Error running code from the manual

[diego-uva]

Good afternoon,

I have started using the KAN library and I got a couple of errors running the code from the https://kindxiaoming.github.io/pykan/intro.html#hello-kan manual page.

The first is that the manual has not yet been updated and the train method is used instead of fit.

The second one was when executing this code from the manual page:

from kan import *
# create a KAN: 2D inputs, 1D output, and 5 hidden neurons. cubic spline (k=3), 5 grid intervals (grid=5).
model = KAN(width=[2,5,1], grid=5, k=3, seed=0)
# create dataset f(x,y) = exp(sin(pi*x)+y^2)
f = lambda x: torch.exp(torch.sin(torch.pi*x[:,[0]]) + x[:,[1]]**2)
dataset = create_dataset(f, n_var=2)
dataset['train_input'].shape, dataset['train_label'].shape
# plot KAN at initialization
model(dataset['train_input']);
model.plot(beta=100)
# train the model
model.fit(dataset, opt="LBFGS", steps=20, lamb=0.01, lamb_entropy=10.) #fit instead train
model.plot()
model.prune() #The error in this line!
model.plot(mask=True)

The error is:

    "name": "AttributeError",
    "message": "'float' object has no attribute 'to'",
    "stack": "---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[13], line 1
----> 1 model.prune()
      2 model.plot(mask=True)

File c:\\Users\\hp\\.conda\\envs\\kan\\lib\\site-packages\\kan\\MultKAN.py:970, in MultKAN.prune(self, node_th, edge_th)
    969 def prune(self, node_th=1e-2, edge_th=3e-2):
--> 970     self = self.prune_node(node_th, log_history=False)
    971     #self.prune_node(node_th, log_history=False)
    972     self.forward(self.cache_data)

File c:\\Users\\hp\\.conda\\envs\\kan\\lib\\site-packages\\kan\\MultKAN.py:942, in MultKAN.prune_node(self, threshold, mode, active_neurons_id, log_history)
    938         model2.symbolic_fun[i].out_dim_mult = num_mult
    940         width_new.append([num_sum, num_mult])
--> 942     model2.act_fun[i] = model2.act_fun[i].get_subset(active_neurons_up[i], active_neurons_down[i])
    943     model2.symbolic_fun[i] = self.symbolic_fun[i].get_subset(active_neurons_up[i], active_neurons_down[i])
    945 model2.cache_data = self.cache_data

File c:\\Users\\hp\\.conda\\envs\\kan\\lib\\site-packages\\kan\\KANLayer.py:305, in KANLayer.get_subset(self, in_id, out_id)
    283 def get_subset(self, in_id, out_id):
    284     '''
    285     get a smaller KANLayer from a larger KANLayer (used for pruning)
    286     
   (...)
    303     (2, 3)
    304     '''
--> 305     spb = KANLayer(len(in_id), len(out_id), self.num, self.k, base_fun=self.base_fun, device=self.device)
    306     spb.grid.data = self.grid[in_id]
    307     spb.coef.data = self.coef[in_id][:,out_id]

File c:\\Users\\hp\\.conda\\envs\\kan\\lib\\site-packages\\kan\\KANLayer.py:132, in KANLayer.__init__(self, in_dim, out_dim, num, k, noise_scale, scale_base, scale_sp, base_fun, grid_eps, grid_range, sp_trainable, sb_trainable, save_plot_data, device, sparse_init)
    129 else:
    130     mask = 1.
--> 132 scale_base = scale_base.to(device)
    133 self.scale_base = torch.nn.Parameter(torch.ones(in_dim, out_dim, device=device) * scale_base * mask).requires_grad_(sb_trainable)  # make scale trainable
    134 #else:
    135 #self.scale_base = torch.nn.Parameter(scale_base.to(device)).requires_grad_(sb_trainable)

AttributeError: 'float' object has no attribute 'to'"

Best regards.

Diego.

[atakansite]

Good afternoon,

I have started using the KAN library and I got a couple of errors running the code from the https://kindxiaoming.github.io/pykan/intro.html#hello-kan manual page.

The first is that the manual has not yet been updated and the train method is used instead of fit.

The second one was when executing this code from the manual page:

from kan import *
# create a KAN: 2D inputs, 1D output, and 5 hidden neurons. cubic spline (k=3), 5 grid intervals (grid=5).
model = KAN(width=[2,5,1], grid=5, k=3, seed=0)
# create dataset f(x,y) = exp(sin(pi*x)+y^2)
f = lambda x: torch.exp(torch.sin(torch.pi*x[:,[0]]) + x[:,[1]]**2)
dataset = create_dataset(f, n_var=2)
dataset['train_input'].shape, dataset['train_label'].shape
# plot KAN at initialization
model(dataset['train_input']);
model.plot(beta=100)
# train the model
model.fit(dataset, opt="LBFGS", steps=20, lamb=0.01, lamb_entropy=10.) #fit instead train
model.plot()
model.prune() #The error in this line!
model.plot(mask=True)

The error is:

  "name": "AttributeError",
  "message": "'float' object has no attribute 'to'",
  "stack": "---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[13], line 1
----> 1 model.prune()
      2 model.plot(mask=True)

File c:\\Users\\hp\\.conda\\envs\\kan\\lib\\site-packages\\kan\\MultKAN.py:970, in MultKAN.prune(self, node_th, edge_th)
    969 def prune(self, node_th=1e-2, edge_th=3e-2):
--> 970     self = self.prune_node(node_th, log_history=False)
    971     #self.prune_node(node_th, log_history=False)
    972     self.forward(self.cache_data)

File c:\\Users\\hp\\.conda\\envs\\kan\\lib\\site-packages\\kan\\MultKAN.py:942, in MultKAN.prune_node(self, threshold, mode, active_neurons_id, log_history)
    938         model2.symbolic_fun[i].out_dim_mult = num_mult
    940         width_new.append([num_sum, num_mult])
--> 942     model2.act_fun[i] = model2.act_fun[i].get_subset(active_neurons_up[i], active_neurons_down[i])
    943     model2.symbolic_fun[i] = self.symbolic_fun[i].get_subset(active_neurons_up[i], active_neurons_down[i])
    945 model2.cache_data = self.cache_data

File c:\\Users\\hp\\.conda\\envs\\kan\\lib\\site-packages\\kan\\KANLayer.py:305, in KANLayer.get_subset(self, in_id, out_id)
    283 def get_subset(self, in_id, out_id):
    284     '''
    285     get a smaller KANLayer from a larger KANLayer (used for pruning)
    286     
   (...)
    303     (2, 3)
    304     '''
--> 305     spb = KANLayer(len(in_id), len(out_id), self.num, self.k, base_fun=self.base_fun, device=self.device)
    306     spb.grid.data = self.grid[in_id]
    307     spb.coef.data = self.coef[in_id][:,out_id]

File c:\\Users\\hp\\.conda\\envs\\kan\\lib\\site-packages\\kan\\KANLayer.py:132, in KANLayer.__init__(self, in_dim, out_dim, num, k, noise_scale, scale_base, scale_sp, base_fun, grid_eps, grid_range, sp_trainable, sb_trainable, save_plot_data, device, sparse_init)
    129 else:
    130     mask = 1.
--> 132 scale_base = scale_base.to(device)
    133 self.scale_base = torch.nn.Parameter(torch.ones(in_dim, out_dim, device=device) * scale_base * mask).requires_grad_(sb_trainable)  # make scale trainable
    134 #else:
    135 #self.scale_base = torch.nn.Parameter(scale_base.to(device)).requires_grad_(sb_trainable)

AttributeError: 'float' object has no attribute 'to'"

Best regards.

Diego. Just comment out the erroneous line in KANLayer.py, and it will work

[KindXiaoming]

Hi, please find the most up-to-date hellokan here: https://github.com/KindXiaoming/pykan/blob/master/hellokan.ipynb model.prune() should be model = model.prune()

[lexmar07]

Thanks for the reply! Unfortunately, it does not help much. I'm facing the analogous problem with the latest hello-kan:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
[<ipython-input-11-d30a6490a38c>](https://localhost:8080/#) in <cell line: 1>()
----> 1 model = model.prune()
      2 model.plot()

3 frames
[/content/pykan/kan/MultKAN.py](https://localhost:8080/#) in prune(self, node_th, edge_th)
    968 
    969     def prune(self, node_th=1e-2, edge_th=3e-2):
--> 970         self = self.prune_node(node_th, log_history=False)
    971         #self.prune_node(node_th, log_history=False)
    972         self.forward(self.cache_data)

[/content/pykan/kan/MultKAN.py](https://localhost:8080/#) in prune_node(self, threshold, mode, active_neurons_id, log_history)
    940                 width_new.append([num_sum, num_mult])
    941 
--> 942             model2.act_fun[i] = model2.act_fun[i].get_subset(active_neurons_up[i], active_neurons_down[i])
    943             model2.symbolic_fun[i] = self.symbolic_fun[i].get_subset(active_neurons_up[i], active_neurons_down[i])
    944 

[/content/pykan/kan/KANLayer.py](https://localhost:8080/#) in get_subset(self, in_id, out_id)
    303         (2, 3)
    304         '''
--> 305         spb = KANLayer(len(in_id), len(out_id), self.num, self.k, base_fun=self.base_fun, device=self.device)
    306         spb.grid.data = self.grid[in_id]
    307         spb.coef.data = self.coef[in_id][:,out_id]

[/content/pykan/kan/KANLayer.py](https://localhost:8080/#) in __init__(self, in_dim, out_dim, num, k, noise_scale, scale_base, scale_sp, base_fun, grid_eps, grid_range, sp_trainable, sb_trainable, save_plot_data, device, sparse_init)
    130             mask = 1.
    131 
--> 132         scale_base = scale_base.to(device)
    133         self.scale_base = torch.nn.Parameter(torch.ones(in_dim, out_dim, device=device) * scale_base * mask).requires_grad_(sb_trainable)  # make scale trainable
    134         #else:

AttributeError: 'float' object has no attribute 'to'

PS 1) also tried different version: # v0.2.1, v0.1.2, the problem repeats 2) a small suggestion: maybe a lot of these compatibility issues will vanish if You can provide a compilable google - colab hellokan, like the one here: https://colab.research.google.com/drive/1YOU7AifdYieMWK2hDfKjlN7l6_n6BkvV?usp=sharing

[KindXiaoming]

line 132 'scale_base = scale_base.to(device)' is the problem, please comment it out. This is already updated in the github repo (so you can also pull the latest github repo), but not on pypi (thanks for your note, I just realized this).