Slow, but it seems it works on the big boy dataset too. Just some issues with I/O most likely. Time to polish it up a bit.

Hehe "big boy" by my standards, that is

I'm getting the distinct impression that performance issues in maxbin2 are a combination of I/O not designed for the data sizes that mag handles, and the tools it calls out to being kinda slow. The IO issues are all very fixable.

CAMI I High-Complexity — from the https://doi.org/10.1038/nmeth.4458 (CAMI) challenge - getting identical output with the re-implementation, all C++ and perl gone.

(there's the external dependencies, still - not touching those; easier to wrangle now, though)

I feel like it finally clicks how EM actually works, and... wtf? I'm sorry... What. That's such a counterintuitive result. Makes sense in hindsight, but what? How's a 4-mer frequency profile indicative of species, rather than similar contigs across species? Like, I understand the explanation, but it's still... how's that a thing you can just use as a classification signal?

TIL

long double

is 80 bits on Linux x86_60. 80 bits, not 64, not 128, it's 80. OS and architecture dependent. Why.

maxbin2 really loves its

printf

calls

I'm staring at about a 14x speedup of the core EM algorithm, Rust vs C++ - which is honestly setting off my bullshit alarms more than anything else. • looks like I missed another sprintf call. Sadge, would've been nice.

The 10x speedup is still looking harder to break than I expected. What is going on.

I would reply in the original thread but it's stuck on "loading"