Skip to content

Latest commit

 

History

History
 
 

text_to_image

README.md

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:

git clone https://github.com/huggingface/diffusers
cd diffusers
pip install .

Then cd in the example folder and run

pip install -r requirements.txt

And initialize an 🤗Accelerate environment with:

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 guide.

Run the following command to authenticate your token

huggingface-cli login

We also use Weights and Biases logging by default, because it is really useful to monitor the training progress by regularly generating sample images during training. To install wandb, run

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


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 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.

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

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

from diffusers import AutoPipelineForText2Image, UNet2DConditionModel

model_path = "path_to_saved_model"

unet = UNet2DConditionModel.from_pretrained(model_path + "/checkpoint-<N>/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.


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.

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 for running distributed training with accelerate. Here is an example command:

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 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 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.
  • 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 was the first to try out LoRA training for Stable Diffusion in the popular 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. Make sure to set the MODEL_NAME and DATASET_NAME environment variables. Here, we will use Kandinsky 2.2 and the Pokemons dataset.

Train decoder

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

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.

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

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 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, 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.