qqingzheng / ai-self-training-dpo-sdxl Goto Github PK

this work means much to me and thanks! I'm not familiar with image, just curious, for --good_image_column="good_jpg"
--bad_image_column="bad_jpg", is good_jpg and bad_jpg the same picture but with goo/bad fidelity/detail or they can be completely different picture? or is bad_jpg model generated ones or can it be real picture but just with bad quality?

Hello, thank you for your open source. How did you adjust the learning rate? The paper suggests 1e-8 while you used 1e-5, which is quite a big difference. Could you share your adjustment strategy(including some other hyperparameters) ?

The training script uses a default dpo_beta=1, which is far lower than the range suggested by the paper (2000-5000). Is this intentional/oversight?

This is the code in build_dataset.py
def gen():
for key, value in results.items():
yield {
"caption": value["prompt"],
"good_jpg": os.path.join(args.image_dir, f"{key}{value['win']}.png"),
"bad_jpg": os.path.join(args.image_dir, f"{key}{value['lose']}.png"),
}

This is the code in train.py
def collate_fn(examples):
print(examples[0])
good_model_input = torch.stack(
[torch.tensor(example[args.good_image_column + "_model_input"]) for example in examples]
)
bad_model_input = torch.stack(
[torch.tensor(example[args.bad_image_column + "_model_input"]) for example in examples]
)
original_sizes = [example[args.good_image_column + "_original_sizes"] for example in examples]
crop_top_lefts = [example[args.good_image_column + "_crop_top_lefts"] for example in examples]

    prompt_embeds = torch.stack(
        [torch.tensor(example["prompt_embeds"]) for example in examples]
    )
    pooled_prompt_embeds = torch.stack(
        [torch.tensor(example["pooled_prompt_embeds"]) for example in examples]
    )

    return {
        "good_model_input": good_model_input,
        "bad_model_input": bad_model_input,
        "prompt_embeds": prompt_embeds,
        "pooled_prompt_embeds": pooled_prompt_embeds,
        "original_sizes": original_sizes,
        "crop_top_lefts": crop_top_lefts,
    }

Hello, thank you very much for your code. I have some questions.. Is the data processing method in build_dataset.py different from here? Is there any complete code that can be aligned with the data processing here?
I found that the data in build_dataset.py has three processed parameters, but there are indeed 6 parameters in train.py. Is there a complete code?