InvalidArgument: [pyAgrum] Invalid argument: Potential to copy has not the same domainSize.

[kenneth-lee-ch]

I am not sure why I am getting this error as the number of states is the same for the bn object and the data.

tr_data_graph.csv

Below is how I got the error.

import pyAgrum as gum
sach_bn = gum.fastBN('PKA->Raf->Mek->Erk->Akt;'
                     'PKA->Mek;'
                     'PKA->Erk;'
                     'PKA->Akt;'
                     'PKA->Jnk;'
                     'PKA->P38;'
                    )

bn_without_S= gum.BayesNet(sach_bn)
learner = gum.BNLearner("tr_data_graph.csv")
learner.fitParameters(bn_without_S)

[phwuil]

You created a BN where all variables are binary. I guess this is not the case in your CSV.

[kenneth-lee-ch]

But the data is indeed binary, that's why I got so confused...I also attached the data file on the thread.

…

On Thu, Apr 13, 2023, 5:08 AM Pierre-Henri Wuillemin < ***@***.***> wrote: You created a BN with all variables are binary. I guess this is not the case in your CSV. [image: image] <https://user-images.githubusercontent.com/5980636/231711438-d25ed193-dd67-4eba-9316-bed00d282ea9.png> — Reply to this email directly, view it on GitHub <#13 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AF3Z3CXGI4BJBEIYJRMBRR3XA67AZANCNFSM6AAAAAAW4XDRMI> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[phwuil]

PKA has only one value (always 0) ... Maybe we should warn in this case ... or automatically add a second value ...

[kenneth-lee-ch]

I thought the fitparameter should just count the frequency based on the given finite sample and look at the BN object for the exact domain size. In this case, wouldn't it make sense to just return 0 for the entry in CPT of PKA state 1? Kenneth

…

On Thu, Apr 13, 2023, 9:51 AM Pierre-Henri Wuillemin < ***@***.***> wrote: [image: image] <https://user-images.githubusercontent.com/5980636/231779233-5cce5fc9-ac53-4c84-b8a6-18f0089c4686.png> PKA has only one value (always 0) ... Maybe we should warn in this case ... or automatically add a second value ... — Reply to this email directly, view it on GitHub <#13 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AF3Z3CRG6Q65TJOZCAWNUXDXBAAENANCNFSM6AAAAAAW4XDRMI> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[phwuil]

How can the BNLearner know how many modalities are not present in the database ? Could PKA be equal to 0, 2, 3 ?
So by default, the BNLearner takes the exact domain size of the data for the CPT it learns. So PKA is learned with only one value. And when fitParameters tries to fill a cpt of size 2*... with a table of size 1*... , it ends up with the error of domainSize...

However, there is a solution, you can bypass the automatic type induction of the BNLearner by explicitly giving it the exact type of each variable. You can add a BN as a parameter when constructing the BNLearner. This BN is not used for the structure but is used as a dictionary of all the variables and their types :

[phwuil]

you can even pass the same bn :

bn=gum.fastBN("Erk->Akt->PKA;Mek->Jnk->Raf->P38[5]") 
learner=gum.BNLearner("tr_data_graph.csv",bn) # use bn as template for the variables
learner.fitParameters(bn)
gnb.flow.row(bn.cpt("PKA"),bn.cpt("P38"))