data.py

import h5py
from collections import namedtuple
import torch
import torch.utils.data as data

ROOT = 'Data/'
dataset_tuple = namedtuple('dataset_tuple', ['I_tr', 'T_tr', 'L_tr',
                                             'I_db', 'T_db', 'L_db',
                                             'I_te', 'T_te', 'L_te'])
paths = {
    'flickr': ROOT + 'mir_cnn_twt.mat',
    'nuswide': ROOT + 'nus_cnn_twt.mat',
    'coco': ROOT + 'coco_cnn_twt_2014.mat'
}

# Check in 2022-1-3
def load_data(DATANAME):
    data = h5py.File(paths[DATANAME], 'r')

    I_tr = data['I_tr'][:].T
    T_tr = data['T_tr'][:].T
    L_tr = data['L_tr'][:].T

    I_db = data['I_db'][:].T
    T_db = data['T_db'][:].T
    L_db = data['L_db'][:].T

    I_te = data['I_te'][:].T
    T_te = data['T_te'][:].T
    L_te = data['L_te'][:].T

    return dataset_tuple(I_tr=I_tr, T_tr=T_tr, L_tr=L_tr,
                         I_db=I_db, T_db=T_db, L_db=L_db,
                         I_te=I_te, T_te=T_te, L_te=L_te)

# Check in 2022-1-3
class my_dataset(data.Dataset):
    def __init__(self, img_feature, txt_feature, label, **kwargs):
        self.img_feature = torch.Tensor(img_feature)
        self.txt_feature = torch.Tensor(txt_feature)
        self.label = torch.Tensor(label)
        self.length = self.img_feature.size(0)
        self.training = False
        if 'B_tr' in kwargs.keys():
            print('Convert the dataset state to "Training"')
            self.training = True
            self.B_tr = kwargs['B_tr']

    def __getitem__(self, item):
        if self.training:
            return item, self.img_feature[item, :], self.txt_feature[item, :], self.label[item, :], self.B_tr[item, :]
        else:
            return item, self.img_feature[item, :], self.txt_feature[item, :], self.label[item, :]

    def __len__(self):
        return self.length

# Check in 2022-1-3
class l2h_dataset(data.Dataset):
    def __init__(self, label_index_sequence, label):
        self.label_index_sequence = label_index_sequence
        self.label = label
        self.length = self.label.size(0)

    def __getitem__(self, item):
        return self.label_index_sequence[item, :], self.label[item, :]

    def __len__(self):
        return self.length

if __name__ == '__main__':
    dset = load_data('nuswide')