/SparseResNet

Sparse ResNet example

Primary LanguagePython

mlws-2019-04-srn

Sparse ResNet for the workshop.

Brief overview

This repo contains quickly put together scripts for training 3D sparse CNN for a ml workshop. Look at models/SparseResNet10.py for the model definition. iotools/spd_*.py are two simple dataloading scheme (event-wise from file vs. loading chunk in RAM). Training step is about 0.1 [s]/iteration on v100, ~80MB tensor size on GPU (yay sparse lib!).

  • Wanted:
    • shared data buffer filling for SPDSimple (in spd_reader.py), which is torch.utils.data.DataSet inherit, so that the already-read data can stay in memory (sparsified so this is fairly small). Maybe w/ multiprocessing module?
    • data augmentation probably needed heavily, or generate or variety of data.
    • point regression is probably a fun extension for the workshop.
    • improve data formatting into sparse 3d tensor: grouping of timing axis not optimal, good to use an idea based on actual geometry.
    • implement graph conv as another way to tackle 3d sparse data. might work better for this classification example.
    • implement segmentation net (least important, but fun if muon+gamma or some mixture event can be generated and separate them at pixel level).

Requirement

  • Software ... Use this Singularity container or this Docker container (tag cuda90-pytorch-dev20181015-scn).
    • Refer to this recipe if you want to check more precisely about dependencies.
  • Data ... consult with Patrick de Perio :)

Running

Example command to run a training:

$HOME/sw/2019-02-mlws//bin/spd.py train -ld log -wp weights/snapshot -chks 100 -nc 2 -it 30000 -bs 32 -mbs 32 --gpus 0  -rs 1 -mn SimpleResNet10 -io spd_buffer -id /scratch/kterao/hkml_data/eminus,/scratch/kterao/hkml_data/gamma/ -nr 10 -dd 3 -ss 128 -uf 16

Example command to run an inference (you'll need your own weights):

$HOME/sw/2019-02-mlws/bin/spd.py inference -ld log -nc 2 -it 300 -bs 32 -mbs 32 --gpus 0  -rs 1 -mn SimpleResNet10 -io spd_buffer -id /scratch/kterao/hkml_data/eminus,/scratch/kterao/hkml_data/gamma/ -nr 10 -dd 3 -ss 128 -uf 16 -mp weights/snapshot-24999.ckpt