Weights & Biases (wandb)

#1

Hello,

I am currently writing some code to setup some experiments. The final goal is to perform different training sweeps (hyperparameter search). I use Weights & Biases (wandb) for experiment and artifact tracking (metrics, plots, model version, …).

My question can I do something similar in EI Studio. Is it possible to perform the training sweeps in EI and use Weights & Biases for the experiment and artifact tracking.

Regards,
J.

Replacing iPython notebook support (in favor of something better)
#2

Hello @Joeri,

I’ve asked our ML team to have a look at your questions. I am not sure if that easily doable (or just doable) but it’s worth having our expert opinion :wink:

Regards,

Louis

#3

@Joeri We’re working on this in combination with Adding custom learning blocks - Edge Impulse Documentation. The idea would be that you can:

  1. Take the full training code (rather than the notebook) out of EI.
  2. We’ll have some default hooks for W&B, MLFlow and Ray already in there, so you just hook your keys in for those services, and run a ton of experiments (you can get your new data from EI through the CLI easily).
  3. Once you’re happy with the hyperparams / architecture etc. you push the model back to EI, so you can retrain / deploy as always.

We’ll open this up for all developers as well at some point. But expected somewhere in Q3…

1 Like
#4

@janjongboom thanks for the answer. I was already thinking into this direction, however I was curious if there where alternatives.

The idea was to start from the notebook (from EI studio), translate it to python script and add the needed wandb code (run everything in Docker container), but you mention the following: take the full training code, we’ll have some default hooks for W&B. Can you explain this in more detail. Do I need to use the API to retrieve the code?

Small extra question. Before building a learning block I need to write a custom dsp block, I assume that I make a separate container for the dsp block and not combine both (dsp and learning) in one container?

Regards,
Joeri