duibi_300.py

import argparse
import os
import torch
import pandas as pd
from diffusers import StableDiffusion3Pipeline

def main(args):

    pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3.5-large", torch_dtype=torch.bfloat16)
    pipe = pipe.to("cuda:2")

    df = pd.read_csv("/DATA/DATA3/zhaoyu/T2I_model_SD35/dataset_AGIQA_3K/data.csv")

    prompts= df["prompt"]
    prompts = prompts[:300]
    cnt = 1
    prompts = prompts[cnt-1:]
    print(len(prompts))
    # 设置一个可复现的随机种子
    seed = 3
    #to(accelerator.device)
    #----------------------------------------------------------------
    # for index,prompt in enumerate(prompts,start=cnt-1):
    #     generator = torch.Generator(device="cuda:0").manual_seed(seed)
    #     image_name = f"zero1_{index:06d}.png"
    #     save_path = os.path.join(save_dir0, image_name)
    #     image = pipe(
    #         prompt=prompt,
    #         num_inference_steps=28,
    #         guidance_scale=3.5,
    #         generator=generator,
    #     ).images[0]
    #     image.save(save_path)

    # # load attention processors
    pipe.load_lora_weights(args.lora_dir)

    for index,prompt in enumerate(prompts):
        generator = torch.Generator(device="cuda:2").manual_seed(seed)
        image_name = f"fintune_2epoch_{index:06d}.png"
        save_path = os.path.join(args.output_dir, image_name)
        image = pipe(
            prompt=prompt,
            num_inference_steps=28,
            guidance_scale=3.5,
            generator=generator,
        ).images[0]
        image.save(save_path)

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--lora_dir", type=str, required=True, help="Path to LoRA checkpoint directory")
    parser.add_argument("--output_dir", type=str, required=True, help="Directory to save generated images")

    args = parser.parse_args()
    main(args)