Now, you can search your commit history to find good results and the associated code state.

Then when your experiment runs, logis will commit your code for you, with a nice commit message and experiment metadata at the bottom.

Or use the (work-in-progress) implicit API, where logis finds your parameters/metrics in the arguments and return value.

The SDK is similar to Weights & Biases, just add relevant data to your experiment's run.

There is an API for building custom @hooks - they're just functions.

You can also pre-process the raw data using @hooks.

The @env:MY_VARIABLE hook loads data from an environment variable. The @value:path.to.val hook references a value in the raw input data. The @import:foo.bar dynamically imports a Python object.

The idea is simple: write an object's Blueprint once, then whenever you need that object in your @pydantic.dev models, just pass in the parameters.

In #machinelearning and #neuroai, we build a lot of Tensors from various parameterisations...

Supports Dale's Law: inhibitory neurons (blue columns) have local within-module connections only.

🔧 Re-sharing an old tool: a recursive function to generate hierarchical, modular, Dalean connectivity matrices in PyTorch.

Modules are locally dense, with increasingly sparse connections to more distal modules.

#neuroai #neuroskyence

gist.github.com/rorybyrne/dd...

I've been using a custom version of this tool to parameterise neuron models for a long time now.

1. Write a parameterisation ("Cast") and register it with the `cast.for_type(Tensor)` decorator.

2. Use `CastModel` to build your @pydantic.dev models.

3. You can now pass the _parameters_ of your tensor, and receive a built tensor.