Visualize the model

[yuyiyi]

How do we visualize the model architecture in CEBRA? eg. using make_dot

[stes]

Hi @yuyiyi , the cebra encoder is a regular PyTorch Module, ie any visualisation library for such modules would work.

However, many standard models only have a few layers without a complicated architecture, so it might also make sense to look at their string representation and go from there with a manual visualisation as well.

Please let me know if that addresses the question already, or if you're looking for something else.

[yuyiyi]

Hi, @stes ,
Thanks for the quick response! I was trying to print the network architecture of the CEBRA model. Usually I can use print(net) to see what's inside the network, linear layer, mlp and so on. ... However, print(cebra model) does not give me that information. It seems that the pytorch module is wrapped inside the custom application. How to get information about network architecture of a cebra model?

[stes]

Hi @yuyiyi , yes this is possible, you need to follow sklearn-logic for this (i.e., you need to call the fit() function once, e.g. by setting the # of steps very low).

>>> import cebra
>>> import numpy as np

>>> X = np.random.randn(1000, 50)
>>> model = cebra.CEBRA(max_iterations=10)
>>> model.fit(X)
CEBRA(max_iterations=10)
>>> print(model)
CEBRA(max_iterations=10)
>>> print(model.model_)
Offset0Model(
  (net): Sequential(
    (0): Flatten(start_dim=1, end_dim=-1)
    (1): Linear(in_features=50, out_features=32, bias=True)
    (2): GELU(approximate='none')
    (3): Linear(in_features=32, out_features=32, bias=True)
    (4): GELU(approximate='none')
    (5): Linear(in_features=32, out_features=16, bias=True)
    (6): GELU(approximate='none')
    (7): Linear(in_features=16, out_features=8, bias=True)
    (8): _Norm()
    (9): Squeeze()
  )
)

does that answer your question? (then please mark as resolved :) )

[stes]

flagging this as resolved, hope ok!