Error: Dimensions of inputs should match

[guelfaure]

This error appear when i try to train my GNN on graph datasets in order to optimize makespan of RCPSP problem.
The error disapear when i clear the dataset and i keep only few graphs (4-5 not more).
I tried to changed every parameter in train_options and model_description but nothing solved my problem.

model_GNN_beta.zip

Image with the full error message in attachment.

[MiquelFerriol]

Dear @guelfaure,
I do not see any problem with your model description.

Following the error image you posted, it looks like the error was raised during the initialization of the hidden states. I would make sure that all the variables you use in the hidden state creation (duration, ressources_needed, duration_other, ressources_needed_other, ressources_available) have the same length in all the samples.

Otherwise, if possible, please share with us some of the graphs used to reproduce the error and fix it.

Thanks,
Miquel

[guelfaure]

Dear @MiquelFerriol,
Thanks for your answer! I have checked that all the variables have the same length, which was the case. I am still not understanding where come from my error but i try to fix it.

I attach to this message some of the graphs from my dataset.

*After some modifications i deduced that the second dimension of shape[1] (12) from the photo of my first message is equal to the parameter "state_dimension" - 4. But i don't know exactly why. The first dimension is still a mistery.

Thanks in advance for your help!!
train.json

[MiquelFerriol]

Hi @guelfaure,

After reviewing the train.json file, I noticed an error concerning nodes of type 'time' lacking the 'ressources_available' variable. This discrepancy affects several graphs, including:

Graph 3: Node ID: time2
Graph 4: Node ID: time2
Graph 5: Node ID: time2
Graph 19: Node ID: time20
Graph 20: Node ID: time20

For IGNNITION to function properly, all nodes within the same entity type should possess the same parameters and lengths.

To verify the impact, I've provided a "corrected" revision of the train.json file, removing the affected graphs.

train_corrected.json

Having said that, thank you for pointing this out. We'll implement an error-checking mechanism to detect these types of issues in the future to provide the user with a clearer explanation of the error.

Best regards,
Miquel

[guelfaure]

Thank you very much for your help! I missed those nodes because they were added automatically by my algorithm during the creation of the graphs as they were appearing in a link (whereas they should not).