How to evaluate a graph regression model with new data?

[alfonsogijon]

Hi, I slightly modified the qm9_ecc.py example (in disjoint mode) to construct a graph regression model to predict the energy of molecules (my dataset is different to QM9). For each molecular geometry I define a graph with graph=Graph(x=x,a=a,e=e,y=energy), which I add to the graph dataset. It works pretty well when I evaluate the model on the test dataset with a disjoint loader, but now I want to use this model to predict the energy on different data.

I would like to make predictions on new molecules (graphs) one by one, because I need to know the energy of a molecule to generate the next molecule. From a molecular geometry I define the data matrices of the graph, (x,a,e,y), but I do not know how to predict its energy with the pre-trained model. I would need something like energy = model(graph, training=False), or energy = model(inputs, training=False) where inputs=[x,a,e,y]. I have tried using a SingleLoader but it does not work. From my graph object or data matrices (x,a,e), how could I evaluate the pre-trained model on it?

Thanks in advance,

Alfonso

[danielegrattarola]

Like you said, you would need to call the model on the matrices describing your graph, so energy = model(inputs, training=False)
You just need to take care of converting the adjacency matrix to a sparse matrix using this function.

[alfonsogijon]

Thanks, but I am still not able to do it. Why the following piece of code does not work?:

x = dataset_tr[10].x
a = dataset_tr[10].a
e = dataset_tr[10].e
y = dataset_tr[10].y

#a = sp_matrix_to_sp_tensor(a)
inputs = [x,a,y]
energy = model(inputs,training=False)

The adjacency matrix is already sparse, because I am reading it from a graph of my training dataset, and it was made sparse before training. This is my model, I do not use the edge features in this case:

def call(self, inputs):
        x, a, i = inputs
        x = self.conv1([x, a])
        x = self.conv2([x, a])
        output = self.global_pool([x, i])
        output = self.dense1(output)
        output = self.dense2(output)
        output = self.dense(output)

An error occurs when doing the first convolution:
'tuple' object has no attribute 'rank'
Call arguments received by layer 'gat_conv' (type GATConv):
• inputs=['tf.Tensor(shape=(228, 4), dtype=float32)', "<228x228 sparse matrix of type ''\n\twith 2364 stored elements in Compressed Sparse Row format>"]
• mask=None

[danielegrattarola]

Sorry, it's actually a bit more involved than that.

Your input must contain the i tensor representing the batch index of the graph. In this case, since you only have one graph, you just need a tensor of all zeros.

In other words:

i = tf.zeros(x.shape[0])
inputs = [x, a, i]

You don't pass y as an input because that's just a target label.

Also, a must be a sparse tensor, not a sparse matrix. So you still need the call to sp_matrix_to_sp_tensor

[alfonsogijon]

Ok, now the following code produces an error in the global pool layer:

a = dataset_tr[10].a
e = dataset_tr[10].e
i = tf.zeros(x.shape[0])

a = sp_matrix_to_sp_tensor(a)
inputs = [x,a,i]
energy = model(inputs,training=False)

The error message is:

Value for attr 'Tindices' of float is not in the list of allowed values: int32, int64
	; NodeDef: {{node SegmentSum}}; Op output:T; attr=T:type,allowed=[DT_FLOAT, DT_DOUBLE, DT_INT32, DT_UINT8, DT_INT16, 6034766930529145842, DT_UINT16, DT_COMPLEX128, DT_HALF, DT_UINT32, DT_UINT64]; attr=Tindices:type,allowed=[DT_INT32, DT_INT64]> [Op:SegmentSum]

Call arguments received by layer 'global_sum_pool' (type GlobalSumPool):
  • inputs=['tf.Tensor(shape=(417, 56), dtype=float32)', 'tf.Tensor(shape=(417,), dtype=float32)']

The only way I can evaluate the energy of a single graph is defining a disjoint loader with a dataset of only that graph. But maybe there is a better way to do it.