https://acmecups.eu

Andersen bakeries use these mugs.

Do you have a source? Everywhere I looked says I need to be registered at the right location. I had unexpected circumstances and am not located in the correct ward today, and seems like I can’t vote.

https://www.elections.org.za/pw/Voter/Moved-Since-You-Registered

This is the only right answer for cheap and reliable. /thread

Bilka has this one.

I prefer knowing, so I don’t waste my time applying. Why not be transparent?

Thanks, I will check them out!

Jumping on this thread, is there anywhere to find good cold meats? I’m tired of buying 150g prepacked ham :/

Hi. Try reach out to Centro Odontoiatrico Grimaldi. Good doctors and they are responsive on WhatsApp — they should be able to help you out.

🤣 did you give him R100

Is he the one with the American accent? Needs cash to get to embassy etc?

Vanish powder also has Activated Oxygen Bleach with TDEA.

Seems very low as a salary. Feel free to share your CV once you’ve updated it, I can share it in a few circles.

Did you check with Tom Campher? They’re the most reasonable. If you speak to Gerald he might be able to point you in the right direction.

The loudest restaurant in Johannesburg lmao

Different types, but the ones most recently are IoT data sources from cell phone towers. Voltages, currents, alarms, that kinda stuff. Proprietary but not particularly interesting. If you’re interested, I would try find industrial IoT data sources (smart washing machines, fridges, etc.). Maybe a bit boring but real and useful for learning time series problems.

When I have independently tracked the reconstruction and KL divergence, I noticed that typically in my specific use case, one loss would completely dominate the equation. So it would only use the reconstruction loss. Using a β-VAE helped, where the loss is started as reconstruction only, and the KL is slowly brought in over a few epochs. This has an additional headache in that you have to tinker to make it work, but it does eventually work.

Maybe this has just been my experience with my own datasets. But VAEs are tricky beasts because there is no single metric that tells you what your performance is like. In my use cases I always have a downstream task where I get that number.

Thanks for sharing. I’ll look into this to add to my toolkit too. I’ve worked a lot with industrial IoT data, and training there has always been quite tricky. Visualizing the two losses side by side was the only way I could guarantee it was actually training correctly. There are no other obvious ways of knowing if a VAE is doing what it’s meant to, particularly with a non 2D latent space. Not that I know of anyway 🤔

Thanks for sharing. Yes that’s what I meant. Practically I view them as two losses due to the way I work with them in TensorFlow. I have a suspicion that most VAEs are actually not training as intended, and the reconstruction loss dominates the training. Thanks for sharing your decoder tips. I’ll look into it.

Non-parametric though. Only good for visualizations or fixed data.

Out of interest, in your course, did you monitor the two losses of VAEs to determine if they were actually learning what they are meant to? I’ve been working with them for years, and it’s a real dark art to train them correctly.

Yeah looks good. Very clean and sensible. I would probably show some visualizations of the model’s performance, but that’s just a nice to have. Maybe a confusion matrix. But the rest looks great.

Recall, Precision, F1 Score. All of these are important when using imbalanced data. If you have a threshold that you can adjust, it would be worth plotting the recall vs precision plot. AUC is also valuable there.

Too many people get comfortable with headphones and don’t realize. I’ve seen it so often and I’ve only been here for 3 months lol