Mask_rcnn: No file for valminusminival for dataset year 2017. How is valminusminival obtained?

[JonathanCMitchell]

I am attempting to train on 2017 coco data. There is no valminusminival2017.json file.

The instances_valminusminival json files are for 2014 data only. There isn't one for 2017 data. I am guessing you can rename file instances_valminusminival2014.json to instances_valminusminival2017.json but I am not sure how this will affect your validation loss.

Does anybody know what the instances_valminusminival dataset consists of ? / How it was computed? I simply commented out L477 in order to train but I don't know if I am missing something.

[fastlater]

@JonathanCMitchell Before rename the file, better go to http://cocodataset.org/#download and download the 2017 data. The json file contains the dataset info, images info and segmentation info of each specific image. Actually, I guess that if you rename the file, I don't think the script will notice that but you will be training using the 2014 dataset, not the 2017 dataset.

[JonathanCMitchell]

Hi @fastlater. I see the validation data, but there isn't anything called valminusminival2017. Are most people simply training on the 2014 dataset? If my train is train2017 and my val has been renamed from 2014->2017 then would that mean I am training on the 2014 dataset even though my train folder is 2017?

[hochthom]

@JonathanCMitchell: Simply replace the following lines (478-488) in coco.py with:

    # Training dataset. Use the training set and 35K from the
    # validation set, as as in the Mask RCNN paper.
    dataset_train = CocoDataset()
    dataset_train.load_coco(args.dataset, "train", year=args.year, auto_download=args.download)
    dataset_train.prepare()

    # Validation dataset
    dataset_val = CocoDataset()
    dataset_val.load_coco(args.dataset, "val", year=args.year, auto_download=args.download)
    dataset_val.prepare()