how to use cgd_util.txt_to_dir? about clip-guided-diffusion

[afiaka87]

@githubarooski I have a colab notebook for this project in the root of the github repository.

It isn't yet public/stable.

This code uses a feature of python called "generators" which are indeed a bit tough to wrap your head around. Aside from that; a lot of this code is very "reasearch-y" because that was its original intended audience. I've done my best to strike a balance between the strange realm of research where single-letter variables are encouraged and the software engineering realm I'm used to; but I'll admit it isn't perfect.

[afiaka87]

@githubarooski Good question! that functionality got refactored a bit recently because I implemented multiple text prompts and image prompts. It is now called clean_and_combine_prompts and works on a list of prompts. If you just have one you can just wrap it in a list [my_prompt]

Since the code still isn't stable here and it sounds like you wanna use the original notebooks as well; perhaps you can just copy-paste this into a notebook?

ALPHANUMERIC_REGEX = r"[^\w\s]"
def alphanumeric_filter(s: str) -> str:
    return re.sub(ALPHANUMERIC_REGEX, "", s).replace(" ", "_")

def clean_and_combine_prompts(base_path, txts, batch_idx, max_length=255) -> str:
    clean_txt = "_".join([alphanumeric_filter(txt) for txt in txts])[:max_length]
    return os.path.join(base_path, clean_txt, f"{batch_idx:02}")

You may also just want the log_image method; which I use to create a directory with the prompt as a name, and a subdirectory for each batch in batch_size. Then an image is saved in both ./current.png and the output directory.

def log_image(image: th.Tensor, base_path: str, txts: list, current_step: int, batch_idx: int) -> str:
    dirname = clean_and_combine_prompts(base_path, txts, batch_idx)
    os.makedirs(dirname, exist_ok=True)
    stem = f"{current_step:04}"
    filename = os.path.join(dirname, f'{stem}.png')
    pil_image = tvf.to_pil_image(image.add(1).div(2).clamp(0, 1))
    pil_image.save(os.path.join(os.getcwd(), f'current.png'))
    pil_image.save(filename)
    return str(filename)

You can call it as such in Katherine's code (pseudocode)

prefix_path = "/content/outputs"
current_timestep = diffusion.num_timesteps - 1
for i, sample in enumerate(samples):
    current_timestep -= 1
    if i % save_frequency == 0 or current_timestep == -1:
        for j, img in enumerate(sample["pred_xstart"]):
            log_image(img, prefix_path, prompts, step, j)

[githubarooski]

Thanks for the response. This looks like similar features I'm working to integrate. It's very slow going because my python is entry level. Are you planning to publish a colab with your helper utilities?