Dreambooth

NeuRIPS2022 Workshop

For NeuRIPS2002 Workshop, please use this guide to train your DreamBooth Model on Lambda Cloud.

Guide

Local install using Python virtual environment.

python -m venv .venv --prompt dreambooth && \
. .venv/bin/activate && \
pip install --upgrade pip && \
pip install -r requirements.txt && \
pip install ipykernel  && \
pip install jupyterlab  && \
python -m ipykernel install --user --name=venv
accelerate config

Run a training job:

export INPUT_DIR=<path-to-input-image-folder>
export MODEL_DIR=<path-to-output-model>
export TOKEN=aabbccddeeffgg
export LR=1e-6

accelerate launch --gpu_ids 0 \
  train_dreambooth.py \
  --config_file config.yaml \
  instance_data_dir="$INPUT_DIR" \
  output_dir="$MODEL_DIR" \
  max_train_steps=1000 \
  learning_rate="$LR" \
  instance_str="$TOKEN" \
  wandb_mode=disabled \
  with_prior_preservation=false \
  use_tf32=true

any arguments passed as <name>=<value> will be added to the config. Arguments passed as --<name> <value> before train_dreambooth.py are expected to be for huggingface accelerate, and those after are options for the script, see python train_dreambooth.py --help

Try the model with some testing prompts:

export MODEL_DIR=<path-to-output-model>
export PRED_DIR=<path-to-save-output-image>
export TOKEN=aabbccddeeffgg
export NUM_PRED=<number-of-predictions-per-prompt>

accelerate launch --gpu_ids 0 python test_dreambooth.py \
--model_path $MODEL_DIR \
--pred_path $PRED_DIR \
--num_preds $NUM_PRED \
--token $TOKEN \
--ddim