Function build_transform

omnilmm/model/utils.py:421–462

(is_train, randaug=True, input_size=224, interpolation='bicubic', std_mode='IMAGENET_INCEPTION')

def build_transform(is_train, randaug=True, input_size=224, interpolation='bicubic', std_mode='IMAGENET_INCEPTION'):
    if std_mode == 'IMAGENET_INCEPTION':
        mean = IMAGENET_INCEPTION_MEAN
        std = IMAGENET_INCEPTION_STD
    elif std_mode == 'OPENAI_CLIP':
        mean = OPENAI_CLIP_MEAN
        std = OPENAI_CLIP_STD
    else:
        raise NotImplementedError

    if is_train:
        crop_scale = float(os.environ.get('TRAIN_CROP_SCALE', 0.9999))
        t = [
            RandomResizedCropAndInterpolation(
                input_size, scale=(crop_scale, 1.0), interpolation='bicubic'),
            # transforms.RandomHorizontalFlip(),
        ]
        if randaug and os.environ.get('TRAIN_DO_AUG', 'False') == 'True':
            print(f'@@@@@ Do random aug during training', flush=True)
            t.append(
                RandomAugment(
                    2, 7, isPIL=True,
                    augs=[
                        'Identity', 'AutoContrast', 'Equalize', 'Brightness', 'Sharpness',
                        'ShearX', 'ShearY', 'TranslateX', 'TranslateY', 'Rotate',
                    ]))
        else:
            print(f'@@@@@ Skip random aug during training', flush=True)
        t += [
            transforms.ToTensor(),
            transforms.Normalize(mean=mean, std=std),
        ]
        t = transforms.Compose(t)
    else:
        t = transforms.Compose([
            transforms.Resize((input_size, input_size),
                              interpolation=transforms.InterpolationMode.BICUBIC),
            transforms.ToTensor(),
            transforms.Normalize(mean=mean, std=std)
        ])

    return t


def img2b64(img_path):