diff --git a/docs/source/en/training/text2image.md b/docs/source/en/training/text2image.md index 9585141c1468..6aa39572ab34 100644 --- a/docs/source/en/training/text2image.md +++ b/docs/source/en/training/text2image.md @@ -281,3 +281,8 @@ image.save("yoda-pokemon.png") * We support fine-tuning the UNet shipped in [Stable Diffusion XL](https://huggingface.co/papers/2307.01952) via the `train_text_to_image_sdxl.py` script. Please refer to the docs [here](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/README_sdxl.md). * We also support fine-tuning of the UNet and Text Encoder shipped in [Stable Diffusion XL](https://huggingface.co/papers/2307.01952) with LoRA via the `train_text_to_image_lora_sdxl.py` script. Please refer to the docs [here](https://github.com/huggingface/diffusers/blob/main/examples/text_to_image/README_sdxl.md). + + +## Kandinsky 2.2 + +* We support fine-tuning both the decoder and prior in Kandinsky2.2 with the `train_text_to_image_prior.py` and `train_text_to_image_decoder.py` scripts. LoRA support is also included. Please refer to the docs [here](https://github.com/huggingface/diffusers/blob/main/examples/kandinsky2_2/text_to_image/README_sdxl.md). \ No newline at end of file diff --git a/examples/kandinsky2_2/text_to_image/README.md b/examples/kandinsky2_2/text_to_image/README.md new file mode 100644 index 000000000000..6e5a1835593f --- /dev/null +++ b/examples/kandinsky2_2/text_to_image/README.md @@ -0,0 +1,317 @@ +# Kandinsky2.2 text-to-image fine-tuning + +Kandinsky 2.2 includes a prior pipeline that generates image embeddings from text prompts, and a decoder pipeline that generates the output image based on the image embeddings. We provide `train_text_to_image_prior.py` and `train_text_to_image_decoder.py` scripts to show you how to fine-tune the Kandinsky prior and decoder models separately based on your own dataset. To achieve the best results, you should fine-tune **_both_** your prior and decoder models. + +___Note___: + +___This script is experimental. The script fine-tunes the whole model and often times the model overfits and runs into issues like catastrophic forgetting. It's recommended to try different hyperparameters to get the best result on your dataset.___ + + +## Running locally with PyTorch + +Before running the scripts, make sure to install the library's training dependencies: + +**Important** + +To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment: +```bash +git clone https://github.com/huggingface/diffusers +cd diffusers +pip install . +``` + +Then cd in the example folder and run +```bash +pip install -r requirements.txt +``` + +And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with: + +```bash +accelerate config +``` +For this example we want to directly store the trained LoRA embeddings on the Hub, so we need to be logged in and add the --push_to_hub flag. + +___ + +### Pokemon example + +For all our examples, we will directly store the trained weights on the Hub, so we need to be logged in and add the `--push_to_hub` flag. In order to do that, you have to be a registered user on the 🤗 Hugging Face Hub, and you'll also need to use an access token for the code to work. For more information on access tokens, please refer to the [User Access Tokens](https://huggingface.co/docs/hub/security-tokens) guide. + +Run the following command to authenticate your token + +```bash +huggingface-cli login +``` + +We also use [Weights and Biases](https://docs.wandb.ai/quickstart) logging by default, because it is really useful to monitor the training progress by regularly generating sample images during training. To install wandb, run + +```bash +pip install wandb +``` + +To disable wandb logging, remove the `--report_to=="wandb"` and `--validation_prompts="A robot pokemon, 4k photo"` flags from below examples + +#### Fine-tune decoder +
+ + +```bash +export DATASET_NAME="lambdalabs/pokemon-blip-captions" + +accelerate launch --mixed_precision="fp16" train_text_to_image_decoder.py \ + --dataset_name=$DATASET_NAME \ + --resolution=768 \ + --train_batch_size=1 \ + --gradient_accumulation_steps=4 \ + --gradient_checkpointing \ + --max_train_steps=15000 \ + --learning_rate=1e-05 \ + --max_grad_norm=1 \ + --checkpoints_total_limit=3 \ + --lr_scheduler="constant" --lr_warmup_steps=0 \ + --validation_prompts="A robot pokemon, 4k photo" \ + --report_to="wandb" \ + --push_to_hub \ + --output_dir="kandi2-decoder-pokemon-model" +``` + + + +To train on your own training files, prepare the dataset according to the format required by `datasets`. You can find the instructions for how to do that in the [ImageFolder with metadata](https://huggingface.co/docs/datasets/en/image_load#imagefolder-with-metadata) guide. +If you wish to use custom loading logic, you should modify the script and we have left pointers for that in the training script. + +```bash +export TRAIN_DIR="path_to_your_dataset" + +accelerate launch --mixed_precision="fp16" train_text_to_image_decoder.py \ + --train_data_dir=$TRAIN_DIR \ + --resolution=768 \ + --train_batch_size=1 \ + --gradient_accumulation_steps=4 \ + --gradient_checkpointing \ + --max_train_steps=15000 \ + --learning_rate=1e-05 \ + --max_grad_norm=1 \ + --checkpoints_total_limit=3 \ + --lr_scheduler="constant" --lr_warmup_steps=0 \ + --validation_prompts="A robot pokemon, 4k photo" \ + --report_to="wandb" \ + --push_to_hub \ + --output_dir="kandi22-decoder-pokemon-model" +``` + + +Once the training is finished the model will be saved in the `output_dir` specified in the command. In this example it's `kandi22-decoder-pokemon-model`. To load the fine-tuned model for inference just pass that path to `AutoPipelineForText2Image` + +```python +from diffusers import AutoPipelineForText2Image +import torch + +pipe = AutoPipelineForText2Image.from_pretrained(output_dir, torch_dtype=torch.float16) +pipe.enable_model_cpu_offload() + +prompt='A robot pokemon, 4k photo' +images = pipe(prompt=prompt).images +images[0].save("robot-pokemon.png") +``` + +Checkpoints only save the unet, so to run inference from a checkpoint, just load the unet +```python +from diffusers import AutoPipelineForText2Image, UNet2DConditionModel + +model_path = "path_to_saved_model" + +unet = UNet2DConditionModel.from_pretrained(model_path + "/checkpoint-/unet") + +pipe = AutoPipelineForText2Image.from_pretrained("kandinsky-community/kandinsky-2-2-decoder", unet=unet, torch_dtype=torch.float16) +pipe.enable_model_cpu_offload() + +image = pipe(prompt="A robot pokemon, 4k photo").images[0] +image.save("robot-pokemon.png") +``` + +#### Fine-tune prior + +You can fine-tune the Kandinsky prior model with `train_text_to_image_prior.py` script. Note that we currently do not support `--gradient_checkpointing` for prior model fine-tuning. + +
+ + +```bash +export DATASET_NAME="lambdalabs/pokemon-blip-captions" + +accelerate launch --mixed_precision="fp16" train_text_to_image_prior.py \ + --dataset_name=$DATASET_NAME \ + --resolution=768 \ + --train_batch_size=1 \ + --gradient_accumulation_steps=4 \ + --max_train_steps=15000 \ + --learning_rate=1e-05 \ + --max_grad_norm=1 \ + --checkpoints_total_limit=3 \ + --lr_scheduler="constant" --lr_warmup_steps=0 \ + --validation_prompts="A robot pokemon, 4k photo" \ + --report_to="wandb" \ + --push_to_hub \ + --output_dir="kandi2-prior-pokemon-model" +``` + + + +To perform inference with the fine-tuned prior model, you will need to first create a prior pipeline by passing the `output_dir` to `DiffusionPipeline`. Then create a `KandinskyV22CombinedPipeline` from a pretrained or fine-tuned decoder checkpoint along with all the modules of the prior pipeline you just created. + +```python +from diffusers import AutoPipelineForText2Image, DiffusionPipeline +import torch + +pipe_prior = DiffusionPipeline.from_pretrained(output_dir, torch_dtype=torch.float16) +prior_components = {"prior_" + k: v for k,v in pipe_prior.components.items()} +pipe = AutoPipelineForText2Image.from_pretrained("kandinsky-community/kandinsky-2-2-decoder", **prior_components, torch_dtype=torch.float16) + +pipe.enable_model_cpu_offload() +prompt='A robot pokemon, 4k photo' +images = pipe(prompt=prompt, negative_prompt=negative_prompt).images +images[0] +``` + +If you want to use a fine-tuned decoder checkpoint along with your fine-tuned prior checkpoint, you can simply replace the "kandinsky-community/kandinsky-2-2-decoder" in above code with your custom model repo name. Note that in order to be able to create a `KandinskyV22CombinedPipeline`, your model repository need to have a prior tag. If you have created your model repo using our training script, the prior tag is automatically included. + +#### Training with multiple GPUs + +`accelerate` allows for seamless multi-GPU training. Follow the instructions [here](https://huggingface.co/docs/accelerate/basic_tutorials/launch) +for running distributed training with `accelerate`. Here is an example command: + +```bash +export DATASET_NAME="lambdalabs/pokemon-blip-captions" + +accelerate launch --mixed_precision="fp16" --multi_gpu train_text_to_image_decoder.py \ + --dataset_name=$DATASET_NAME \ + --resolution=768 \ + --train_batch_size=1 \ + --gradient_accumulation_steps=4 \ + --gradient_checkpointing \ + --max_train_steps=15000 \ + --learning_rate=1e-05 \ + --max_grad_norm=1 \ + --checkpoints_total_limit=3 \ + --lr_scheduler="constant" --lr_warmup_steps=0 \ + --validation_prompts="A robot pokemon, 4k photo" \ + --report_to="wandb" \ + --push_to_hub \ + --output_dir="kandi2-decoder-pokemon-model" +``` + + +#### Training with Min-SNR weighting + +We support training with the Min-SNR weighting strategy proposed in [Efficient Diffusion Training via Min-SNR Weighting Strategy](https://arxiv.org/abs/2303.09556) which helps achieve faster convergence +by rebalancing the loss. Enable the `--snr_gamma` argument and set it to the recommended +value of 5.0. + + +## Training with LoRA + +Low-Rank Adaption of Large Language Models was first introduced by Microsoft in [LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685) by *Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen*. + +In a nutshell, LoRA allows adapting pretrained models by adding pairs of rank-decomposition matrices to existing weights and **only** training those newly added weights. This has a couple of advantages: + +- Previous pretrained weights are kept frozen so that model is not prone to [catastrophic forgetting](https://www.pnas.org/doi/10.1073/pnas.1611835114). +- Rank-decomposition matrices have significantly fewer parameters than original model, which means that trained LoRA weights are easily portable. +- LoRA attention layers allow to control to which extent the model is adapted toward new training images via a `scale` parameter. + +[cloneofsimo](https://github.com/cloneofsimo) was the first to try out LoRA training for Stable Diffusion in the popular [lora](https://github.com/cloneofsimo/lora) GitHub repository. + +With LoRA, it's possible to fine-tune Kandinsky 2.2 on a custom image-caption pair dataset +on consumer GPUs like Tesla T4, Tesla V100. + +### Training + +First, you need to set up your development environment as explained in the [installation](#installing-the-dependencies). Make sure to set the `MODEL_NAME` and `DATASET_NAME` environment variables. Here, we will use [Kandinsky 2.2](https://huggingface.co/kandinsky-community/kandinsky-2-2-decoder) and the [Pokemons dataset](https://huggingface.co/datasets/lambdalabs/pokemon-blip-captions). + + +#### Train decoder + +```bash +export DATASET_NAME="lambdalabs/pokemon-blip-captions" + +accelerate launch --mixed_precision="fp16" train_text_to_image_decoder_lora.py \ + --dataset_name=$DATASET_NAME --caption_column="text" \ + --resolution=768 \ + --train_batch_size=1 \ + --num_train_epochs=100 --checkpointing_steps=5000 \ + --learning_rate=1e-04 --lr_scheduler="constant" --lr_warmup_steps=0 \ + --seed=42 \ + --rank=4 \ + --gradient_checkpointing \ + --output_dir="kandi22-decoder-pokemon-lora" \ + --validation_prompt="cute dragon creature" --report_to="wandb" \ + --push_to_hub \ +``` + +#### Train prior + +```bash +export DATASET_NAME="lambdalabs/pokemon-blip-captions" + +accelerate launch --mixed_precision="fp16" train_text_to_image_prior_lora.py \ + --dataset_name=$DATASET_NAME --caption_column="text" \ + --resolution=768 \ + --train_batch_size=1 \ + --num_train_epochs=100 --checkpointing_steps=5000 \ + --learning_rate=1e-04 --lr_scheduler="constant" --lr_warmup_steps=0 \ + --seed=42 \ + --rank=4 \ + --output_dir="kandi22-prior-pokemon-lora" \ + --validation_prompt="cute dragon creature" --report_to="wandb" \ + --push_to_hub \ +``` + +**___Note: When using LoRA we can use a much higher learning rate compared to non-LoRA fine-tuning. Here we use *1e-4* instead of the usual *1e-5*. Also, by using LoRA, it's possible to run above scripts in consumer GPUs like T4 or V100.___** + + +### Inference + +#### Inference using fine-tuned LoRA checkpoint for decoder + +Once you have trained a Kandinsky decoder model using the above command, inference can be done with the `AutoPipelineForText2Image` after loading the trained LoRA weights. You need to pass the `output_dir` for loading the LoRA weights, which in this case is `kandi22-decoder-pokemon-lora`. + + +```python +from diffusers import AutoPipelineForText2Image +import torch + +pipe = AutoPipelineForText2Image.from_pretrained("kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16) +pipe.unet.load_attn_procs(output_dir) +pipe.enable_model_cpu_offload() + +prompt='A robot pokemon, 4k photo' +image = pipe(prompt=prompt).images[0] +image.save("robot_pokemon.png") +``` + +#### Inference using fine-tuned LoRA checkpoint for prior + +```python +from diffusers import AutoPipelineForText2Image +import torch + +pipe = AutoPipelineForText2Image.from_pretrained("kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16) +pipe.prior_prior.load_attn_procs(output_dir) +pipe.enable_model_cpu_offload() + +prompt='A robot pokemon, 4k photo' +image = pipe(prompt=prompt).images[0] +image.save("robot_pokemon.png") +image +``` + +### Training with xFormers: + +You can enable memory efficient attention by [installing xFormers](https://huggingface.co/docs/diffusers/main/en/optimization/xformers) and passing the `--enable_xformers_memory_efficient_attention` argument to the script. + +xFormers training is not available for fine-tuning the prior model. + +**Note**: + +According to [this issue](https://github.com/huggingface/diffusers/issues/2234#issuecomment-1416931212), xFormers `v0.0.16` cannot be used for training in some GPUs. If you observe that problem, please install a development version as indicated in that comment. \ No newline at end of file diff --git a/examples/kandinsky2_2/text_to_image/requirements.txt b/examples/kandinsky2_2/text_to_image/requirements.txt new file mode 100644 index 000000000000..31b9026efdc2 --- /dev/null +++ b/examples/kandinsky2_2/text_to_image/requirements.txt @@ -0,0 +1,7 @@ +accelerate>=0.16.0 +torchvision +transformers>=4.25.1 +datasets +ftfy +tensorboard +Jinja2 diff --git a/examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py b/examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py new file mode 100644 index 000000000000..364ed7e03189 --- /dev/null +++ b/examples/kandinsky2_2/text_to_image/train_text_to_image_decoder.py @@ -0,0 +1,936 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and + +import argparse +import logging +import math +import os +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 PIL import Image +from tqdm import tqdm +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection +from transformers.utils import ContextManagers + +import diffusers +from diffusers import AutoPipelineForText2Image, DDPMScheduler, UNet2DConditionModel, VQModel +from diffusers.optimization import get_scheduler +from diffusers.training_utils import EMAModel +from diffusers.utils import check_min_version, is_wandb_available, make_image_grid +from diffusers.utils.import_utils import is_xformers_available + + +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("0.21.0.dev0") + +logger = get_logger(__name__, log_level="INFO") + + +def save_model_card( + args, + repo_id: str, + images=None, + repo_folder=None, +): + img_str = "" + if len(images) > 0: + image_grid = make_image_grid(images, 1, len(args.validation_prompts)) + image_grid.save(os.path.join(repo_folder, "val_imgs_grid.png")) + img_str += "![val_imgs_grid](./val_imgs_grid.png)\n" + + yaml = f""" +--- +license: creativeml-openrail-m +base_model: {args.pretrained_decoder_model_name_or_path} +datasets: +- {args.dataset_name} +prior: +- {args.pretrained_prior_model_name_or_path} +tags: +- kandinsky +- text-to-image +- diffusers +inference: true +--- + """ + model_card = f""" +# Finetuning - {repo_id} + +This pipeline was finetuned from **{args.pretrained_decoder_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 = AutoPipelineForText2Image.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_card += wandb_info + + with open(os.path.join(repo_folder, "README.md"), "w") as f: + f.write(yaml + model_card) + + +def log_validation(vae, image_encoder, image_processor, unet, args, accelerator, weight_dtype, epoch): + logger.info("Running validation... ") + + pipeline = AutoPipelineForText2Image.from_pretrained( + args.pretrained_decoder_model_name_or_path, + vae=accelerator.unwrap_model(vae), + prior_image_encoder=accelerator.unwrap_model(image_encoder), + prior_image_processor=image_processor, + unet=accelerator.unwrap_model(unet), + 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.warn(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 finetuning Kandinsky 2.2.") + parser.add_argument( + "--pretrained_decoder_model_name_or_path", + type=str, + default="kandinsky-community/kandinsky-2-2-decoder", + required=False, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--pretrained_prior_model_name_or_path", + type=str, + default="kandinsky-community/kandinsky-2-2-prior", + required=False, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--dataset_name", + type=str, + default=None, + help=( + "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private," + " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem," + " or to a folder containing files that 🤗 Datasets can understand." + ), + ) + parser.add_argument( + "--dataset_config_name", + type=str, + default=None, + help="The config of the Dataset, leave as None if there's only one config.", + ) + parser.add_argument( + "--train_data_dir", + type=str, + default=None, + help=( + "A folder containing the training data. Folder contents must follow the structure described in" + " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file" + " must exist to provide the captions for the images. Ignored if `dataset_name` is specified." + ), + ) + parser.add_argument( + "--image_column", type=str, default="image", help="The column of the dataset containing an image." + ) + parser.add_argument( + "--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="kandi_2_2-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( + "--train_batch_size", type=int, default=1, 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="learning rate", + ) + 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( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument( + "--adam_weight_decay", + type=float, + default=0.0, + required=False, + 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( + "--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( + "--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" + ), + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + # Sanity checks + if args.dataset_name is None and args.train_data_dir is None: + raise ValueError("Need either a dataset name or a training folder.") + + return args + + +def main(): + args = parse_args() + logging_dir = os.path.join(args.output_dir, args.logging_dir) + accelerator_project_config = ProjectConfiguration( + total_limit=args.checkpoints_total_limit, 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 + noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="scheduler") + image_processor = CLIPImageProcessor.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="image_processor" + ) + + 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)] + + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + with ContextManagers(deepspeed_zero_init_disabled_context_manager()): + vae = VQModel.from_pretrained( + args.pretrained_decoder_model_name_or_path, subfolder="movq", torch_dtype=weight_dtype + ).eval() + image_encoder = CLIPVisionModelWithProjection.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="image_encoder", torch_dtype=weight_dtype + ).eval() + unet = UNet2DConditionModel.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="unet") + + # Freeze vae and image_encoder + vae.requires_grad_(False) + image_encoder.requires_grad_(False) + + # Create EMA for the unet. + if args.use_ema: + ema_unet = UNet2DConditionModel.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="unet") + ema_unet = EMAModel(ema_unet.parameters(), model_cls=UNet2DConditionModel, model_config=ema_unet.config) + ema_unet.to(accelerator.device) + if args.enable_xformers_memory_efficient_attention: + if is_xformers_available(): + import xformers + + xformers_version = version.parse(xformers.__version__) + if xformers_version == version.parse("0.0.16"): + logger.warn( + "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details." + ) + unet.enable_xformers_memory_efficient_attention() + else: + raise ValueError("xformers is not available. Make sure it is installed correctly") + + def compute_snr(timesteps): + """ + Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849 + """ + alphas_cumprod = noise_scheduler.alphas_cumprod + sqrt_alphas_cumprod = alphas_cumprod**0.5 + sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5 + + # Expand the tensors. + # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026 + sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None] + alpha = sqrt_alphas_cumprod.expand(timesteps.shape) + + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None] + sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape) + + # Compute SNR. + snr = (alpha / sigma) ** 2 + return snr + + # `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 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 i 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.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`" + ) + + optimizer_cls = bnb.optim.AdamW8bit + else: + optimizer_cls = torch.optim.AdamW + + optimizer = optimizer_cls( + 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, + ) + 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 + + 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)}") + + def center_crop(image): + width, height = image.size + new_size = min(width, height) + left = (width - new_size) / 2 + top = (height - new_size) / 2 + right = (width + new_size) / 2 + bottom = (height + new_size) / 2 + return image.crop((left, top, right, bottom)) + + def train_transforms(img): + img = center_crop(img) + img = img.resize((args.resolution, args.resolution), resample=Image.BICUBIC, reducing_gap=1) + img = np.array(img).astype(np.float32) / 127.5 - 1 + img = torch.from_numpy(np.transpose(img, [2, 0, 1])) + return img + + 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["clip_pixel_values"] = image_processor(images, return_tensors="pt").pixel_values + return examples + + with accelerator.main_process_first(): + if args.max_train_samples is not None: + dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples)) + # Set the training transforms + train_dataset = dataset["train"].with_transform(preprocess_train) + + def collate_fn(examples): + pixel_values = torch.stack([example["pixel_values"] for example in examples]) + pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float() + clip_pixel_values = torch.stack([example["clip_pixel_values"] for example in examples]) + clip_pixel_values = clip_pixel_values.to(memory_format=torch.contiguous_format).float() + return {"pixel_values": pixel_values, "clip_pixel_values": clip_pixel_values} + + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + shuffle=True, + collate_fn=collate_fn, + batch_size=args.train_batch_size, + num_workers=args.dataloader_num_workers, + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, + num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, + ) + unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + unet, optimizer, train_dataloader, lr_scheduler + ) + # Move image_encode and vae to gpu and cast to weight_dtype + image_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) + + # 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 + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + else: + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + global_step = int(path.split("-")[1]) + + resume_global_step = global_step * args.gradient_accumulation_steps + first_epoch = global_step // num_update_steps_per_epoch + resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps) + + progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process) + progress_bar.set_description("Steps") + for epoch in range(first_epoch, args.num_train_epochs): + unet.train() + train_loss = 0.0 + for step, batch in enumerate(train_dataloader): + # Skip steps until we reach the resumed step + if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step: + if step % args.gradient_accumulation_steps == 0: + progress_bar.update(1) + continue + + with accelerator.accumulate(unet): + # Convert images to latent space + images = batch["pixel_values"].to(weight_dtype) + clip_images = batch["clip_pixel_values"].to(weight_dtype) + latents = vae.encode(images).latents + image_embeds = image_encoder(clip_images).image_embeds + # Sample noise that we'll add to the latents + noise = torch.randn_like(latents) + 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() + + noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) + + target = noise + + # Predict the noise residual and compute loss + added_cond_kwargs = {"image_embeds": image_embeds} + + model_pred = unet(noisy_latents, timesteps, None, added_cond_kwargs=added_cond_kwargs).sample[:, :4] + + if args.snr_gamma is None: + loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") + else: + # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556. + # Since we predict the noise instead of x_0, the original formulation is slightly changed. + # This is discussed in Section 4.2 of the same paper. + snr = compute_snr(timesteps) + mse_loss_weights = ( + torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr + ) + # We first calculate the original loss. Then we mean over the non-batch dimensions and + # rebalance the sample-wise losses with their respective loss weights. + # Finally, we take the mean of the rebalanced loss. + loss = F.mse_loss(model_pred.float(), target.float(), reduction="none") + loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights + loss = loss.mean() + + # Gather the losses across all processes for logging (if we use distributed training). + avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean() + train_loss += avg_loss.item() / args.gradient_accumulation_steps + + # Backpropagate + accelerator.backward(loss) + if accelerator.sync_gradients: + 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 + + if global_step % args.checkpointing_steps == 0: + if accelerator.is_main_process: + # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` + if args.checkpoints_total_limit is not None: + checkpoints = os.listdir(args.output_dir) + checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] + checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) + + # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints + if len(checkpoints) >= args.checkpoints_total_limit: + num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 + removing_checkpoints = checkpoints[0:num_to_remove] + + logger.info( + f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" + ) + logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") + + for removing_checkpoint in removing_checkpoints: + removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) + shutil.rmtree(removing_checkpoint) + + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_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, + image_encoder, + image_processor, + 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 = accelerator.unwrap_model(unet) + if args.use_ema: + ema_unet.copy_to(unet.parameters()) + + pipeline = AutoPipelineForText2Image.from_pretrained( + args.pretrained_decoder_model_name_or_path, + vae=vae, + unet=unet, + ) + pipeline.decoder_pipe.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) + pipeline.enable_model_cpu_offload() + + 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/examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py b/examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py new file mode 100644 index 000000000000..9d96a936d0ca --- /dev/null +++ b/examples/kandinsky2_2/text_to_image/train_text_to_image_lora_decoder.py @@ -0,0 +1,820 @@ +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Fine-tuning script for Kandinsky with support for LoRA.""" + +import argparse +import logging +import math +import os +import shutil +from pathlib import Path + +import datasets +import numpy as np +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +import transformers +from accelerate import Accelerator +from accelerate.logging import get_logger +from accelerate.utils import ProjectConfiguration, set_seed +from datasets import load_dataset +from huggingface_hub import create_repo, upload_folder +from PIL import Image +from tqdm import tqdm +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection + +import diffusers +from diffusers import AutoPipelineForText2Image, DDPMScheduler, UNet2DConditionModel, VQModel +from diffusers.loaders import AttnProcsLayers +from diffusers.models.attention_processor import LoRAAttnAddedKVProcessor +from diffusers.optimization import get_scheduler +from diffusers.utils import check_min_version, is_wandb_available + + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.21.0.dev0") + +logger = get_logger(__name__, log_level="INFO") + + +def save_model_card(repo_id: str, images=None, base_model=str, dataset_name=str, repo_folder=None): + img_str = "" + for i, image in enumerate(images): + image.save(os.path.join(repo_folder, f"image_{i}.png")) + img_str += f"![img_{i}](./image_{i}.png)\n" + + yaml = f""" +--- +license: creativeml-openrail-m +base_model: {base_model} +tags: +- kandinsky +- text-to-image +- diffusers +- lora +inference: true +--- + """ + model_card = f""" +# LoRA text2image fine-tuning - {repo_id} +These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n +{img_str} +""" + with open(os.path.join(repo_folder, "README.md"), "w") as f: + f.write(yaml + model_card) + + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of finetuning Kandinsky 2.2 with LoRA.") + parser.add_argument( + "--pretrained_decoder_model_name_or_path", + type=str, + default="kandinsky-community/kandinsky-2-2-decoder", + required=False, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--pretrained_prior_model_name_or_path", + type=str, + default="kandinsky-community/kandinsky-2-2-prior", + required=False, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--dataset_name", + type=str, + default=None, + help=( + "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private," + " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem," + " or to a folder containing files that 🤗 Datasets can understand." + ), + ) + parser.add_argument( + "--dataset_config_name", + type=str, + default=None, + help="The config of the Dataset, leave as None if there's only one config.", + ) + parser.add_argument( + "--train_data_dir", + type=str, + default=None, + help=( + "A folder containing the training data. Folder contents must follow the structure described in" + " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file" + " must exist to provide the captions for the images. Ignored if `dataset_name` is specified." + ), + ) + parser.add_argument( + "--image_column", type=str, default="image", help="The column of the dataset containing an image." + ) + parser.add_argument( + "--validation_prompt", type=str, default=None, help="A prompt that is sampled during training for inference." + ) + parser.add_argument( + "--num_validation_images", + type=int, + default=4, + help="Number of images that should be generated during validation with `validation_prompt`.", + ) + parser.add_argument( + "--validation_epochs", + type=int, + default=1, + help=( + "Run fine-tuning validation every X epochs. The validation process consists of running the prompt" + " `args.validation_prompt` multiple times: `args.num_validation_images`." + ), + ) + parser.add_argument( + "--max_train_samples", + type=int, + default=None, + help=( + "For debugging purposes or quicker training, truncate the number of training examples to this " + "value if set." + ), + ) + parser.add_argument( + "--output_dir", + type=str, + default="kandi_2_2-model-finetuned-lora", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument( + "--cache_dir", + type=str, + default=None, + help="The directory where the downloaded models and datasets will be stored.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--resolution", + type=int, + default=512, + help=( + "The resolution for input images, all the images in the train/validation dataset will be resized to this" + " resolution" + ), + ) + parser.add_argument( + "--train_batch_size", type=int, default=1, 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( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--snr_gamma", + type=float, + default=None, + help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. " + "More details here: https://arxiv.org/abs/2303.09556.", + ) + parser.add_argument( + "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." + ) + parser.add_argument( + "--allow_tf32", + action="store_true", + help=( + "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see" + " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices" + ), + ) + parser.add_argument( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument("--adam_weight_decay", type=float, default=0.0, 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( + "--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( + "--rank", + type=int, + default=4, + help=("The dimension of the LoRA update matrices."), + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + # Sanity checks + if args.dataset_name is None and args.train_data_dir is None: + raise ValueError("Need either a dataset name or a training folder.") + + return args + + +def main(): + args = parse_args() + logging_dir = Path(args.output_dir, args.logging_dir) + accelerator_project_config = ProjectConfiguration( + total_limit=args.checkpoints_total_limit, 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 scheduler, tokenizer and models. + noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="scheduler") + image_processor = CLIPImageProcessor.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="image_processor" + ) + image_encoder = CLIPVisionModelWithProjection.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="image_encoder" + ) + + vae = VQModel.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="movq") + + unet = UNet2DConditionModel.from_pretrained(args.pretrained_decoder_model_name_or_path, subfolder="unet") + # freeze parameters of models to save more memory + unet.requires_grad_(False) + vae.requires_grad_(False) + + image_encoder.requires_grad_(False) + + # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision + # as these weights are only used for inference, keeping weights in full precision is not required. + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + + # Move unet, vae and text_encoder to device and cast to weight_dtype + unet.to(accelerator.device, dtype=weight_dtype) + vae.to(accelerator.device, dtype=weight_dtype) + image_encoder.to(accelerator.device, dtype=weight_dtype) + + lora_attn_procs = {} + for name in unet.attn_processors.keys(): + cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim + if name.startswith("mid_block"): + hidden_size = unet.config.block_out_channels[-1] + elif name.startswith("up_blocks"): + block_id = int(name[len("up_blocks.")]) + hidden_size = list(reversed(unet.config.block_out_channels))[block_id] + elif name.startswith("down_blocks"): + block_id = int(name[len("down_blocks.")]) + hidden_size = unet.config.block_out_channels[block_id] + + lora_attn_procs[name] = LoRAAttnAddedKVProcessor( + hidden_size=hidden_size, + cross_attention_dim=cross_attention_dim, + rank=args.rank, + ) + + unet.set_attn_processor(lora_attn_procs) + + def compute_snr(timesteps): + """ + Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849 + """ + alphas_cumprod = noise_scheduler.alphas_cumprod + sqrt_alphas_cumprod = alphas_cumprod**0.5 + sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5 + + # Expand the tensors. + # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026 + sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None] + alpha = sqrt_alphas_cumprod.expand(timesteps.shape) + + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None] + sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape) + + # Compute SNR. + snr = (alpha / sigma) ** 2 + return snr + + lora_layers = AttnProcsLayers(unet.attn_processors) + + if args.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`" + ) + + optimizer_cls = bnb.optim.AdamW8bit + else: + optimizer_cls = torch.optim.AdamW + + optimizer = optimizer_cls( + lora_layers.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, + ) + 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 + + 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)}") + + def center_crop(image): + width, height = image.size + new_size = min(width, height) + left = (width - new_size) / 2 + top = (height - new_size) / 2 + right = (width + new_size) / 2 + bottom = (height + new_size) / 2 + return image.crop((left, top, right, bottom)) + + def train_transforms(img): + img = center_crop(img) + img = img.resize((args.resolution, args.resolution), resample=Image.BICUBIC, reducing_gap=1) + img = np.array(img).astype(np.float32) / 127.5 - 1 + img = torch.from_numpy(np.transpose(img, [2, 0, 1])) + return img + + 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["clip_pixel_values"] = image_processor(images, return_tensors="pt").pixel_values + return examples + + with accelerator.main_process_first(): + if args.max_train_samples is not None: + dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples)) + # Set the training transforms + train_dataset = dataset["train"].with_transform(preprocess_train) + + def collate_fn(examples): + pixel_values = torch.stack([example["pixel_values"] for example in examples]) + pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float() + clip_pixel_values = torch.stack([example["clip_pixel_values"] for example in examples]) + clip_pixel_values = clip_pixel_values.to(memory_format=torch.contiguous_format).float() + return {"pixel_values": pixel_values, "clip_pixel_values": clip_pixel_values} + + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + shuffle=True, + collate_fn=collate_fn, + batch_size=args.train_batch_size, + num_workers=args.dataloader_num_workers, + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, + num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, + ) + # Prepare everything with our `accelerator`. + lora_layers, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + lora_layers, optimizer, train_dataloader, lr_scheduler + ) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + accelerator.init_trackers("text2image-fine-tune", config=vars(args)) + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + # Potentially load in the weights and states from a previous save + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + else: + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + global_step = int(path.split("-")[1]) + + resume_global_step = global_step * args.gradient_accumulation_steps + first_epoch = global_step // num_update_steps_per_epoch + resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps) + + # Only show the progress bar once on each machine. + progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process) + progress_bar.set_description("Steps") + + for epoch in range(first_epoch, args.num_train_epochs): + unet.train() + train_loss = 0.0 + for step, batch in enumerate(train_dataloader): + # Skip steps until we reach the resumed step + if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step: + if step % args.gradient_accumulation_steps == 0: + progress_bar.update(1) + continue + + with accelerator.accumulate(unet): + # Convert images to latent space + images = batch["pixel_values"].to(weight_dtype) + clip_images = batch["clip_pixel_values"].to(weight_dtype) + latents = vae.encode(images).latents + image_embeds = image_encoder(clip_images).image_embeds + # Sample noise that we'll add to the latents + noise = torch.randn_like(latents) + 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() + + noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps) + + target = noise + + # Predict the noise residual and compute loss + added_cond_kwargs = {"image_embeds": image_embeds} + + model_pred = unet(noisy_latents, timesteps, None, added_cond_kwargs=added_cond_kwargs).sample[:, :4] + + if args.snr_gamma is None: + loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") + else: + # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556. + # Since we predict the noise instead of x_0, the original formulation is slightly changed. + # This is discussed in Section 4.2 of the same paper. + snr = compute_snr(timesteps) + mse_loss_weights = ( + torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr + ) + # We first calculate the original loss. Then we mean over the non-batch dimensions and + # rebalance the sample-wise losses with their respective loss weights. + # Finally, we take the mean of the rebalanced loss. + loss = F.mse_loss(model_pred.float(), target.float(), reduction="none") + loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights + loss = loss.mean() + + # Gather the losses across all processes for logging (if we use distributed training). + avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean() + train_loss += avg_loss.item() / args.gradient_accumulation_steps + + # Backpropagate + accelerator.backward(loss) + if accelerator.sync_gradients: + params_to_clip = lora_layers.parameters() + accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm) + optimizer.step() + lr_scheduler.step() + optimizer.zero_grad() + + # Checks if the accelerator has performed an optimization step behind the scenes + if accelerator.sync_gradients: + progress_bar.update(1) + global_step += 1 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + + if global_step % args.checkpointing_steps == 0: + if accelerator.is_main_process: + # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` + if args.checkpoints_total_limit is not None: + checkpoints = os.listdir(args.output_dir) + checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] + checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) + + # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints + if len(checkpoints) >= args.checkpoints_total_limit: + num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 + removing_checkpoints = checkpoints[0:num_to_remove] + + logger.info( + f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" + ) + logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") + + for removing_checkpoint in removing_checkpoints: + removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) + shutil.rmtree(removing_checkpoint) + + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_prompt is not None and epoch % args.validation_epochs == 0: + logger.info( + f"Running validation... \n Generating {args.num_validation_images} images with prompt:" + f" {args.validation_prompt}." + ) + # create pipeline + pipeline = AutoPipelineForText2Image.from_pretrained( + args.pretrained_decoder_model_name_or_path, + unet=accelerator.unwrap_model(unet), + torch_dtype=weight_dtype, + ) + pipeline = pipeline.to(accelerator.device) + pipeline.set_progress_bar_config(disable=True) + + # run inference + generator = torch.Generator(device=accelerator.device) + if args.seed is not None: + generator = generator.manual_seed(args.seed) + images = [] + for _ in range(args.num_validation_images): + images.append( + pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0] + ) + + for tracker in accelerator.trackers: + if tracker.name == "tensorboard": + np_images = np.stack([np.asarray(img) for img in images]) + tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC") + if tracker.name == "wandb": + tracker.log( + { + "validation": [ + wandb.Image(image, caption=f"{i}: {args.validation_prompt}") + for i, image in enumerate(images) + ] + } + ) + + del pipeline + torch.cuda.empty_cache() + + # Save the lora layers + accelerator.wait_for_everyone() + if accelerator.is_main_process: + unet = unet.to(torch.float32) + unet.save_attn_procs(args.output_dir) + + if args.push_to_hub: + save_model_card( + repo_id, + images=images, + base_model=args.pretrained_decoder_model_name_or_path, + dataset_name=args.dataset_name, + 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_*"], + ) + + # Final inference + # Load previous pipeline + pipeline = AutoPipelineForText2Image.from_pretrained( + args.pretrained_decoder_model_name_or_path, torch_dtype=weight_dtype + ) + pipeline = pipeline.to(accelerator.device) + + # load attention processors + pipeline.unet.load_attn_procs(args.output_dir) + + # run inference + generator = torch.Generator(device=accelerator.device) + if args.seed is not None: + generator = generator.manual_seed(args.seed) + images = [] + for _ in range(args.num_validation_images): + images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]) + + if accelerator.is_main_process: + for tracker in accelerator.trackers: + if len(images) != 0: + 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) + ] + } + ) + + accelerator.end_training() + + +if __name__ == "__main__": + main() diff --git a/examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py b/examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py new file mode 100644 index 000000000000..e4aec111b8f7 --- /dev/null +++ b/examples/kandinsky2_2/text_to_image/train_text_to_image_lora_prior.py @@ -0,0 +1,850 @@ +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Fine-tuning script for Stable Diffusion for text2image with support for LoRA.""" + +import argparse +import logging +import math +import os +import random +import shutil +from pathlib import Path + +import datasets +import numpy as np +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +import transformers +from accelerate import Accelerator +from accelerate.logging import get_logger +from accelerate.utils import ProjectConfiguration, set_seed +from datasets import load_dataset +from huggingface_hub import create_repo, upload_folder +from tqdm import tqdm +from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection + +import diffusers +from diffusers import AutoPipelineForText2Image, DDPMScheduler, PriorTransformer +from diffusers.loaders import AttnProcsLayers +from diffusers.models.attention_processor import LoRAAttnProcessor +from diffusers.optimization import get_scheduler +from diffusers.utils import check_min_version, is_wandb_available + + +# Will error if the minimal version of diffusers is not installed. Remove at your own risks. +check_min_version("0.21.0.dev0") + +logger = get_logger(__name__, log_level="INFO") + + +def save_model_card(repo_id: str, images=None, base_model=str, dataset_name=str, repo_folder=None): + img_str = "" + for i, image in enumerate(images): + image.save(os.path.join(repo_folder, f"image_{i}.png")) + img_str += f"![img_{i}](./image_{i}.png)\n" + + yaml = f""" +--- +license: creativeml-openrail-m +base_model: {base_model} +tags: +- kandinsky +- text-to-image +- diffusers +- lora +inference: true +--- + """ + model_card = f""" +# LoRA text2image fine-tuning - {repo_id} +These are LoRA adaption weights for {base_model}. The weights were fine-tuned on the {dataset_name} dataset. You can find some example images in the following. \n +{img_str} +""" + with open(os.path.join(repo_folder, "README.md"), "w") as f: + f.write(yaml + model_card) + + +def parse_args(): + parser = argparse.ArgumentParser(description="Simple example of finetuning Kandinsky 2.2.") + parser.add_argument( + "--pretrained_decoder_model_name_or_path", + type=str, + default="kandinsky-community/kandinsky-2-2-decoder", + required=False, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--pretrained_prior_model_name_or_path", + type=str, + default="kandinsky-community/kandinsky-2-2-prior", + required=False, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--dataset_name", + type=str, + default=None, + help=( + "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private," + " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem," + " or to a folder containing files that 🤗 Datasets can understand." + ), + ) + parser.add_argument( + "--dataset_config_name", + type=str, + default=None, + help="The config of the Dataset, leave as None if there's only one config.", + ) + parser.add_argument( + "--train_data_dir", + type=str, + default=None, + help=( + "A folder containing the training data. Folder contents must follow the structure described in" + " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file" + " must exist to provide the captions for the images. Ignored if `dataset_name` is specified." + ), + ) + parser.add_argument( + "--image_column", type=str, default="image", help="The column of the dataset containing an image." + ) + parser.add_argument( + "--caption_column", + type=str, + default="text", + help="The column of the dataset containing a caption or a list of captions.", + ) + parser.add_argument( + "--validation_prompt", type=str, default=None, help="A prompt that is sampled during training for inference." + ) + parser.add_argument( + "--num_validation_images", + type=int, + default=4, + help="Number of images that should be generated during validation with `validation_prompt`.", + ) + parser.add_argument( + "--validation_epochs", + type=int, + default=1, + help=( + "Run fine-tuning validation every X epochs. The validation process consists of running the prompt" + " `args.validation_prompt` multiple times: `args.num_validation_images`." + ), + ) + parser.add_argument( + "--max_train_samples", + type=int, + default=None, + help=( + "For debugging purposes or quicker training, truncate the number of training examples to this " + "value if set." + ), + ) + parser.add_argument( + "--output_dir", + type=str, + default="kandi_2_2-model-finetuned-lora", + help="The output directory where the model predictions and checkpoints will be written.", + ) + parser.add_argument( + "--cache_dir", + type=str, + default=None, + help="The directory where the downloaded models and datasets will be stored.", + ) + parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.") + parser.add_argument( + "--resolution", + type=int, + default=512, + help=( + "The resolution for input images, all the images in the train/validation dataset will be resized to this" + " resolution" + ), + ) + parser.add_argument( + "--train_batch_size", type=int, default=1, 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( + "--learning_rate", + type=float, + default=1e-4, + help="learning rate", + ) + parser.add_argument( + "--lr_scheduler", + type=str, + default="constant", + help=( + 'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",' + ' "constant", "constant_with_warmup"]' + ), + ) + parser.add_argument( + "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler." + ) + parser.add_argument( + "--snr_gamma", + type=float, + default=None, + help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. " + "More details here: https://arxiv.org/abs/2303.09556.", + ) + parser.add_argument( + "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes." + ) + parser.add_argument( + "--allow_tf32", + action="store_true", + help=( + "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see" + " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices" + ), + ) + parser.add_argument( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument( + "--adam_weight_decay", + type=float, + default=0.0, + required=False, + 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( + "--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( + "--rank", + type=int, + default=4, + help=("The dimension of the LoRA update matrices."), + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + # Sanity checks + if args.dataset_name is None and args.train_data_dir is None: + raise ValueError("Need either a dataset name or a training folder.") + + return args + + +DATASET_NAME_MAPPING = { + "lambdalabs/pokemon-blip-captions": ("image", "text"), +} + + +def main(): + args = parse_args() + logging_dir = Path(args.output_dir, args.logging_dir) + + accelerator_project_config = ProjectConfiguration( + total_limit=args.checkpoints_total_limit, 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 scheduler, image_processor, tokenizer and models. + noise_scheduler = DDPMScheduler(beta_schedule="squaredcos_cap_v2", prediction_type="sample") + image_processor = CLIPImageProcessor.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="image_processor" + ) + tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_prior_model_name_or_path, subfolder="tokenizer") + image_encoder = CLIPVisionModelWithProjection.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="image_encoder" + ) + text_encoder = CLIPTextModelWithProjection.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="text_encoder" + ) + prior = PriorTransformer.from_pretrained(args.pretrained_prior_model_name_or_path, subfolder="prior") + # freeze parameters of models to save more memory + image_encoder.requires_grad_(False) + prior.requires_grad_(False) + text_encoder.requires_grad_(False) + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + + # Move image_encoder, text_encoder and prior to device and cast to weight_dtype + prior.to(accelerator.device, dtype=weight_dtype) + image_encoder.to(accelerator.device, dtype=weight_dtype) + text_encoder.to(accelerator.device, dtype=weight_dtype) + lora_attn_procs = {} + for name in prior.attn_processors.keys(): + lora_attn_procs[name] = LoRAAttnProcessor(hidden_size=2048, rank=args.rank) + + prior.set_attn_processor(lora_attn_procs) + + def compute_snr(timesteps): + """ + Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849 + """ + alphas_cumprod = noise_scheduler.alphas_cumprod + sqrt_alphas_cumprod = alphas_cumprod**0.5 + sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5 + + # Expand the tensors. + # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026 + sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None] + alpha = sqrt_alphas_cumprod.expand(timesteps.shape) + + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None] + sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape) + + # Compute SNR. + snr = (alpha / sigma) ** 2 + return snr + + lora_layers = AttnProcsLayers(prior.attn_processors) + + if args.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`" + ) + + optimizer_cls = bnb.optim.AdamW8bit + else: + optimizer_cls = torch.optim.AdamW + + optimizer = optimizer_cls( + lora_layers.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, + ) + 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" + ) + text_input_ids = inputs.input_ids + text_mask = inputs.attention_mask.bool() + return text_input_ids, text_mask + + def preprocess_train(examples): + images = [image.convert("RGB") for image in examples[image_column]] + examples["clip_pixel_values"] = image_processor(images, return_tensors="pt").pixel_values + examples["text_input_ids"], examples["text_mask"] = tokenize_captions(examples) + return examples + + with accelerator.main_process_first(): + if args.max_train_samples is not None: + dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples)) + # Set the training transforms + train_dataset = dataset["train"].with_transform(preprocess_train) + + def collate_fn(examples): + clip_pixel_values = torch.stack([example["clip_pixel_values"] for example in examples]) + clip_pixel_values = clip_pixel_values.to(memory_format=torch.contiguous_format).float() + text_input_ids = torch.stack([example["text_input_ids"] for example in examples]) + text_mask = torch.stack([example["text_mask"] for example in examples]) + return {"clip_pixel_values": clip_pixel_values, "text_input_ids": text_input_ids, "text_mask": text_mask} + + # DataLoaders creation: + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + shuffle=True, + collate_fn=collate_fn, + batch_size=args.train_batch_size, + num_workers=args.dataloader_num_workers, + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, + num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, + ) + clip_mean = prior.clip_mean.clone() + clip_std = prior.clip_std.clone() + lora_layers, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + lora_layers, optimizer, train_dataloader, lr_scheduler + ) + + # We need to recalculate our total training steps as the size of the training dataloader may have changed. + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if overrode_max_train_steps: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + # Afterwards we recalculate our number of training epochs + args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch) + + # We need to initialize the trackers we use, and also store our configuration. + # The trackers initializes automatically on the main process. + if accelerator.is_main_process: + accelerator.init_trackers("text2image-fine-tune", config=vars(args)) + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + # Potentially load in the weights and states from a previous save + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + else: + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + global_step = int(path.split("-")[1]) + + resume_global_step = global_step * args.gradient_accumulation_steps + first_epoch = global_step // num_update_steps_per_epoch + resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps) + + # Only show the progress bar once on each machine. + progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process) + progress_bar.set_description("Steps") + clip_mean = clip_mean.to(weight_dtype).to(accelerator.device) + clip_std = clip_std.to(weight_dtype).to(accelerator.device) + for epoch in range(first_epoch, args.num_train_epochs): + prior.train() + train_loss = 0.0 + for step, batch in enumerate(train_dataloader): + # Skip steps until we reach the resumed step + if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step: + if step % args.gradient_accumulation_steps == 0: + progress_bar.update(1) + continue + + with accelerator.accumulate(prior): + # Convert images to latent space + text_input_ids, text_mask, clip_images = ( + batch["text_input_ids"], + batch["text_mask"], + batch["clip_pixel_values"].to(weight_dtype), + ) + with torch.no_grad(): + text_encoder_output = text_encoder(text_input_ids) + prompt_embeds = text_encoder_output.text_embeds + text_encoder_hidden_states = text_encoder_output.last_hidden_state + + image_embeds = image_encoder(clip_images).image_embeds + # Sample noise that we'll add to the image_embeds + noise = torch.randn_like(image_embeds) + bsz = image_embeds.shape[0] + # Sample a random timestep for each image + timesteps = torch.randint( + 0, noise_scheduler.config.num_train_timesteps, (bsz,), device=image_embeds.device + ) + timesteps = timesteps.long() + image_embeds = (image_embeds - clip_mean) / clip_std + noisy_latents = noise_scheduler.add_noise(image_embeds, noise, timesteps) + + target = image_embeds + + # Predict the noise residual and compute loss + model_pred = prior( + noisy_latents, + timestep=timesteps, + proj_embedding=prompt_embeds, + encoder_hidden_states=text_encoder_hidden_states, + attention_mask=text_mask, + ).predicted_image_embedding + + if args.snr_gamma is None: + loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") + else: + # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556. + # Since we predict the noise instead of x_0, the original formulation is slightly changed. + # This is discussed in Section 4.2 of the same paper. + snr = compute_snr(timesteps) + mse_loss_weights = ( + torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr + ) + # We first calculate the original loss. Then we mean over the non-batch dimensions and + # rebalance the sample-wise losses with their respective loss weights. + # Finally, we take the mean of the rebalanced loss. + loss = F.mse_loss(model_pred.float(), target.float(), reduction="none") + loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights + loss = loss.mean() + + # Gather the losses across all processes for logging (if we use distributed training). + avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean() + train_loss += avg_loss.item() / args.gradient_accumulation_steps + + # Backpropagate + accelerator.backward(loss) + if accelerator.sync_gradients: + accelerator.clip_grad_norm_(prior.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: + progress_bar.update(1) + global_step += 1 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + + if global_step % args.checkpointing_steps == 0: + if accelerator.is_main_process: + # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` + if args.checkpoints_total_limit is not None: + checkpoints = os.listdir(args.output_dir) + checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] + checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) + + # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints + if len(checkpoints) >= args.checkpoints_total_limit: + num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 + removing_checkpoints = checkpoints[0:num_to_remove] + + logger.info( + f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" + ) + logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") + + for removing_checkpoint in removing_checkpoints: + removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) + shutil.rmtree(removing_checkpoint) + + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_prompt is not None and epoch % args.validation_epochs == 0: + logger.info( + f"Running validation... \n Generating {args.num_validation_images} images with prompt:" + f" {args.validation_prompt}." + ) + # create pipeline + pipeline = AutoPipelineForText2Image.from_pretrained( + args.pretrained_decoder_model_name_or_path, + prior_prior=accelerator.unwrap_model(prior), + torch_dtype=weight_dtype, + ) + pipeline = pipeline.to(accelerator.device) + pipeline.set_progress_bar_config(disable=True) + + # run inference + generator = torch.Generator(device=accelerator.device) + if args.seed is not None: + generator = generator.manual_seed(args.seed) + images = [] + for _ in range(args.num_validation_images): + images.append( + pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0] + ) + + for tracker in accelerator.trackers: + if tracker.name == "tensorboard": + np_images = np.stack([np.asarray(img) for img in images]) + tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC") + if tracker.name == "wandb": + tracker.log( + { + "validation": [ + wandb.Image(image, caption=f"{i}: {args.validation_prompt}") + for i, image in enumerate(images) + ] + } + ) + + del pipeline + torch.cuda.empty_cache() + + # Save the lora layers + accelerator.wait_for_everyone() + if accelerator.is_main_process: + prior = prior.to(torch.float32) + prior.save_attn_procs(args.output_dir) + + if args.push_to_hub: + save_model_card( + repo_id, + images=images, + base_model=args.pretrained_prior_model_name_or_path, + dataset_name=args.dataset_name, + 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_*"], + ) + + # Final inference + # Load previous pipeline + pipeline = AutoPipelineForText2Image.from_pretrained( + args.pretrained_decoder_model_name_or_path, torch_dtype=weight_dtype + ) + pipeline = pipeline.to(accelerator.device) + + # load attention processors + pipeline.prior_prior.load_attn_procs(args.output_dir) + + # run inference + generator = torch.Generator(device=accelerator.device) + if args.seed is not None: + generator = generator.manual_seed(args.seed) + images = [] + for _ in range(args.num_validation_images): + images.append(pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0]) + + if accelerator.is_main_process: + for tracker in accelerator.trackers: + if len(images) != 0: + 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) + ] + } + ) + + accelerator.end_training() + + +if __name__ == "__main__": + main() diff --git a/examples/kandinsky2_2/text_to_image/train_text_to_image_prior.py b/examples/kandinsky2_2/text_to_image/train_text_to_image_prior.py new file mode 100644 index 000000000000..d451e1bfe40d --- /dev/null +++ b/examples/kandinsky2_2/text_to_image/train_text_to_image_prior.py @@ -0,0 +1,966 @@ +#!/usr/bin/env python +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and + +import argparse +import 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 tqdm import tqdm +from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection +from transformers.utils import ContextManagers + +import diffusers +from diffusers import AutoPipelineForText2Image, DDPMScheduler, PriorTransformer +from diffusers.optimization import get_scheduler +from diffusers.training_utils import EMAModel +from diffusers.utils import check_min_version, is_wandb_available, make_image_grid + + +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("0.21.0.dev0") + +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=None, + repo_folder=None, +): + img_str = "" + if len(images) > 0: + image_grid = make_image_grid(images, 1, len(args.validation_prompts)) + image_grid.save(os.path.join(repo_folder, "val_imgs_grid.png")) + img_str += "![val_imgs_grid](./val_imgs_grid.png)\n" + + yaml = f""" +--- +license: creativeml-openrail-m +base_model: {args.pretrained_prior_model_name_or_path} +datasets: +- {args.dataset_name} +tags: +- kandinsky +- text-to-image +- diffusers +inference: true +--- + """ + model_card = f""" +# Finetuning - {repo_id} + +This pipeline was finetuned from **{args.pretrained_prior_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 + +pipe_prior = DiffusionPipeline.from_pretrained("{repo_id}", torch_dtype=torch.float16) +pipe_t2i = DiffusionPipeline.from_pretrained("{args.pretrained_decoder_model_name_or_path}", torch_dtype=torch.float16) +prompt = "{args.validation_prompts[0]}" +image_embeds, negative_image_embeds = pipe_prior(prompt, guidance_scale=1.0).to_tuple() +image = pipe_t2i(image_embeds=image_embeds, negative_image_embeds=negative_image_embeds).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_card += wandb_info + + with open(os.path.join(repo_folder, "README.md"), "w") as f: + f.write(yaml + model_card) + + +def log_validation( + image_encoder, image_processor, text_encoder, tokenizer, prior, args, accelerator, weight_dtype, epoch +): + logger.info("Running validation... ") + + pipeline = AutoPipelineForText2Image.from_pretrained( + args.pretrained_decoder_model_name_or_path, + prior_image_encoder=accelerator.unwrap_model(image_encoder), + prior_image_processor=image_processor, + prior_text_encoder=accelerator.unwrap_model(text_encoder), + prior_tokenizer=tokenizer, + prior_prior=accelerator.unwrap_model(prior), + torch_dtype=weight_dtype, + ) + pipeline = pipeline.to(accelerator.device) + pipeline.set_progress_bar_config(disable=True) + + 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.warn(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 finetuning Kandinsky 2.2.") + parser.add_argument( + "--pretrained_decoder_model_name_or_path", + type=str, + default="kandinsky-community/kandinsky-2-2-decoder", + required=False, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--pretrained_prior_model_name_or_path", + type=str, + default="kandinsky-community/kandinsky-2-2-prior", + required=False, + help="Path to pretrained model or model identifier from huggingface.co/models.", + ) + parser.add_argument( + "--dataset_name", + type=str, + default=None, + help=( + "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private," + " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem," + " or to a folder containing files that 🤗 Datasets can understand." + ), + ) + parser.add_argument( + "--dataset_config_name", + type=str, + default=None, + help="The config of the Dataset, leave as None if there's only one config.", + ) + parser.add_argument( + "--train_data_dir", + type=str, + default=None, + help=( + "A folder containing the training data. Folder contents must follow the structure described in" + " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file" + " must exist to provide the captions for the images. Ignored if `dataset_name` is specified." + ), + ) + parser.add_argument( + "--image_column", type=str, default="image", help="The column of the dataset containing an image." + ) + parser.add_argument( + "--caption_column", + type=str, + default="text", + help="The column of the dataset containing a caption or a list of captions.", + ) + parser.add_argument( + "--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="kandi_2_2-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( + "--train_batch_size", type=int, default=1, 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( + "--learning_rate", + type=float, + default=1e-4, + help="learning rate", + ) + 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( + "--dataloader_num_workers", + type=int, + default=0, + help=( + "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process." + ), + ) + parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.") + parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.") + parser.add_argument( + "--adam_weight_decay", + type=float, + default=0.0, + required=False, + 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( + "--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( + "--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" + ), + ) + + args = parser.parse_args() + env_local_rank = int(os.environ.get("LOCAL_RANK", -1)) + if env_local_rank != -1 and env_local_rank != args.local_rank: + args.local_rank = env_local_rank + + # Sanity checks + if args.dataset_name is None and args.train_data_dir is None: + raise ValueError("Need either a dataset name or a training folder.") + + return args + + +def main(): + args = parse_args() + logging_dir = os.path.join(args.output_dir, args.logging_dir) + accelerator_project_config = ProjectConfiguration( + total_limit=args.checkpoints_total_limit, 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, image_processor, tokenizer and models. + noise_scheduler = DDPMScheduler(beta_schedule="squaredcos_cap_v2", prediction_type="sample") + image_processor = CLIPImageProcessor.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="image_processor" + ) + tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_prior_model_name_or_path, subfolder="tokenizer") + + 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)] + + weight_dtype = torch.float32 + if accelerator.mixed_precision == "fp16": + weight_dtype = torch.float16 + elif accelerator.mixed_precision == "bf16": + weight_dtype = torch.bfloat16 + with ContextManagers(deepspeed_zero_init_disabled_context_manager()): + image_encoder = CLIPVisionModelWithProjection.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="image_encoder", torch_dtype=weight_dtype + ).eval() + text_encoder = CLIPTextModelWithProjection.from_pretrained( + args.pretrained_prior_model_name_or_path, subfolder="text_encoder", torch_dtype=weight_dtype + ).eval() + + prior = PriorTransformer.from_pretrained(args.pretrained_prior_model_name_or_path, subfolder="prior") + + # Freeze text_encoder and image_encoder + text_encoder.requires_grad_(False) + image_encoder.requires_grad_(False) + + # Create EMA for the prior. + if args.use_ema: + ema_prior = PriorTransformer.from_pretrained(args.pretrained_prior_model_name_or_path, subfolder="prior") + ema_prior = EMAModel(ema_prior.parameters(), model_cls=PriorTransformer, model_config=ema_prior.config) + ema_prior.to(accelerator.device) + + def compute_snr(timesteps): + """ + Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849 + """ + alphas_cumprod = noise_scheduler.alphas_cumprod + sqrt_alphas_cumprod = alphas_cumprod**0.5 + sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5 + + # Expand the tensors. + # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026 + sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None] + alpha = sqrt_alphas_cumprod.expand(timesteps.shape) + + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None] + sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape) + + # Compute SNR. + snr = (alpha / sigma) ** 2 + return snr + + # `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 args.use_ema: + ema_prior.save_pretrained(os.path.join(output_dir, "prior_ema")) + + for i, model in enumerate(models): + model.save_pretrained(os.path.join(output_dir, "prior")) + + # 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, "prior_ema"), PriorTransformer) + ema_prior.load_state_dict(load_model.state_dict()) + ema_prior.to(accelerator.device) + del load_model + + for i in range(len(models)): + # pop models so that they are not loaded again + model = models.pop() + + # load diffusers style into model + load_model = PriorTransformer.from_pretrained(input_dir, subfolder="prior") + 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.allow_tf32: + torch.backends.cuda.matmul.allow_tf32 = True + + if args.use_8bit_adam: + try: + import bitsandbytes as bnb + except ImportError: + raise ImportError( + "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`" + ) + + optimizer_cls = bnb.optim.AdamW8bit + else: + optimizer_cls = torch.optim.AdamW + optimizer = optimizer_cls( + prior.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, + ) + 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" + ) + text_input_ids = inputs.input_ids + text_mask = inputs.attention_mask.bool() + return text_input_ids, text_mask + + def preprocess_train(examples): + images = [image.convert("RGB") for image in examples[image_column]] + examples["clip_pixel_values"] = image_processor(images, return_tensors="pt").pixel_values + examples["text_input_ids"], examples["text_mask"] = tokenize_captions(examples) + return examples + + with accelerator.main_process_first(): + if args.max_train_samples is not None: + dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples)) + # Set the training transforms + train_dataset = dataset["train"].with_transform(preprocess_train) + + def collate_fn(examples): + clip_pixel_values = torch.stack([example["clip_pixel_values"] for example in examples]) + clip_pixel_values = clip_pixel_values.to(memory_format=torch.contiguous_format).float() + text_input_ids = torch.stack([example["text_input_ids"] for example in examples]) + text_mask = torch.stack([example["text_mask"] for example in examples]) + return {"clip_pixel_values": clip_pixel_values, "text_input_ids": text_input_ids, "text_mask": text_mask} + + # DataLoaders creation: + train_dataloader = torch.utils.data.DataLoader( + train_dataset, + shuffle=True, + collate_fn=collate_fn, + batch_size=args.train_batch_size, + num_workers=args.dataloader_num_workers, + ) + + # Scheduler and math around the number of training steps. + overrode_max_train_steps = False + num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps) + if args.max_train_steps is None: + args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch + overrode_max_train_steps = True + + lr_scheduler = get_scheduler( + args.lr_scheduler, + optimizer=optimizer, + num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps, + num_training_steps=args.max_train_steps * args.gradient_accumulation_steps, + ) + + clip_mean = prior.clip_mean.clone() + clip_std = prior.clip_std.clone() + + prior, optimizer, train_dataloader, lr_scheduler = accelerator.prepare( + prior, optimizer, train_dataloader, lr_scheduler + ) + + image_encoder.to(accelerator.device, dtype=weight_dtype) + text_encoder.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) + + # Train! + total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps + + logger.info("***** Running training *****") + logger.info(f" Num examples = {len(train_dataset)}") + logger.info(f" Num Epochs = {args.num_train_epochs}") + logger.info(f" Instantaneous batch size per device = {args.train_batch_size}") + logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}") + logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}") + logger.info(f" Total optimization steps = {args.max_train_steps}") + global_step = 0 + first_epoch = 0 + + # Potentially load in the weights and states from a previous save + if args.resume_from_checkpoint: + if args.resume_from_checkpoint != "latest": + path = os.path.basename(args.resume_from_checkpoint) + else: + # Get the most recent checkpoint + dirs = os.listdir(args.output_dir) + dirs = [d for d in dirs if d.startswith("checkpoint")] + dirs = sorted(dirs, key=lambda x: int(x.split("-")[1])) + path = dirs[-1] if len(dirs) > 0 else None + + if path is None: + accelerator.print( + f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run." + ) + args.resume_from_checkpoint = None + else: + accelerator.print(f"Resuming from checkpoint {path}") + accelerator.load_state(os.path.join(args.output_dir, path)) + global_step = int(path.split("-")[1]) + + resume_global_step = global_step * args.gradient_accumulation_steps + first_epoch = global_step // num_update_steps_per_epoch + resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps) + + # Only show the progress bar once on each machine. + progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process) + progress_bar.set_description("Steps") + + clip_mean = clip_mean.to(weight_dtype).to(accelerator.device) + clip_std = clip_std.to(weight_dtype).to(accelerator.device) + + for epoch in range(first_epoch, args.num_train_epochs): + prior.train() + train_loss = 0.0 + for step, batch in enumerate(train_dataloader): + # Skip steps until we reach the resumed step + if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step: + if step % args.gradient_accumulation_steps == 0: + progress_bar.update(1) + continue + + with accelerator.accumulate(prior): + # Convert images to latent space + text_input_ids, text_mask, clip_images = ( + batch["text_input_ids"], + batch["text_mask"], + batch["clip_pixel_values"].to(weight_dtype), + ) + with torch.no_grad(): + text_encoder_output = text_encoder(text_input_ids) + prompt_embeds = text_encoder_output.text_embeds + text_encoder_hidden_states = text_encoder_output.last_hidden_state + + image_embeds = image_encoder(clip_images).image_embeds + # Sample noise that we'll add to the image_embeds + noise = torch.randn_like(image_embeds) + bsz = image_embeds.shape[0] + # Sample a random timestep for each image + timesteps = torch.randint( + 0, noise_scheduler.config.num_train_timesteps, (bsz,), device=image_embeds.device + ) + timesteps = timesteps.long() + image_embeds = (image_embeds - clip_mean) / clip_std + noisy_latents = noise_scheduler.add_noise(image_embeds, noise, timesteps) + + target = image_embeds + + # Predict the noise residual and compute loss + model_pred = prior( + noisy_latents, + timestep=timesteps, + proj_embedding=prompt_embeds, + encoder_hidden_states=text_encoder_hidden_states, + attention_mask=text_mask, + ).predicted_image_embedding + + if args.snr_gamma is None: + loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean") + else: + # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556. + # Since we predict the noise instead of x_0, the original formulation is slightly changed. + # This is discussed in Section 4.2 of the same paper. + snr = compute_snr(timesteps) + mse_loss_weights = ( + torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr + ) + # We first calculate the original loss. Then we mean over the non-batch dimensions and + # rebalance the sample-wise losses with their respective loss weights. + # Finally, we take the mean of the rebalanced loss. + loss = F.mse_loss(model_pred.float(), target.float(), reduction="none") + loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights + loss = loss.mean() + + # Gather the losses across all processes for logging (if we use distributed training). + avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean() + train_loss += avg_loss.item() / args.gradient_accumulation_steps + + # Backpropagate + accelerator.backward(loss) + if accelerator.sync_gradients: + accelerator.clip_grad_norm_(prior.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_prior.step(prior.parameters()) + progress_bar.update(1) + global_step += 1 + accelerator.log({"train_loss": train_loss}, step=global_step) + train_loss = 0.0 + + if global_step % args.checkpointing_steps == 0: + if accelerator.is_main_process: + # _before_ saving state, check if this save would set us over the `checkpoints_total_limit` + if args.checkpoints_total_limit is not None: + checkpoints = os.listdir(args.output_dir) + checkpoints = [d for d in checkpoints if d.startswith("checkpoint")] + checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1])) + + # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints + if len(checkpoints) >= args.checkpoints_total_limit: + num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1 + removing_checkpoints = checkpoints[0:num_to_remove] + + logger.info( + f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints" + ) + logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}") + + for removing_checkpoint in removing_checkpoints: + removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint) + shutil.rmtree(removing_checkpoint) + + save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}") + accelerator.save_state(save_path) + logger.info(f"Saved state to {save_path}") + + logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]} + progress_bar.set_postfix(**logs) + + if global_step >= args.max_train_steps: + break + + if accelerator.is_main_process: + if args.validation_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_prior.store(prior.parameters()) + ema_prior.copy_to(prior.parameters()) + log_validation( + image_encoder, + image_processor, + text_encoder, + tokenizer, + prior, + args, + accelerator, + weight_dtype, + global_step, + ) + if args.use_ema: + # Switch back to the original UNet parameters. + ema_prior.restore(prior.parameters()) + + # Create the pipeline using the trained modules and save it. + accelerator.wait_for_everyone() + if accelerator.is_main_process: + prior = accelerator.unwrap_model(prior) + if args.use_ema: + ema_prior.copy_to(prior.parameters()) + + pipeline = AutoPipelineForText2Image.from_pretrained( + args.pretrained_decoder_model_name_or_path, + prior_image_encoder=image_encoder, + prior_text_encoder=text_encoder, + prior_prior=prior, + ) + pipeline.prior_pipe.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.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/src/diffusers/models/prior_transformer.py b/src/diffusers/models/prior_transformer.py index fd744e9d195e..8ada0a7c08a5 100644 --- a/src/diffusers/models/prior_transformer.py +++ b/src/diffusers/models/prior_transformer.py @@ -6,6 +6,7 @@ from torch import nn from ..configuration_utils import ConfigMixin, register_to_config +from ..loaders import UNet2DConditionLoadersMixin from ..utils import BaseOutput from .attention import BasicTransformerBlock from .attention_processor import ( @@ -32,7 +33,7 @@ class PriorTransformerOutput(BaseOutput): predicted_image_embedding: torch.FloatTensor -class PriorTransformer(ModelMixin, ConfigMixin): +class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin): """ A Prior Transformer model.