2022-10-06 1319+0000 I’m back for a bit after some things. I feel like looking at the code some more. I have a couple hours to my appointments, and am ready enough to go. 1349 I’ve gotten through the attention output. Torch has an existing function for calculating cosine similarity. Cosine similarity is simply the normalized dot product of the vectors: i.e. the cosine of the n-dimensional angle between them. I’ve pushed my draft to git. Next is organisation, testing, and then often a bit of recoding when I find my mistakes. Noting I still have time before my appointments. 1430 I’ve quickly organised the draft a little and tested it imports and constructs without parse failures. Pushed to git. I’m spinning out a little. My appointments are soon. Next is relearning to use this library to test my added algorithm. 1436 Since I copied an existing algorithm almost exactly, I can run all its tests on my copied code. The tests also include an example of using it, which seems the simplest way to learn this. I’m running into pytest crashing weirdly after updating a dependency. 1457 looks like i need to enable the [dev] extra on mosaicml composer to pull in the test dependencies. the tests run now! i have 5 mistakes to resolve. 1503 i pushed a change to git. the vanilla tests are segfaulting on my system associated with tensorflow. my algorithm doesn’t use tensorflow so i drill down into just it. the tests over the algorithm all pass — but having looked at them i know they don’t verify its output, just its structure. next is to try to use the code to do something simple, and compare a result to without the change. my appt is at 1600; i’m in the habit of going a little early to protect against missing it from weird states of mind i can have.