MLP not learning (Learn state attibutes instead of forcing attributes etc.)

[RY4GIT]

Discussion with @jmframe today

• [x] Use scheduling for learning rate, step size 1 and try gamma 0.7
  • https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.StepLR.html#torch.optim.lr_scheduler.StepLR
  • Probably learning rate 1e-03 is good
• [x] Check to make sure initialization is correct
• [x] Use at least 3 years of observation data for training
• [x] Use state variables (e.g., GIUH queues, soil bucket level, or GW level) as MLP learning input, because we want to produce different responses to the same situation (forcing) given different initial conditions (state in the model)
• [x] Change self.c from NLDAS attributes to forcing?

[RY4GIT]

Addressed tasks in NWC-CUAHSI-Summer-Institute/dCFE/pull/29 Testing in progress

[RY4GIT]

Discussion with @jmframe and @taddyb today

• [x] Normalize by max & min state of reservoir
• [x] Run CFE once before MLP in epoch 1
• [x] Try Cgw instead of refkdt
• [x] Everytime step run MLP, use normalized c t-24 to t-0 & states at t-0 because Sequence-to-sequence doesn't work (currently timestep 0 has access to 3600)
• [ ] Move to LSTM

[RY4GIT]

Now I get errors about gradient overwritten

[RY4GIT]

Debug session with @jmframe today

What's done

• Remove shuffle from the dataloader
• Add activation function in between the Linear layers
• Increase the hidden size (2/3 of input)

Todo

• Learn against real streamflow case (hard to make static prediction with MLP)
• Create a plot
• Loop through each basin

[RY4GIT]

Addressed in 67e119fcce24f831846837a1da39aedb95b97d5a