Most recent stuff up here, older down there.
March 9, 2021
Okeydokey. I’ve got my Shark Tank going and YouTube closed, so it’s time to be productive.
Today’s will be all about twins work
- Finish up the methodology
- GridSearch for best model (Keira)
- Discuss results
- That nasty frizzy-haired lady is back again (ugh)
- Maybe think about how to deal with resolution-based portion.
I will work on Arcface feature extraction on Jetson Xavier after twins work is done (hopefully within a week). Note: use C instead of the Python apps… I know. It hurts. Combine redaction pipeline with the advice from NVIDIA forums/TensorRTx guy. I also need to proofread Oge’s paper… let’s do that right now before the rest of the night starts. Ok, that’s done, and tbh was a hellish experience… I need to download MS Office when I have the chance, since the online version is infuriating to the nth degree.
In other news, I have likely spent more than 1/2 of the past 24 hours on YouTube watching “maker” videos, so we need to think of something to build. Maybe I’ll make a separate article on this site for ideas and progress outside of the academic sphere. Saw lots of superhero stuff, so maybe I’ll go watch the MCU again and find some interesting tidbits for inspiration (idk, so much in there).
Maybe use less pineapple in smoothies (too many of those d*mn carbs). More ginger, too.
March 8, 2021
Let’s get some Spotify crankin’, and some work movin’.
Things to do: twins at different resolutions/add to paper, dashboard stuff (Jetson and CINA), Oge work, trajectory work.
Almost done with the methodology portion of the paper… still need to work on the different resolutions of all the images.
I imagine that the best way to compare the original images and the synthetic samples is going to be with perceptual loss or another face matcher. I probably shouldn’t use cosine distance between Arcface feature vectors. What would that even look like? What’s the best way to compare the feature vectors across the different resolutions? I’ll be thinking about this more later.
Should we use streaming dataframes from the streamz package? Perhaps… will have to wait to see what BL says about progress.
Maybe, we can put the period streaming df (pdf) in the flask API portion. Would that work? Only one way to find out I suppose.