openclimatefix / predict_pv_yield

Using optical flow & machine learning to predict PV yield
MIT License
52 stars 11 forks source link

Try 2d, 3d CNN before attention #63

Open JackKelly opened 3 years ago

JackKelly commented 3 years ago

related: #35

3D cnn: sees two timesteps at once

peterdudfield commented 3 years ago

https://github.com/openclimatefix/predict_pv_yield/blob/issue/85-perciever-nwp/predict_pv_yield/models/perceiver/perceiver_conv3d_nwp_sat.py

This model does now

JackKelly commented 3 years ago

Nice! How well does the new model perform?! :)

peterdudfield commented 3 years ago

I dont have a direct comparision to running it without the cnn.

results: https://github.com/openclimatefix/predict_pv_yield/blob/issue/85-perciever-nwp/experiments/2021-09/2021-09-28/experiments.txt

JackKelly commented 3 years ago

Cool beans. No rush, but it might be nice to have a direct comparison, so we can get a feel for whether it's useful to place a CNN before the attention. What do you think?

peterdudfield commented 3 years ago

yea, I think it would be a useful to do. Could also see if 01, ... N layers of CNN are useful?

JackKelly commented 3 years ago

sounds great!