I wanna repeat your algorithm for search policies but don't understand some moments in paper. about fast-autoaugment

[ildoonet]

1. Yes, search algorithm run only with 6k pictures. Below is the partial code snippet I used.

    elif dataset == 'reduced_imagenet':
        # randomly chosen indices
        idx120 = [904, 385, 759, 884, 784, 844, 132, 214, 990, 786, 979, 582, 104, 288, 697, 480, 66, 943, 308, 282, 118, 926, 882, 478, 133, 884, 570, 964, 825, 656, 661, 289, 385, 448, 705, 609, 955, 5, 703, 713, 695, 811, 958, 147, 6, 3, 59, 354, 315, 514, 741, 525, 685, 673, 657, 267, 575, 501, 30, 455, 905, 860, 355, 911, 24, 708, 346, 195, 660, 528, 330, 511, 439, 150, 988, 940, 236, 803, 741, 295, 111, 520, 856, 248, 203, 147, 625, 589, 708, 201, 712, 630, 630, 367, 273, 931, 960, 274, 112, 239, 463, 355, 955, 525, 404, 59, 981, 725, 90, 782, 604, 323, 418, 35, 95, 97, 193, 690, 869, 172]
        total_trainset = torchvision.datasets.ImageFolder(root=os.path.join(dataroot, 'imagenet/train'), transform=transform_train)
        testset = torchvision.datasets.ImageFolder(root=os.path.join(dataroot, 'imagenet/val'), transform=transform_test)

        # compatibility
        total_trainset.train_labels = [lb for _, lb in total_trainset.samples]

        sss = StratifiedShuffleSplit(n_splits=1, test_size=len(total_trainset) - 500000, random_state=0)  # 4000 trainset
        sss = sss.split(list(range(len(total_trainset))), total_trainset.train_labels)
        train_idx, valid_idx = next(sss)

        # filter out
        train_idx = list(filter(lambda x: total_trainset.train_labels[x] in idx120, train_idx))
        valid_idx = list(filter(lambda x: total_trainset.train_labels[x] in idx120, valid_idx))
        test_idx = list(filter(lambda x: testset.samples[x][1] in idx120, range(len(testset))))

        train_labels = [idx120.index(total_trainset.train_labels[idx]) for idx in train_idx]
        for idx in range(len(total_trainset.samples)):
            if total_trainset.samples[idx][1] not in idx120:
                continue
            total_trainset.samples[idx] = (total_trainset.samples[idx][0], idx120.index(total_trainset.samples[idx][1]))
        total_trainset = Subset(total_trainset, train_idx)
        total_trainset.train_labels = train_labels

        for idx in range(len(testset.samples)):
            if testset.samples[idx][1] not in idx120:
                continue
            testset.samples[idx] = (testset.samples[idx][0], idx120.index(testset.samples[idx][1]))
        testset = Subset(testset, test_idx)
        print('reduced_imagenet train=', len(total_trainset))

2. I didn't understand fully. Please elaborate more.

3. Yes, TTA(Test Time Augmentation) performance was used for evaluate each policies and we select top-k. In our ablation study, the number of policies you select(we used 10 for each run) was not that important. You can reduce it like 5. Also, I believe that finding many policies in a efficient manner is a key reason for our performance.

4. We just choose best top-k policies for each cv set(=5) and we combine them.