GroupViT/datasets/build.py

import logging
import torch
import torchvision.transforms as T
from torch.utils.data import DataLoader

# from datasets.sampler import RandomIdentitySampler
# from datasets.sampler_ddp import RandomIdentitySampler_DDP
from utils.comm import get_world_size
from .cuhkpedes import CUHKPEDES
from .bases import ImageDataset, TextDataset, ImageTextDataset, ImageTextMLMDataset

# __factory = {'CUHK-PEDES': CUHKPEDES, 'ICFG-PEDES': ICFGPEDES, 'RSTPReid': RSTPReid}
__factory = {'CUHK-PEDES': CUHKPEDES}

def build_transforms(img_size=(384, 128), aug=False, is_train=True):
    height, width = img_size

    mean = [0.48145466, 0.4578275, 0.40821073]
    std = [0.26862954, 0.26130258, 0.27577711]

    if not is_train:
        transform = T.Compose([
            T.Resize((height, width)),
            T.ToTensor(),
            T.Normalize(mean=mean, std=std),
        ])
        return transform

    # transform for training
    if aug:
        transform = T.Compose([
            T.Resize((height, width)),
            T.RandomHorizontalFlip(0.5),
            T.Pad(10),
            T.RandomCrop((height, width)),
            T.ToTensor(),
            T.Normalize(mean=mean, std=std),
            T.RandomErasing(scale=(0.02, 0.4), value=mean),
        ])
    else:
        transform = T.Compose([
            T.Resize((height, width)),
            T.RandomHorizontalFlip(0.5),
            T.ToTensor(),
            T.Normalize(mean=mean, std=std),
        ])
    return transform

def collate(batch):
    keys = set([key for b in batch for key in b.keys()])
    # turn list of dicts data structure to dict of lists data structure
    dict_batch = {k: [dic[k] if k in dic else None for dic in batch] for k in keys}

    batch_tensor_dict = {}
    for k, v in dict_batch.items():
        if isinstance(v[0], int):
            batch_tensor_dict.update({k: torch.tensor(v)})
        elif torch.is_tensor(v[0]):
             batch_tensor_dict.update({k: torch.stack(v)})
        else:
            raise TypeError(f"Unexpect data type: {type(v[0])} in a batch.")

    return batch_tensor_dict

def build_dataloader(args, tranforms=None):
    logger = logging.getLogger("IRRA.dataset")

    num_workers = args.data.num_workers
    dataset = __factory[args.data.dataset.meta.cuhkpedes_val.name](root=args.data.dataset.meta.cuhkpedes_val.raw_path)
    num_classes = len(dataset.train_id_container)

    val_transforms = build_transforms(img_size=(args.data.img_aug.img_size * 3, args.data.img_aug.img_size),
                                    is_train=False)

    # use test set as validate set
    ds = dataset.val
    val_img_set = ImageDataset(ds['image_pids'], ds['img_paths'],
                                val_transforms)
    val_txt_set = TextDataset(ds['caption_pids'],
                                ds['captions'],
                                text_length=args.data.text_aug.max_seq_len)

    val_img_loader = DataLoader(val_img_set,
                                batch_size=args.data.batch_size,
                                shuffle=False,
                                num_workers=num_workers)
    val_txt_loader = DataLoader(val_txt_set,
                                batch_size=args.data.batch_size,
                                shuffle=False,
                                num_workers=num_workers)

    return val_img_loader, val_txt_loader, num_classes