code/model_vis_tools/extract_sd_self_attn.py

import logging
import os
import torch
from tqdm import tqdm
from accelerate import Accelerator
from accelerate.utils import ProjectConfiguration
import torch.utils.data as data
from torchvision.transforms import ToTensor,Compose
from diffusion_model.stable_diffusion import diffusion
from pycocotools.coco import COCO
from accelerate import DistributedDataParallelKwargs
import h5py
from tqdm import tqdm
import torch
from torch.utils.data.distributed import DistributedSampler
from torch.utils.data import Dataset
from open_clip.transform import ResizeLongest, _convert_to_rgb
from PIL import Image
from functools import reduce
import torch.nn.functional as F
os.environ["NCCL_P2P_DISABLE"] = "1"
os.environ["NCCL_IB_DISABLE"] = "1"

def save_features_to_h5(group, key, self_att):
    group.create_dataset(
        key,
        data=self_att.cpu().numpy().astype('float16'),
        dtype='float16',
        compression='lzf')

class SDNormalize(object):
    def __call__(self, img):
        return 2.0 * img - 1.0

class COCODataset(Dataset):
    def __init__(self,
                 input_filename,
                 transforms,
                 image_root):
        self.coco = COCO(input_filename)
        logging.info('Done loading data.')
        self.transforms = transforms
        self.image_root = image_root
        self.image_ids = list(self.coco.imgs.keys())
        
    def read_image(self, image_name):
        image_path = os.path.join(self.image_root, image_name)
        try:
            image = Image.open(image_path)
        except:
            print(f"Cannot load {image_path}", flush=True)
            return None
        width, height = image.size
        if width < 10 or height < 10:
            print(f"Invalid image, size {image.size}", flush=True)
            return None
        return image

    def __len__(self):
        return len(self.image_ids)

    def _load_target(self, id: int):
        return self.coco.loadAnns(self.coco.getAnnIds(id))
    
    def __getitem__(self, idx):
        image_id = self.image_ids[idx]
        image_info = self.coco.imgs[image_id]
        if 'file_name' in image_info:
            image_name = image_info['file_name']
        else:
            assert 'coco_url' in image_info
            coco_url = image_info['coco_url'].split('/')
            image_name = os.path.join(coco_url[-2], coco_url[-1])
        old_image = self.read_image(image_name)
        new_image = self.transforms(old_image)
        return new_image, image_name
    
if __name__ == '__main__':
 
    attention_layers_to_use=[-4, -6]
    sd_version='v2.1'
    time_step=45
    cache_dir='sd_self_attn_cache'
    half_precision=True
    thr=0.1
    preprocess =Compose([ResizeLongest(512, fill=0), _convert_to_rgb, ToTensor(), SDNormalize()])
    dataset = COCODataset( input_filename='/mnt/SSD8T/home/wjj/dataset/standard_coco/annotations/instances_train2017.json',
                            transforms=preprocess,
                            image_root='/mnt/SSD8T/home/wjj/dataset/standard_coco/train2017')

   

    accelerator = Accelerator(log_with="tensorboard", project_config=ProjectConfiguration(
            project_dir=".",
            logging_dir="logs"
        ),
        mixed_precision="fp16" if half_precision else "no", # bf16 for A100
        kwargs_handlers=[DistributedDataParallelKwargs(find_unused_parameters=True)] # True for full fine-tune
    )

    print("Waiting for everyone!")
    accelerator.wait_for_everyone()
    print("All processes synchronized!")

    with accelerator.main_process_first():
        if not os.path.exists(cache_dir):
            os.mkdir(cache_dir)
    rank = accelerator.process_index  
    world_size = accelerator.num_processes  
    sampler = DistributedSampler(
        dataset,
        num_replicas=world_size,
        rank=rank,
        shuffle=False,
        drop_last=False
    )

    dataloader = data.DataLoader(dataset,
                                batch_size=1,
                                sampler=sampler,
                                shuffle=False,
                                num_workers=4)
    
    cache_path = os.path.join(cache_dir,f"rank_{rank}.h5")
    teacher=diffusion(attention_layers_to_use=attention_layers_to_use,model=sd_version, time_step=time_step, device=accelerator.device,dtype=torch.float16)
    teacher.eval()

    with h5py.File(cache_path, 'w') as h5f:
        # 只有主进程显示 tqdm
        if rank == 0:
            batch_iter = tqdm(dataloader, desc=f"Rank {rank} extracting features")
        else:
            batch_iter = dataloader
        for batch in batch_iter:
            with torch.no_grad():
                images, key = batch
                images = images.to(accelerator.device, dtype=torch.float16 if half_precision else torch.float32)
                teacher.forward_wo_preprocess(images, "")
                self_att = torch.cat([teacher.attention_maps[idx] for idx in attention_layers_to_use]).float() # 10,1024
                self_att /= torch.amax(self_att, dim=-2, keepdim=True) + 1e-5
                self_att = torch.where(self_att < thr, 0, self_att)
                self_att /= self_att.sum(dim=-1, keepdim=True) + 1e-5
                self_att = reduce(torch.matmul, self_att, torch.eye(self_att.shape[-1], device=self_att.device))
                save_features_to_h5(h5f, key[0], self_att)

    print(f"Rank {rank} finished caching to {cache_path}")

# accelerate launch --num_processes 8 --mixed_precision fp16 extract_features.py