Diary 2023-10-01
9/30
nishio I'm going to get a certain competition to the point where I can submit a baseline for now, and then I'll figure out a strategy for after that. nishio Reading the chat log while it's running, I'm reading a conversation from 4 days ago now. nishio ah, I'm easily falling into the trap of device map auto and "1 GPU for now" w nishio Oh, I saw someone who got to 100 and then fell off, but I guess the task switches to a summary task from 100, it's taking me a long time too, I'm on 109 right now. If it takes too long, I'll stop and start over. nishio I've made 5 progress in 14 minutes, so, well, I guess I have 30 minutes left. Let's wait then. nishio I missed that there is also a generation task at the end! nishio I've been experimenting in parallel in the lab environment and in Colab Pro, and I've got the code for the third experiment in hand, but I can't seem to finish the other two, so I'm going to have to use the company I'll have to VPN into the network... 10/1
nishio I reaffirmed that I am not good at making do with limited resources or remembering to do something properly. I reaffirmed that I am not good at making do with limited resources or remembering to do something properly (I used up all the GPUs in the exercise environment and charged for Colab Pro+). I took a bath while it was running, and then I just fell asleep.
nishio I also thought that the prioritization of "we should look into the uncertainties first" was conditional on having enough resources (the additional GPU credit requirement I ran out before I could meet them). We should have first ensured that the low-risk option was borderline and then experimented with the high-risk option.
Up at 6:00 a.m. tomorrow.
---
This page is auto-translated from /nishio/日記2023-10-01 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.