Config file of few supervision.

[Huiimin5]

It is a very nice work.
Thank you so much for your contribution.
Is it possible to release scripts for few-supervision on PASCAL VOC 2012, too?

[charlesCXK]

Hi, the config file of few supervision is here.
(1) You could start from this config. Please note that we only show the difference in the below. Other files such as network.py, train.py need no change.
(2) Following the official code of PseudoSeg, we view all the 10582 images in the VOC Aug set as the unlabeled set. The image IDs are recorded in the file train_aug.txt. To generate this file, you could simply merge two files: train_aug_unlabeled_1-2.txt and train_aug_labeled_1-2.txt.
(3) We train 30 epochs for few supervision. The cps_weight is 0.1 here, but if you use 0.5, you would get significantly higher performance than the number in our paper (1%~4% IoU gain for 732, 366, 183 labeled settings).

C.dataset_path = osp.join(C.volna, 'DATA/pascal_voc')
C.img_root_folder = C.dataset_path
C.gt_root_folder = C.dataset_path
C.labeled_ratio = 2
C.train_source = osp.join(C.dataset_path,'subset_train_aug/pseudoseg_labeled_1-{}.txt'.format(C.labeled_ratio))
C.unsup_source = osp.join(C.dataset_path, 'train_aug.txt')
C.eval_source = osp.join(C.dataset_path, "val.txt")
C.is_test = False

C.num_train_imgs = 1464 // C.labeled_ratio
C.num_eval_imgs = 1449
C.num_unsup_imgs = 10582     # unsupervised samples

C.max_samples = max(C.num_train_imgs, C.num_unsup_imgs)


"""Train Config"""
if os.getenv('learning_rate'):
    C.lr = float(os.environ['learning_rate'])
else:
    C.lr = 0.001

if os.getenv('batch_size'):
    C.batch_size = int(os.environ['batch_size'])
else:
    C.batch_size = 8

C.nepochs = 30
C.niters_per_epoch = int(math.ceil(C.max_samples * 1.0 // C.batch_size))
C.num_workers = 8
C.train_scale_array = [0.5, 0.75, 1, 1.5, 1.75, 2.0]
C.warm_up_epoch = 0

C.cps_weight = 0.1
C.cutmix_mask_prop_range=(0.25, 0.5)
C.cutmix_boxmask_n_boxes=3
C.cutmix_boxmask_fixed_aspect_ratio=False
C.cutmix_boxmask_by_size=False
C.cutmix_boxmask_outside_bounds=False
C.cutmix_boxmask_no_invert=False

[Huiimin5]

Hi, Thank you so much for your reply.
But changing C.train_source to subset_train_aug/pseudoseg_labeled_1-{}.txt only leads me dataloader errors.

I think the problem is with file format because when I check overlap ids between subset_train_aug/pseudoseg_labeled_1-2.txt and train_aug.txt with uniq command I can only find one but there are definitely identical ids across these two files.

Could you find overlapped ids using uniq in your environment?

[charlesCXK]

Could you provide more details?

[Huiimin5]

It seems the problem is caused by missing directory: TorchSemiSeg/DATA/pascal_voc/pseudoseg_labeled_1-2.
Could you please provide that?

[charlesCXK]

Hi, there is no need to create such a directory. All the images could be found in the VOC Aug set.