A graph is one expression
a >> (b | c) >> d, and a node is one thing: forward(input, ctx) takes
what arrived along the edges and returns what it produced. No filter type,
no step type, no return value that tells them apart. The graph is compiled
and executed in Rust.
Where it runs is a suffix
.on("cuda:0") is a device, .at("worker1") is another machine,
.cached() is a value worth keeping. The graph says what exists and
those say where and what is remembered — five orthogonal facts, and
confusing them is the easy mistake.
It tells you what happened
A run emits facts as it goes, including from the machines it was spread
over — a Watcher is a hole in the core, so a notebook is told live and a
store is written for later. The record is read back with a price list, and
it draws.
And whether the network is learning
Dead units, saturation, a step-to-weight ratio that has gone. A diagnosis
is an opinion about the stored record, which is what makes arguing with
a threshold cost a scan instead of an afternoon of GPU. And probe asks
what can be said before the first step is taken.
A search that needs no coordinator
A trial is a number, ask is a function of it, and a shared folder settles
who gets which — so a machine that claimed trial 7 derives it without
replaying six. Halton and Sobol cover the space for every prefix, which is
what stops two machines proposing neighbours.
And what you were trying to find out
The half a repository loses. somatize-tree writes down questions,
hypotheses, attempts, findings and decisions in a DAG beside the code —
then go takes you back to a move by the name you remember, instead of a
hash you do not.
pip install 'somatize[viz]'The Rust is compiled inside the wheel; viz adds plotly, which every figure
here needs. Then a first graph in five minutes, or
the thirteen notebooks, which ship executed with
their outputs.