This repo contains assignments of CS598 / IE534 Deep Learning @ UIUC (Fall 2018 semester).

• HW1: Train a perceptron network from scratch for MNIST dataset
• HW2: Train a multi-channel CNN from scratch for MNIST dataset
• HW3: Train a deep convolution network on a GPU with PyTorch for the CIFAR10 dataset
• HW4: Implement a deep residual neural network for CIFAR100
• HW5: Implement a deep learning model for image ranking
• HW6: Generative adversarial networks (GANs)
• HW7: Natural Language Processing A
• HW8: Natural Language Processing B
• HW9: Video recognition I

• Login

  ssh <bw-username>@bwbay.ncsa.illinois.edu
  

• Mount the remote file system into your local machine

  Note: on Bluewaters, this approach works much better than scp since you don't have to manually scp everytime you update a file.

  At your local machine

  mkdir <local-path>
  sshfs -o ssh_command="ssh <bw-username>@bwbay.ncsa.illinois.edu ssh" h2ologin:<remote-path> <local-path>
  

• Load interactive mode

  qsub -I -l gres=ccm -l nodes=1:ppn=16:xk -l walltime=02:00:00
  

• Login to the compute node

  In compute node, you can directly run your program.

  module add ccm
  ccmlogin
  module load bwpy/2.0.0-pre1
  

• Submit a job to BW using a .pbs file

  This can be done simply via login node. See below for a template.

  qsub run.pbs
  

• Check job status

  qstat | grep <username>
  

• Delete a job

  qdel <job number>.bw
  

• Check your remaining training quota

  usage
  

#!/bin/bash
#PBS -l nodes=1:ppn=16:xk
#PBS -N <job name>
#PBS -l walltime=40:00:00
#PBS -e $PBS_JOBNAME.$PBS_JOBID.err
#PBS -o $PBS_JOBNAME.$PBS_JOBID.out
#PBS -M <your email to receive messages about your job>
cd <path to the folder that your program locates>
. /opt/modules/default/init/bash
module load bwpy/2.0.0-pre1
module load cudatoolkit
aprun -n 1 -N 1 python3.6 <program name>