Skip to content
This repository has been archived by the owner on Nov 29, 2023. It is now read-only.

Model updates/CoordConv/Anti-aliased training #38

Closed
wants to merge 7 commits into from

Conversation

jacobbieker
Copy link
Member

No description provided.

@jacobbieker jacobbieker self-assigned this Jul 5, 2021
@jacobbieker
Copy link
Member Author

Relates to #32 somewhat, training DeepLabV3 on segmenting out the current images' clouds. Idea being to then adapt it to predict future timestep clouds as well. Need to figure out a way to make the segmentation models predict multi-frame output. Having it work as a single timestep per output class didn't work.

@jacobbieker
Copy link
Member Author

https://github.com/tcapelle/moving_mnist/blob/master/nbs/01_models.conv_rnn.ipynb seems to have a way around it, with the StackLoss and things like that. Could try converting it to not rely on fastai. Fastai does seem to have quite a bit of nice stuff built in though.

…xample.

Only training to predict mask from satellite channels for current image, so pretraining, but not predicting the future. Still seems to be some error in training on multiple timesteps, might have to add extra input to condition on time, as segmenting stack of images is a little finicky.
@jacobbieker jacobbieker closed this Sep 9, 2021
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
enhancement New feature or request
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant