2D-DDNet with VGG Loss and architecture aware sparse optimizations

2D-DDNet with optimizations

instructions on running at ARC ( or similar system with Slurm and Singularity)

Read Access to "/projects/synergy_lab/garvit217/enhancement_data/"
Pytorch container 22.04

Pulling container at ARC

Run below commands:

module load containers/singulartiy
singularity pull pytorch22.04.sif docker://nvcr.io/nvidia/pytorch:22.04-py3

Other versions of the container can also be used, but driver compatibility needs to be checked with the driver version on GPU nodes at ARC. Carefully verify from documentation the follow:

git checkout
cd 2dnet
source params.sh
sbatch job_tinker.sh
output will be generated in slurm-.out

Instructions on profiling

source params.sh
export enable_profile="true"
sbatch job_tinker.sh

Read carefully the comments in job_tinker.sh

Provide access to all users to Python file

  chmod 766 sparse_ddnet.py
  chmod 766 trainers.py

DDNet baseline hyper-parameters

batch size: 1
learning rate: 0.0001
epochs: 50
decay rate: 0.95

Enabling Architecture-Aware Optimizations:

Mixed precision: export mp=true
DoLL Data Loader for Small Datasets export new_load=true
Graph Optimizations:

#to enable graph change pytorch version above otherwise, the below two parameters won't be respected

export gr_mode="reduce-overhead"
export gr_back="aot-eager"

Enabling Sparse Optimizations:

export retrain=0 # should be >0 options for prune_t (prune type) mag, l1_struc or random_unstru (default) will be set otherwise
export prune_t="random_unstru" 
export prune_amt=0.5

Enabling VGG loss

export model="ddnet" # choice ddnet, vgg16 (ddnet with vgg-16 based loss), vgg19 (ddnet with vgg-16 based loss)

Scaling

Make the following changes in job_tinker.sh file

export MASTER_PORT=<some unique value>

And update below SLURM headers

#SBATCH --ntasks-per-node P
#SBATCH --gpus-per-node G
#SBATCH --nodes N

P: number of parallel process G: Number of GPUs per node

G=P

N: number of nodes

ayushchatur/2dnet