Handle "store_true" bool flags more elegantly

[nateraw]

Right now, bool args are converted and passed to the script like --do_the_thing True if you passed do_the_thing=True as a kwarg to run. Sometimes, like in the case of action='store_true', argparse doesn't like this. It also doesn't like --do_the_thing=True sometimes as well. Note I ran into this while running this script from diffusers.

---

Reproducible Example

You'll need a local instance data dir with some images, and may want to update the instance prompt/validation prompt.

fuego_run.py

from pathlib import Path

import fuego


base_model = 'stabilityai/stable-diffusion-2-1-base'
instance_data_dir = './rexxliferaj_images'
output_dir = './outputs'
instance_prompt = "a photo of sks rexxliferaj"
resolution = 512 # 768
gradient_accumulation = 1
learning_rate = 1e-4
n_steps = 1000
checkpointing_steps = 100
validation_prompt = "a photo of sks rexxliferaj at the Eiffel Tower"
validation_epochs = 100
seed = 42
fp16 = True
use_8bit_adam = True
use_wandb=False

space_url, dataset_url = fuego.run(
    script=Path(__file__).parent / "run_lora_finetuning.py",
    requirements_file=Path(__file__).parent / "requirements.txt",
    delete_space_on_completion=False,
    space_hardware="t4-small",
    downgrade_hardware_on_completion=False,
    # Kwargs
    pretrained_model_name_or_path=base_model,
    instance_data_dir=instance_data_dir,
    output_dir=output_dir,
    instance_prompt=instance_prompt,
    resolution=resolution,
    train_batch_size=1,
    gradient_accumulation_steps=gradient_accumulation,
    learning_rate=learning_rate,
    lr_scheduler="constant",
    lr_warmup_steps=0,
    max_train_steps=n_steps,
    checkpointing_steps=checkpointing_steps,
    validation_prompt=validation_prompt,
    validation_epochs=validation_epochs,
    seed=seed,
    mixed_precision="fp16" if fp16 else None,
    use_8bit_adam=use_8bit_adam,
    report_to="wandb" if use_wandb else "tensorboard",
)

print(f"Space: {space_url}")
print(f"Dataset: {dataset_url}")

requirements.txt

accelerate==0.15.0
bitsandbytes==0.36.0.post2
datasets==2.8.0
git+https://github.com/huggingface/diffusers@31be42209ddfdb69d9640a777b32e9b5c6259bf0#egg=diffusers
ftfy==6.1.1
gradio==3.16.2
huggingface-hub==0.12.0
Pillow==9.4.0
python-slugify==7.0.0
tensorboard==2.11.2
torch==1.13.1
torchvision==0.14.1
transformers==4.26.0
wandb==0.13.9