Okay, I know this is a hot take because everyone seems to love Weights & Biases. I'm a big fan too, for the experiment tracking and collaboration—it's fantastic for team visibility.
But I've been hitting consistent snags specifically inside Jupyter notebooks. It feels like the integration introduces a weird latency. My main gripes:
* **Kernel Interrupts:** When I run `wandb.init()`, especially with a config logged, it sometimes just... pauses. The cell execution light stays solid for 10-15 seconds before anything happens. It's not network-related; it happens offline too.
* **Plotting Jitter:** Using `wandb.log()` inside a training loop (even a simple mock one) makes the notebook feel sluggish. The plots update, but the UI hangs briefly each time. Compared to a barebones Matplotlib update, it's noticeably slower.
* **Automatic Re-initialization:** This is the big one. If my kernel restarts and I re-run cells, I've had old wandb runs re-activate and log to the wrong experiment. I've had to add explicit `wandb.finish()` and `wandb.init(reinit=True)` checks everywhere, which defeats the "seamless" promise.
I'm using the standard `%wandb` magic commands and the API. My workflow is typical: data prep in a notebook, then prototype training loops.
Has anyone else experienced this? I've switched to using W&B mostly via script runs (`!python train.py`) from the notebook, which works fine, but I miss the inline interactivity.
Maybe I'm missing a configuration trick? Or is this just the price for the real-time dashboard sync?
cheers,
danielp