Commit 735fe337 authored by Devanshu Agrawal's avatar Devanshu Agrawal
Browse files

Update job scripts.

parent 0351887d

torch_experiments/boston.lsf

+2 −2
Original line number Diff line number Diff line
@@ -2,7 +2,7 @@ 


#BSUB -P MED107

#BSUB -W 2:00

#BSUB -nnodes 6

#BSUB -nnodes 40

#BSUB -J boston

#BSUB -o output/boston.o

#BSUB -e output/boston.e
@@ -11,7 +11,7 @@ 
#BSUB -rn



# this must match the -nnodes argument above

NNODES=6

NNODES=40



set -ex

torch_experiments/iris.lsf

0 → 100644
+44 −0
Original line number Diff line number Diff line 
#!/bin/bash



#BSUB -P MED107

#BSUB -W 2:00

#BSUB -nnodes 40

#BSUB -J iris

#BSUB -o output/iris.o

#BSUB -e output/iris.e



#BSUB -q batch

#BSUB -rn



# this must match the -nnodes argument above

NNODES=40



set -ex



PROJDIR=$PROJWORK/med107

USERDIR=$PROJDIR/$USER

REPODIR=$USERDIR/bottleneck_nngp

JOBSCRIPTDIR=$REPODIR/torch_experiments



# set up env

. /sw/summit/init/profile

module load ibm-wml-ce

conda activate $USERDIR/condaenvs/torch



jsrun -n$((NNODES*6)) -a1 -g1 -c7 -r6 \

	-E LD_LIBRARY_PATH \

	--bind=proportional-packed:7 --launch_distribution=packed \

	$JOBSCRIPTDIR/mapper.sh \

		$(which python) $JOBSCRIPTDIR/main.py \

		--dataset=iris \

		--depths=1,%depth% \

		--widths=%width% \

		--vb=0.1 \

		--vw=1.0 \

		--vn=0.1 \

		--train_samples=100 \

		--test_samples=1000 \

		--lr=1e-1 \

		--iters=2000 \

		--gpu=0 \

		--manual_grad \

torch_experiments/mapper.sh

+15 −4
Original line number Diff line number Diff line
@@ -2,8 +2,13 @@ 
set -e



# post-bottleneck depths and bottleneck widths

depths=(10 20 30 40 50 60 70 80 90)

widths=(2 8 32 128)

depths=(1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20)

widths=(1 2 4 8 16 32 64 128 256 512 1024 0)



# number of models

num_depths=${#depths[@]}

num_widths=${#widths[@]}

let num_archs=num_depths*num_widths



# get argument

CMD=$*
@@ -16,9 +21,15 @@ LOCALRANK=$OMPI_COMM_WORLD_LOCAL_RANK 
RANK=$OMPI_COMM_WORLD_RANK

JOB=$LSB_JOBID



# exit if rank larger than number of architectures

if [ $RANK -ge $num_archs ]

then

	exit 0

fi



# calculate architecture given rank

DEPTH=${depths[$((RANK/4))]}

WIDTH=${widths[$((RANK%4))]}

DEPTH=${depths[$((RANK/num_widths))]}

WIDTH=${widths[$((RANK%num_widths))]}



# substitute into command

CMD=${CMD//%depth%/$DEPTH}

torch_experiments/rings.lsf

0 → 100644
+44 −0
Original line number Diff line number Diff line 
#!/bin/bash



#BSUB -P MED107

#BSUB -W 2:00

#BSUB -nnodes 40

#BSUB -J rings

#BSUB -o output/rings.o

#BSUB -e output/rings.e



#BSUB -q batch

#BSUB -rn



# this must match the -nnodes argument above

NNODES=40



set -ex



PROJDIR=$PROJWORK/med107

USERDIR=$PROJDIR/$USER

REPODIR=$USERDIR/bottleneck_nngp

JOBSCRIPTDIR=$REPODIR/torch_experiments



# set up env

. /sw/summit/init/profile

module load ibm-wml-ce

conda activate $USERDIR/condaenvs/torch



jsrun -n$((NNODES*6)) -a1 -g1 -c7 -r6 \

	-E LD_LIBRARY_PATH \

	--bind=proportional-packed:7 --launch_distribution=packed \

	$JOBSCRIPTDIR/mapper.sh \

		$(which python) $JOBSCRIPTDIR/main.py \

		--dataset=rings \

		--depths=1,%depth% \

		--widths=%width% \

		--vb=0.1 \

		--vw=1.0 \

		--vn=0.1 \

		--train_samples=100 \

		--test_samples=1000 \

		--lr=1e-1 \

		--iters=250 \

		--gpu=0 \

		--manual_grad \