how to Change SynBN to regular BN ? about fastfcn HOT 13 CLOSED

1. Modify this line to se_loss = args.se_loss, norm_layer = torch.nn.BatchNorm2d,
2. Modify this line to loss = self.criterion(*outputs, target)
3. Remove this line
4. Modify this line to nn.BatchNorm1d(ncodes),
5. Reinstall FastFCN

pip uninstall fast_fcn
PATH=.:$PATH
python setup.py install

from fastfcn.

model = get_segmentation_model()
define norm_layer to be torch.nn.BatchNorm2d solve my problem.

from fastfcn.

@alphaccw i have change the norm_layer ,but it also occur the problem. like below:
AttributeError: 'NoneType' object has no attribute 'run_slave'

Could you help me ?
thank you very much

from fastfcn.

@10183308
by this changing, i can successfully run psp and deeplab models, enc seems still have some problem, i didnot go into it yet.
BTW, i remember i need to set as 'norm_layer=torch.nn.BatchNorm2d', the full name...i canot accesss to the machine now, so i cant check at the moment

from fastfcn.

Still doesn't work

from fastfcn.

Please give the error message in detail.

from fastfcn.

first, I have this problem, #10 when train
I modified four places you mentioned, and then run in docker，use single gpu

CUDA_VISIBLE_DEVICES=0 python train.py --dataset pcontext \
    --model encnet --jpu --aux --se-loss \
    --backbone resnet50 --checkname encnet_res50_pcontext \
    --batch-size 1

get

then I add --workers 1
still error

from fastfcn.

5. Remove DataParallelModel from this line.
6. Remove DataParallelCriterion from this line.
7. Undo 2: Modify this line to loss = self.criterion(outputs, target)

from fastfcn.

still the same error.
my environment，python3.7, pytorch1.0.1, cuda 9.0, cudnn 7.4.2, GTX1080，using docker
is there anything wrong with my environment?
I can run test, test_fps and predict scripts successfully.

from fastfcn.

Have you reinstalled the package after modifying the code?

from fastfcn.

No. which package need to be reinstalled?

from fastfcn.

pip uninstall fast_fcn
PATH=.:$PATH
python setup.py install

If you want to modify the code without reinstalling every time, then

pip uninstall fast_fcn
PATH=.:$PATH
python setup.py develop

Besides, I think step 5-7 is unnecessary, which can be skipped.

from fastfcn.

Step 5-7 is unnecessary.
Thank you very much.

from fastfcn.