run on our dataset, but not convergence

[liangxuejingjing]

@MengyuWang826 hi~

When I ran your code on our dataset, a floorplan, we got white outputs...I don't know the reason...

And the total loss kept rising， here is my training log.
20240221_075018.log.json

How do we resolve this problem? retrain your model on our dataset? or generate a new coarse mask dataset only about the edge?
Please reply to me ASAP.

[yanrihong]

can i check how you load your custom dataset？

[liangxuejingjing]

can i check how you load your custom dataset？

I generated the coarse mask as the authors have released and then collected them into JSON format.

`base = [
'./segrefiner_lr.py'
]

object_size = 256
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=False, with_label=False, with_mask=True),
dict(type='LoadPatchData', object_size=object_size, patch_size = object_size),
dict(type='Resize', img_scale=(object_size, object_size), keep_ratio=False),
dict(type='RandomFlip', flip_ratio=0.5),
dict(type='Normalize', **img_norm_cfg),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['object_img', 'object_gt_masks', 'object_coarse_masks',
'patch_img', 'patch_gt_masks', 'patch_coarse_masks'])]

dataset_type = 'HRCollectionDataset'
data_root = '/mmdetection/SegRefiner-main/data/'
train_dataloader=dict(
samples_per_gpu=6,
workers_per_gpu=1)
data = dict(
delete=True,
train=dict(
type=dataset_type,
pipeline=train_pipeline,
data_root=data_root,
collection_datasets=['vanyi'],
collection_json=data_root + 'collection_vanyi.json'),
train_dataloader=train_dataloader,
val=dict(),
test=dict())`

[yanrihong]

Great! I just want to test the performance of segrefiner, since i use mmsegmentation before, the dataloading is a really big problem for me, thank u for sharing how you do on your custom dataset, now i gonna check the model and see if it will occur the same problem as what you happen, 3Q

[yanrihong]

我也遇到了一样的问题，到最后IOU会直接变成0

[wfantastic]

Hello, I encountered the same problem during training, the iou became 0. Do you have any good way to solve this problem?