Vision: command line to train FCN and DeepLab

[a-maci]

What are the command lines to train the segmentation (FCN and DeepLabV3) networks to accuracy mentioned in torchvision website (63.7 and 67.4 IoU).

[a-maci]

Also, when running the training script with default parameters, I get this error:

model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
AttributeError: type object 'SyncBatchNorm' has no attribute 'convert_sync_batchnorm'

[fmassa]

Hi,

Sorry for the delay in replying.

Here are the commands I used

python -m torch.distributed.launch --nproc_per_node=8 --use_env train.py --lr 0.02 --dataset coco -b 4 --model fcn_resnet101 --aux-loss

for fcn_resnet101 and

python -m torch.distributed.launch --nproc_per_node=8 --use_env train.py --lr 0.02 --dataset coco -b 4 --model deeplabv3_resnet101 --aux-loss

for deeplabv3_resnet101.

About your error, it looks like you are not using a recent enough version of PyTorch? What's your PyTorch version? It should be at least 1.2.

Let me know if you have further questions

[isalirezag]

@a-maci can I know with what code you're going to use that command line? which dataset? where is the file for train.py?

[a-maci]

Coco dataset with train.py in references/segmentation folder.

[fmassa]

Thanks for raising this up again @isalirezag , I'm updating the README in https://github.com/pytorch/vision/pull/1864 with the commands for training the model