setr_naive_512x512_160k_b16_ade20k.py for binary segmentation occring Miou = 0

[rouchoo]

Hi, when i use setr_naive_512x512_160k_b16_ade20k.py for my binary segmentation task, i found Miou = 0 and Acc = 0 at the Iter(val) [1000]. (The first Iter(val) [100] can yield Miou = 23.3 and Acc = 60.7) I don't know why. Please help me. Thanks!

1. Here is my /home/rouch/Projects/pythonProject/mmsegmentation/configs/setr/setr_naive_512x512_160k_b16_ade20k.py:
   base = [
   '../base/models/setr_naive.py', '../base/datasets/leaf.py',
   '../base/default_runtime.py', '../base/schedules/schedule_160k.py'
   ]
   norm_cfg = dict(type='BN', requires_grad=True)
   model = dict(
   pretrained=None,
   backbone=dict(
   img_size=(512, 512),
   drop_rate=0.,
   init_cfg=dict(
   type='Pretrained', checkpoint='/home/rouch/Projects/pythonProject/mmsegmentation/pretrain/vit_large_p16.pth')),
   decode_head=dict(num_classes=2),
   auxiliary_head=[
   dict(
   type='SETRUPHead',
   in_channels=1024,
   channels=256,
   in_index=0,
   num_classes=2,
   out_channels=2,
   dropout_ratio=0,
   norm_cfg=norm_cfg,
   act_cfg=dict(type='ReLU'),
   num_convs=2,
   kernel_size=1,
   align_corners=False,
   loss_decode=dict(
   type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
   dict(
   type='SETRUPHead',
   in_channels=1024,
   channels=256,
   in_index=1,
   num_classes=2,
   out_channels=2,
   dropout_ratio=0,
   norm_cfg=norm_cfg,
   act_cfg=dict(type='ReLU'),
   num_convs=2,
   kernel_size=1,
   align_corners=False,
   loss_decode=dict(
   type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4)),
   dict(
   type='SETRUPHead',
   in_channels=1024,
   channels=256,
   in_index=2,
   num_classes=2,
   out_channels = 2,
   dropout_ratio=0,
   norm_cfg=norm_cfg,
   act_cfg=dict(type='ReLU'),
   num_convs=2,
   kernel_size=1,
   align_corners=False,
   loss_decode=dict(
   type='CrossEntropyLoss', use_sigmoid=False, loss_weight=0.4))
   ],
   test_cfg=dict(mode='slide', crop_size=(512, 512), stride=(341, 341)),
   )

optimizer = dict(
lr=0.01,
weight_decay=0.0,
paramwise_cfg=dict(custom_keys={'head': dict(lr_mult=10.)}))

num_gpus: 8 -> batch_size: 16

data = dict(samples_per_gpu=2)

[MengzhangLI]

Hi, hope this doc may help you.

Best,

[rouchoo]

Thanks for your answer. However i still can not address the problem. When training ealier, the Miou and Acc is normal. But after hundreds iters, the Miou and Acc scores are tending to 0.
I try another model like vit, FCN, the results are normal.

[MengzhangLI]

I think this phenomenon is very wired. From my point of view, SETR should perform better than FCN in many cases, perhaps some hyperparameters of SETR are not the best match up in your dataset.

To sum up, you can try to find suitable model in your customized cases.

[rouchoo]

Thanks for your reply. when i replace the setr_pup model with setr_naive model, it works!
As a result, it is maybe a bug in the setr_naive model.
Thanks for your attention. MMseg is especially useful !