naiqili / dgpg Goto Github PK

Hello!

I am interesting in using DGPG for predicting a graph signal.
I have read the examples and jupyter notebooks and tried to use DGPG with my dataset but I am not able to get the input shapes correct.

My input is a graph with 31 nodes and 7 attributes per node. Considering 10 samples of data my input is trX in 10x31x7.
Then I use 5 inducing points, a batch size of 1 and as number of samples I use the 10 (since it is my number of data samples)
The output training is one attribute of the signal on the graph, therefore trY in 10x 31 x 1.

I tried doing the following:

m_dgpg = DGPG(trX, trY, Z, input_dims=[7], likelihood=Gaussian(), adj=gmat0,
                      agg_op_name='concat3d', ARD=True,
                      is_Z_forward=True, mean_trainable=False, out_mf0=True,
                      num_samples=10, minibatch_size=1,
                      kern_type='RBF'
                      )

The training worked fine.
However when performing predictions this gave me an error of Incompatible shapes.
My input for performing a prediction was 1x31x7

According to the documentation of DGPG the input shape should be:

""": param X: (s1, n, d_in)
:param Y: (s1, n, d_out)
:param Z: (s2, n, d_in) """

On the paper it is written that the input is modified so that the features are concatenated.
From that I understand that I should concatenate the 7 attributes and get an input X of 10x217.
I also tried this using as input_dims=[1] and I got the error
Incompatible shapes: [31,10,217] vs. [31,1,31]

Could you please explain what is the right input to be used for training DGPG and predict_y?

Many thanks!

Where can I find this module jack_utils?