 2.1](multimodal/diffusion/stable-diffusion/sd_2.1/README.md) | PyTorch | pokemon-images +[Stable Diffusion XL](multimodal/diffusion/stable-diffusion/sd_xl/README.md) | PyTorch | pokemon-images ### NLP -- Gitee