Dataloader for coco

vision
1

i need to change the datasets here to coco
https://github.com/facebookresearch/deit/blob/main/datasets.py

in method

build_dataset(is_train, args):

i tried to write this line

elif args.data_set == ‘coco’:
dataset = ??? (args.data_path, train=is_train, year=2017,
category=args.inat_category, transform=transform)

    nb_classes = 1000

but couldn’t know what is the replacement of question marks

2

Does using an out-of-the-box version of COCO work? torchvision.datasets — Torchvision master documentation

3

does data_loader need specific settings or just change name of IMNET to COCO
elif args.data_set == ‘COCO’:
if not args.use_mcloader:
root = os.path.join(args.data_path, ‘train’ if is_train else ‘val’)
dataset = datasets.ImageFolder(root, transform=transform)
else:
dataset = ClassificationDataset(
‘train’ if is_train else ‘val’,
pipeline=transform
)
nb_classes = 80

4

It depends on if the datasets are expected to produce the same kind of labels/annotations (e.g., are you doing classification or object detection)? If you are just doing classification I guess you can just use an ImageFolder provided the organization of the dataset is classes separated by folders.

5

excuse me another question do you mean the result of training the model will depend on the items that i specify detection of classifications ? if yes can you give me an example please

6

Taking another look at the source code, it looks like it is intended for classification. However, I’m not sure COCO is really a classification dataset but really a detection dataset. Do you have a link to the dataset you are using or can you describe the organization of the images/folders?

1 Like
7

I’m confusing in it … it suppose to use a model doing detection but i freeze layers to not make a classification as i need features of images not classify them … so should i use detection for coco .
datasets i have contains two folders (train and val)

8

You might want to take a look at the finetuning object detection tutorial: TorchVision Object Detection Finetuning Tutorial — PyTorch Tutorials 1.8.1+cu102 documentation as that provides a walkthrough of the scenario you are describing