In order to run the scripts, make sure that all of thes eenvironment variables are provided:
export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0" && export VAE_NAME="madebyollin/sdxl-vae-fp16-fix" && export DATASET_NAME="lambdalabs/pokemon-blip-captions"
export HF_DATASETS_CACHE=/future/u/suppakit/hf-cache export WANDB_DIR=future/u/suppakit/wandblogs export HF_HOME=/future/u/suppakit/hf-home export TRANSFORMERS_CACHE=/future/u/suppakit/transformers-cache export WANDB_CONFIG_DIR=/future/u/suppakit/wandb_config
To run training:
accelerate launch train_text_to_image_lora_sdxl.py
--pretrained_model_name_or_path=$MODEL_NAME
--pretrained_vae_model_name_or_path=$VAE_NAME
--dataset_name=$DATASET_NAME
--resolution=1024
--center_crop
--train_batch_size=1
--gradient_accumulation_steps=4
--gradient_checkpointing
--max_train_steps=10000
--use_8bit_adam
--learning_rate=1e-3
--lr_scheduler="constant"
--lr_warmup_steps=0
--mixed_precision="fp16"
--report_to="wandb"
--validation_prompt="a cute Sundar Pichai creature"
--validation_epochs 5
--checkpointing_steps=5000
--output_dir="./model-output"
--rank 256
--cache_dir data-cache