Roasting my CPU
I run calculations for image embeddings for the complete Observing the Coming of Age of Video Game Graphics dataset over the last weekend, on a dedicated little computer in the DH Lab in Bern. The process went smoothly and way quicker than anticipated.
Having the complete dataset now with embeddings meant also having to re-do the dimension reduction and clustering. I went about that first on my old laptop, which just about roasted the CPU. I found out that cat /proc/acpi/ibm/thermal will not return values higher than 75° … and I guess it went beyond that. I would have been able to hurt myself touching the device. So, I needed to set up the repository on the newer laptop, a managed device from my employer that I don’t like using much. I run into some problems creating an interesting layout and fiddled quite a lot with the jupyter notebooks. In the end, I came up with a setup of hyperparameters that mirror different interesting aspects, from global form to local oddities. Next step is to do a clustering/hotspots detection and label them.
Meanwhile, I also had the chance to work on the paper and reflect on it. Once again trying to explain what I’m doing here I figured that I need to concentrate more on what I have at hand. I initially had the feeling that I need to present a fantastic finding, but that pressure doesn’t make sense. What I need to do is to propose the method, do the method and discuss the results. No need to unnecessarily blow up the findings. Having shifted my perspective, I was able to think better about the results and came up with one or two very interesting thoughts.
Initially, I found that the DINOv2 embeddings and UMAP dimension reduction to two showed up as what FAVR calls modes. But there are other similarities by formal aspects, such as in the 60/90° angle cluster, or by semantics, like in the NSFW hotspot. And, I guess, somewhere in between is what one might call a genre.
Besides that I worked on dumping the 4A BASIC code from the tapes I got from Laro Schatzer. I was able to recover close to everything, with only one tape showing some corruption. Everything went into a repository and I’ll see what’s next.