arkadaw9/pgnn

Physics-guided Neural Networks (PGNN) : An Application In Lake Temperature Modelling

pgnn's Issues

Error while running the code on Colab

Hello I am trying to run your code on colab.
But I am getting the following error I can't see how I may fix it. Any ideas?

Traceback (most recent call last):
  File "PGNN.py", line 148, in <module>
    PGNN_train_test(optimizer_name, optimizer_val, drop_frac, use_YPhy, iteration, n_layers, n_nodes, tr_size, lamda, lake_name)
  File "PGNN.py", line 110, in PGNN_train_test
    spio.savemat(results_name, {'train_loss_1':history.history['loss_1'], 'val_loss_1':history.history['val_loss_1'], 'train_rmse':history.history['root_mean_squared_error'], 'val_rmse':history.history['val_root_mean_squared_error'], 'test_rmse':test_score[2]})
KeyError: 'loss_1'

The phyloss doesn't pass any gradient to the model parameters

Hello, I have tried running your code following your article.
However, I found that tensorflow only calculates the gradients based on the MSE part in the combined_loss, meanwhile neglecting the phy_loss term.

For instance, I tried training the model only with the phy_loss: model.compile(loss=phyloss, ...), it will return "ValueError: No gradients provided for any variable, XXXXX".

I also tried tf.GradientTape() to calculated the gradients. The gradients calculated using the phyloss is a list of None value (i.e., [None, None, None, None, None, None, None, None])

with tf.GradientTape() as tape:
    Y_pred = model(trainX)
    #loss = mean_squared_error(trainY, Y_pred)
    #loss = totloss(trainY, Y_pred)
    loss = phyloss(trainY, Y_pred)
grads = tape.gradient(loss, model.trainable_variables)

What are the names of the columns in Xc and uX1,2 ?

Hello,
I have tried running your code following your article (https://arxiv.org/abs/1710.11431).
It runs fine but I don't understand how each of the 12 columns in the Xc and uX1,2 arrays relate to the names of the input drivers listed in table 1 of the article (p.6).
Which column is the 'Day of Year', 'Depth', etc ?
Is there any way to find the original dataset, before any transformation (standardization) is performed on the data ?
Thanks,
Best regards,
Clément

What is the actual data of Xc_doy1 and Xc_doy2?

I am trying to use your approach for another Physics-guided problem and have problems replicating it. I have backtracked your physical loss to Xc_doy1 and Xc_doy2, which are in the MatLab files with dimensions 649723x12. I have trouble understanding what these values actually stand for. Are these 649723 temperature samples at 12 different depths in the lake?

Loading mat-files not possible

Hi,

I am having the problem, that i can't open the mat-files:

mat = sio.loadmat("mendota.mat", squeeze_me=True,
    variable_names=['Y','Xc_doy','Modeled_temp']) 

leads to:
ValueError: Unknown mat file type, version 97, 116

Keras v2.6.0
Scipy v1.8.0
Tensorflow 2.6.0