BAIT "Learning to Learn"

Pdf of Bachelor Thesis

Bachelor Thesis

Installation

Dev environment is best set up via VS Code remote server (SSH & DevContainer Extensions)

Building the VM (Docker Host)

1. Get access to https://apu.cloudlab.zhaw.ch
2. Create a Key Pair via Project/Compute/Key Pairs "Create Key Pair"
   • Name: My-SSH-KeyPair
   • Key Type: SSH
3. Create SSH Security Group via Project/Network/Security Group "Create Security Group"
   • Name: SSH -> Create Security Group Add Rule via "Add Rule":
     • Description: SSH
     • Directions: Ingress
     • Open Port: Port
     • Port: 22
     • Remote: CIDR
     • CIDR: 0.0.0.0/0
4. Launch Instance Project/Compute/Instances "Launch Instance"
   • Instance Name: any
   • Source: 2022-01_Ubuntu_Focal_nVidia-Cuda_Docker
   • Flavour: g1.xlarge.t4
   • Network: Allocate "internal"
   • Security Group: SSH
   • Key Pair: My-SSH-KeyPair -> Launch Instance
5. Add FloatingIp Project/Compute/Instances
   • On Instance: Action dropdown (far right) -> Allocate Floating IP
   • choose any
6. Reboot machine to avoid "Driver/ Library version mismatch" Error

Dev Environment (VS Code DevContainer)

VS Code Remote - SSH & Remote - Containers Extensions necessary Connect to Docker Host

1. Add private key (*.pem file) to VS Code
   • copy .pem file to ~/.ssh
2. Add SSH Host
   • ctrl-p -> Remote-SSH: Add new SSH Host
   • Add Floating IP from Docker Host
3. Connect
4. Pull git repo onto Docker host
5. configure git user name

git config --global user.email "lutzurb1@students.zhaw.ch"
git config --global user.name "Urban Lutz"
git config --global http.sslVerify false

Launch DevContainer

1. ctrl+p -> Remote Containers: Reopen in Container

Experiment Specification

Experiments can be specified in CSV format in experiments.csv. Valid columns are all parammeter available in any class from params.py. If a parameter for a given class is omitted, the default value from params.py will be set.

The mechanism allows to specify an arbitrary number of pruning phases by prefixing the pruning parameter with "X_" (X -> phase number, starting at 1). 3 kinds of pruning phases are supported:

• "train": only training, no pruning, (set the train_epochs param)
• "one-shot": only pruning (set strategy, sparsity)
• "iterative": prune and train (set strategy, sparsity, prune_epochs, train_epochs)

Dataset

lives on gdrive https://medium.com/geekculture/how-to-upload-file-to-google-drive-from-linux-command-line-69668fbe4937