Open lexmar07 opened 1 month ago
OK, actually, one can add something like
model = model.prune(node_th = 1e-1)
to avoid hanging vertices in version 0.2.2
The same issue here.
/usr/local/lib/python3.10/dist-packages/kan/MultKAN.py in fit(self, dataset, opt, steps, log, lamb, lamb_l1, lamb_entropy, lamb_coef, lamb_coefdiff, update_grid, grid_update_num, loss_fn, lr, start_grid_update_step, stop_grid_update_step, batch, metrics, save_fig, in_vars, out_vars, beta, save_fig_freq, img_folder, singularity_avoiding, y_th, reg_metric, displaymetrics) 938 if == steps-1 and old_save_act: 939 #self.save_act = True --> 940 self.recover_save_act_in_fit() 941 942 train_id = np.random.choice(dataset['train_input'].shape[0], batch_size, replace=False)
TypeError: MultKAN.recover_save_act_in_fit() missing 1 required positional argument: 'old_save_act'
It was working properly a couple of days ago.
I'm having the same issue. The problem goes away when I use save_act=False
to override the default (True) as a kwarg for KAN. I only just started playing with this stuff today, so I have no idea whether that's a dumb thing to do.
Problem 1 goes away in version 0.2.3 by changing MultKAN.py line 940 to self.recover_save_act_in_fit(True)
Problem 2 persists in version 0.2.3. Training seems to proceed, but inevitably all losses turn to nans.
I find the same issue! The multiKan is easy to nan.
Has anyone managed to fix problem 2?
Has anyone managed to fix problem 2?
I've rolled back to the July 22 version of PyKAN, and, so far, training proceeds with no nans.
However, I still don't know what is causing problem 2 in later versions of PyKAN.
I I'm having the same issue! Has anyone managed to fix problem 2?
Has anyone managed to fix problem 2?
I've rolled back to the July 22 version of PyKAN, and, so far, training proceeds with no nans.
However, I still don't know what is causing problem 2 in later versions of PyKAN.
How do you use the July 22 version? You just downloaded the folder and is running like a function normally ?
Bug fixed in latest version (at least in HelloKAN).
Good afternoon!
HelloKAN does not work as intended:
1) with version 0.2.3
TypeError: MultKAN.recover_save_act_in_fit() missing 1 required positional argument: 'old_save_act' mentioned above
mentioned here #375 and here #3722) with version 0.2.2:
`train_loss: nan | test_loss: nan | reg: nan
resulting:What is funny is that 0.2.2 used to work just a few days ago. I can hardly understand why this happens. Here is the Google- colab replication code: https://colab.research.google.com/drive/1YOU7AifdYieMWK2hDfKjlN7l6_n6BkvV?usp=sharing