[Question] validation cross entropy about comma10k HOT 11 CLOSED

Yes @dorkmo, @geohot I will work on it a bit more this weekend and release the weights too.

from comma10k.

Thanks!
I used a unet structure too with efficientnet-B4 as encoder, no zooming or TTA. Getting val CE = 0.045.
I can probably do better with weather/lighting augmentations.
I trained first on (H//2, W//2) then continued on full res. Training on full res from scratch gives me around 0.05x too.
Great dataset! I can't wait for the full release.

from comma10k.

Great stuff! Can you add a LICENSE and release weights?

And I didn't know about "segmentation_models.pytorch" or "albumentations", pytorch is fire!

from comma10k.

It's crazy that this doesn't overfit, you aren't scale augmenting I don't see.

Added link to our README with the new record.

from comma10k.

The pytorch ecosystem is 🔥
Sure, I will release weights and a new version ASAP. (I'm slowed down by a conference I'm attending this week)
Thanks for linking to the repo!! I added a MIT license.

from comma10k.

very cool! ill share with the #comma-pencil guys on discord

from comma10k.

I released the weights here: https://github.com/YassineYousfi/comma10k-baseline/releases/tag/v0.1-alpha

from comma10k.

Yes, and that was with a unet. But if you can do a lot better with more FLOPS, go for it!

from comma10k.

Sweet! Want to open source it?

from comma10k.

is this it? https://github.com/YassineYousfi/comma10k-baseline

from comma10k.

There is some regularization (weight_decay), but I would guess that the robustness to overfitting mostly comes from Imagenet pre-trained encoder and the augmentations.

from comma10k.