How to use IceVisionTransformAdapter properly in Object Detection task?

[IncubatorShokuhou]

I am trying to use IceVisionTransformAdapter for data enhancement in Object Detection task, but I am not sure how to use it properly. Is it equivalent to albumentations.Compose? Can someone give me a more complex example compared with the example on the doc?

[ethanwharris]

Hi @IncubatorShokuhou The IceVisionTransformAdapter is intended to let you adapt the transforms from IceVision / Albumentations with the object / keypoint detection + instance segmentation tasks. Internally, it just changes the format of your samples so that the boxes, masks, etc. are all transformed correctly.

So it's not quite the same as albumentations.Compose. A general recipe using alb.Compose would look like this:

from dataclasses import dataclass
import albumentations as alb
from icevision.tfms import A

from flash import InputTransform
from flash.core.integrations.icevision.transforms import IceVisionTransformAdapter
from flash.image import ObjectDetectionData


train_transform = A.Compose([
    A.RandomCrop(width=256, height=256),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.2),
])

transform = A.Compose([
    A.CenterCrop(width=256, height=256),
])


@dataclass
class CustomTransform(InputTransform):

    def per_sample_transform(self):
        return IceVisionTransformAdapter([transform])

    def train_per_sample_transform(self):
        return IceVisionTransformAdapter([train_transform])


datamodule = ObjectDetectionData.from_coco(
    train_folder="data/coco128/images/train2017/",
    train_ann_file="data/coco128/annotations/instances_train2017.json",
    val_split=0.1,
    batch_size=4,
    # NOTE: the below will be simplifed in 0.8.0
    train_transform=CustomTransform,
    val_transform=CustomTransform,
    test_transform=CustomTransform,
    predict_transform=CustomTransform,
    transform_kwargs={},
)

Hope that helps 😃 Let me know if you have any specific example requests

[IncubatorShokuhou]

Thank you @ethanwharris. Does train_per_sample_transform overwrite per_sample_transform in CustomTransform?