Motivation

I wanted to try out a large-scale language model (LLM) for Japanese, so I used rinna, which was released in May. To save installation time, I ran rinna under a docker container environment.

I ran into some problems in doing so, which are summarized below.

Introduction

It has been a year since Ubuntu 22.04 LTS Jammy Jellyfish was released, and I decided to switch from my previous 20.04 to 22.04, thinking that I would be able to get used to the various software and other features.

I installed ubuntu 22.04 Japanese Remix on several workstations so that I can use Japanese easily. I have gotten a bit stuck trying to install ubuntu 22.04 on a NVIDIA GPU-equipped workstation, so here is the situation and how I responded.

Motivation

In a previous post I summarized how to start a docker container in user mode.

When using multiple PCs (WorkStation, hereafter WS), it is necessary to consider how to realize a mechanism to share containers among multiple WS. In the case of singularity, we could store sif files on an NFS server and use them from other WS.

In case of docker, I decided to build a registry server, thinking that I could set up a private registry in my home network and operate it.

Motivation

I have been using singularity containers instead of docker. the reason I have avoided using docker is that it requires root privileges to launch containers. I don’t have to worry about it because I’m running it at home.

I forgot the reason why, but I found out that docker has a rootless mode that allows you to run docker in normal user mode, so I decided to give it a try this time.

Motivation

In my home environment where I use NFS to store JupyterLab notebooks, I measured the performance when the NFS server is a RaspberryPi or HP Z240, and found that in the learning loop (state in which epochs are stacked), there is no significant difference whether the NB is stored on an NFS server or locally. I found that there was no significant difference between NBs stored on an NFS server or locally.

Therefore, I have challenged to speed up the learning process, and I summarize the progress/results here.

In the beginning

In the five articles from this article to this article Galaxy shape(morphological) classification was performed using CNN (VGG16, ResNet) and ViT. In this article, I would like to consider the dataset used for the galaxy morphological classification, as I would like to re-examine the dataset when analyzing errors in the future.

Introduction.

In this article, we performed galaxy shape classification using the Vision Transformer (ViT) of Transformer’s adaptation to image processing. This time, we performed the same classification using a model already trained as a model for ViT.

Introduction

In this article at the end of last year, we mentioned that we performed galaxy shape classification with CNN (VGG16, ResNet). In this article, we will perform galaxy shape classification with a Transformer-based Vison Transformer (ViT).

Introduction

I have been studying Deep Learning for a while and thought I would try it out in my field of interest. I like astronomy and am particularly interested in stellar evolution, the formation of elements, and galaxy formation and evolution. I tried to classify the shape of galaxies, which seemed to be relatively easy to do.

Continued from previous article.

Why doesn’t the training/testing error improve after repeated epochs? The results of various attempts to answer this question are summarized below.