temigo / uresnet_pytorch

PyTorch implementations of dense and sparse U-ResNet

uresnet_pytorch

PyTorch implementations of dense and sparse UResNet

This README is very rough and will be completed soon. Stay tuned!

Software containers

Singularity containers are available on https://www.singularity-hub.org/containers/6596.

Dataset

LArTPC simulation dataset are publicly available on https://osf.io/9b3cv/.

Run

All options can be found in uresnet/flags.py.

To train sparse U-ResNet you can use for example:

python bin/uresnet.py train -chks 500 -wp weights/snapshot -io larcv_sparse -bs 1 --gpus 0 -nc 5 -rs 1 -ss 512 -dd 3 -uns 5 -uf 16 -dkeys data,fivetypes -mn uresnet_sparse -it 10 -ld log -if your_data.root

To run the inference:

python bin/uresnet.py inference --full -mp weights/snapshot-1000.ckpt -io larcv_sparse -bs 1 --gpus 0 -nc 5 -rs 1 -ss 512 -dd 3 -uns 5 -uf 16 -dkeys data,fivetypes -mn uresnet_sparse -it 10 -ld log -if your_data.root

Main command-line parameters:

• -mn model name, can be uresnet_dense or uresnet_sparse
• -io I/O type, can be larcv_sparse or larcv_dense
• -nc number of classes
• -chks save checkpoint every N iterations
• -wp weights directory
• -bs batch size
• --gpus list gpus
• -rs report every N steps in stdout
• -ss spatial size of images
• -dd data dimension (2 or 3)
• -uns U-ResNet depth
• -uf U-ResNet initial number of filters
• -dkeys data keys in LArCV ROOT file
• -it number of iterations
• -ld log directory
• -if input file
• -mp weight files to load for inference

Authors

Laura Domine & Kazuhiro Terao

uresnet_pytorch's Issues

Inference does not work without --full flag

Within the inference_loop(flags, handlers) function, the print functions raise an error.

Here's the command I run:
python bin/uresnet.py inference -mp weights/snapshot-2999.ckpt -io larcv_sparse -bs 1 --gpus 0 -nc 5 -rs 1 -ss 512 -dd 3 -uns 5 -uf 16 -dkeys data,fivetypes -mn uresnet_sparse -it 100 -ld log -if '/gpfs/slac/staas/fs1/g/neutrino/kterao/data/dlprod_ppn_v10/shower_relabel/dlprod_512px_09.root'
Here's the error:

Traceback (most recent call last):
  File "bin/uresnet.py", line 14, in <module>
    main()
  File "bin/uresnet.py", line 11, in main
    flags.parse_args()
  File "/u/nu/drielsma/uresnet/uresnet_pytorch/uresnet/flags.py", line 153, in parse_args
    args.func(self)
  File "/u/nu/drielsma/uresnet/uresnet_pytorch/uresnet/main_funcs.py", line 56, in inference
    inference_loop(flags, handlers)
  File "/u/nu/drielsma/uresnet/uresnet_pytorch/uresnet/main_funcs.py", line 242, in inference_loop
    np.unique(np.argmax(res['segmentation'][0], axis=0)[(data_blob['data'][0]>0)[0, 0, ...]], return_counts=True),
TypeError: 'bool' object has no attribute '__getitem__'

NB: data_blob['data'][0]>0 is bool