Skip to content

Latest commit

 

History

History
130 lines (113 loc) · 6.11 KB

data_augmentation.md

File metadata and controls

130 lines (113 loc) · 6.11 KB
Raw

Data Augmentation

AutoAugment

timm AutoAugment Help plsc TimmAutoAugment
aa Use AutoAugment policy. "v0" or "original". " + "(default: rand-m9-mstd0.5-inc1) config_str
Interpolation mode interpolation
Image size to apply resize img_size
fillcolor mean

Mixup and Cutmix

timm Mixup help plsc Mixup plsc Cutmix
mixup_alpha mixup alpha value, mixup is active if > 0. alpha
cutmix_alpha cutmix alpha value, cutmix is active if > 0. alpha
cutmix_minmax cutmix min/max image ratio, cutmix is active and uses this vs alpha if not None. does not support
prob probability of applying mixup or cutmix per batch or element Does not need, always 1.0
switch_prob probability of switching to cutmix instead of mixup when both are active prob prob
mode how to apply mixup/cutmix params (per 'batch', 'pair' (pair of elements), 'elem' (element) Does not need, always 'batch'
correct_lam apply lambda correction when cutmix bbox clipped by image borders Does not need, always True
label_smoothing apply label smoothing to the mixed target tensor epsilon epsilon
num_classes number of classes for target class_num class_num

RandomErasing

timm RandomErasing help plsc RandomErasing
probability Probability that the Random Erasing operation will be performed. EPSILON
min_area Minimum percentage of erased area wrt input image area. sl
max_area Maximum percentage of erased area wrt input image area. sh
min_aspect Minimum aspect ratio of erased area. r1
mode pixel color mode, one of 'const', 'rand', or 'pixel' mode
min_count minimum number of erasing blocks per image, area per box is scaled by count. Does not need, always 1
max_count maximum number of erasing blocks per image, area per box is scaled by count. per-image count is randomly chosen between 1 and this value. Does not need, always 1
the number of try to random erase attempt

Example configuration in yaml

DataLoader:
  Train:
    dataset:
      name: ImageNetDataset
      image_root: ./dataset/ILSVRC2012/
      cls_label_path: ./dataset/ILSVRC2012/train_list.txt
      transform_ops:
        - DecodeImage:
            to_rgb: True
            channel_first: False
        - RandCropImage:
            size: 224
            interpolation: bicubic
            backend: pil
        - RandFlipImage:
            flip_code: 1
        - TimmAutoAugment:
            config_str: rand-m9-mstd0.5-inc1
            interpolation: bicubic
            img_size: 224
            mean: [0.485, 0.456, 0.406]
        - NormalizeImage:
            scale: 1.0/255.0
            mean: [0.485, 0.456, 0.406]
            std: [0.229, 0.224, 0.225]
            order: ''
        - RandomErasing:
            EPSILON: 0.25
            sl: 0.02
            sh: 1.0/3.0
            r1: 0.3
            attempt: 10
            use_log_aspect: True
            mode: pixel
        - ToCHWImage:
      batch_transform_ops:
        - TransformOpSampler:
            Mixup:
              alpha: 0.8
              prob: 0.5
              epsilon: 0.1
              class_num: 1000
            Cutmix:
              alpha: 1.0
              prob: 0.5
              epsilon: 0.1
              class_num: 1000
    sampler:
      name: RepeatedAugSampler
      batch_size: 64
      drop_last: False
      shuffle: True
    loader:
      num_workers: 8
      use_shared_memory: True

Example code in py

transform_train = transforms.Compose([
        transforms.DecodeImage(to_rgb=True, channel_first=False),
        transforms.RandCropImage(args.input_size, interpolation="bicubic"),  # 3 is bicubic
        transforms.RandFlipImage(),
        transforms.TimmAutoAugment(config_str=args.aa, interpolation="bicubic", img_size=args.input_size, mean=[0.485, 0.456, 0.406]),
        transforms.NormalizeImage(scale=1.0/255.0, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], order='hwc'),
        transforms.RandomErasing(EPSILON=args.reprob, sl=0.02, sh=1.0/3.0, r1=0.3, attempt=10, use_log_aspect=True, mode=args.remode),
        transforms.ToCHWImage()])


batch_transform_ops = {}
batch_transform_ops['Mixup'] = {"alpha": args.mixup, "prob": args.mixup_switch_prob, "epsilon": args.smoothing, "class_num": args.nb_classes}
batch_transform_ops['Cutmix'] = {"alpha": args.cutmix, "prob": args.mixup_switch_prob, "epsilon": args.smoothing, "class_num": args.nb_classes}
mixup_fn = transforms.TransformOpSampler(**batch_transform_ops)

def mixup_collate_fn(batch):
    batch = mixup_fn(batch)
    batch = collate_fn(batch)
    return batch
    
data_loader_train = paddle.io.DataLoader(
    dataset_train, batch_sampler=sampler_train,
    num_workers=8,
    collate_fn=mixup_collate_fn,
)