ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type numpy.ndarray)

[herman-nside]

I'm having trouble running my Spektral model. I am aware of this issue which might have something to do with Windows paths, but I am on a Mac, and I think this is a different issue.

My code is as follows:

import spektral
import tensorflow as tf
from dataset_class import GNN_Dataset
from spektral.data.dataset import Dataset
from spektral.layers import GraphSageConv
from tensorflow.keras.layers import Dense
from tensorflow.keras.models import Model
from tensorflow.keras.regularizers import l1_l2

class MyFirstGNN(Model):

    def __init__(self):
        super().__init__()
        self.X_1 = GraphSageConv(512,
                                 activation='relu',
                                 kernel_regularizer=l1_l2(l1=0.01, l2=0.01),
                                 bias_regularizer=l1_l2(l1=0.01, l2=0.01),
                                 kernel_initializer=tf.keras.initializers.GlorotUniform(),
                                 )

        self.last = Dense(14, activation='softmax')  # (X_4)

    def call(self, inputs):
        out = self.X_1()(inputs)
        out = self.last(out)
        return out

dataset = GNN_Dataset('/dataset/sn/')
batch_loader = spektral.data.loaders.BatchLoader(dataset, batch_size=2)
model = MyFirstGNN()
model.compile(optimizer='Adam', loss='binary_cross_entropy')
model.fit(batch_loader.load(), steps_per_epoch=batch_loader.steps_per_epoch, epochs=10)

I get the following error:

test.py 48 <module>
model.fit(batch_loader.load(), steps_per_epoch=batch_loader.steps_per_epoch, epochs=10)

training.py 108 _method_wrapper
return method(self, *args, **kwargs)

training.py 1049 fit
data_handler = data_adapter.DataHandler(

data_adapter.py 1105 __init__
self._adapter = adapter_cls(

data_adapter.py 788 __init__
peek = _process_tensorlike(peek)

data_adapter.py 1021 _process_tensorlike
inputs = nest.map_structure(_convert_numpy_and_scipy, inputs)

nest.py 635 map_structure
structure[0], [func(*x) for x in entries],

nest.py 635 <listcomp>
structure[0], [func(*x) for x in entries],

data_adapter.py 1016 _convert_numpy_and_scipy
return ops.convert_to_tensor(x, dtype=dtype)

ops.py 1499 convert_to_tensor
ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)

tensor_conversion_registry.py 52 _default_conversion_function
return constant_op.constant(value, dtype, name=name)

constant_op.py 263 constant
return _constant_impl(value, dtype, shape, name, verify_shape=False,

constant_op.py 275 _constant_impl
return _constant_eager_impl(ctx, value, dtype, shape, verify_shape)

constant_op.py 300 _constant_eager_impl
t = convert_to_eager_tensor(value, ctx, dtype)

constant_op.py 98 convert_to_eager_tensor
return ops.EagerTensor(value, ctx.device_name, dtype)

ValueError:
Failed to convert a NumPy array to a Tensor (Unsupported object type numpy.ndarray).

My dataset looks like this:

dataset[0]
Out[2]: Graph(n_nodes=9, n_node_features=36, n_edge_features=None, n_labels=9)

Inside batch_loader, it looks like this:

for i in batch_loader:
    print(i)
    break

    
((array([[[0., 1., 0., ..., 0., 0., 0.],
        [1., 0., 0., ..., 0., 0., 0.],
        [1., 0., 0., ..., 0., 0., 0.],
        ...,
        [0., 1., 0., ..., 0., 0., 0.],
        [1., 0., 0., ..., 0., 0., 0.],
        [0., 0., 0., ..., 0., 0., 0.]],
       [[0., 1., 0., ..., 0., 0., 0.],
        [1., 0., 0., ..., 0., 0., 0.],
        [1., 0., 0., ..., 0., 0., 0.],
        ...,
        [0., 1., 0., ..., 0., 0., 0.],
        [0., 1., 0., ..., 0., 0., 0.],
        [1., 0., 0., ..., 0., 0., 0.]]]), array([[[0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0],
        [0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 0],
        [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0],
        [0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0],
        [0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0],
        [1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0],
        [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0],
        [0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0],
        [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],
       [[0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0],
        [0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1],
        [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0],
        [0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0],
        [0, 1, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0],
        [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0],
        [0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0],
        [0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
        [0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0],
        [1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 0, 0],
        [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]])), array([array([0., 1., 1., 0., 0., 1., 1., 0., 0., 1., 0., 1., 0., 0.]),
       array([1., 1., 1., 0., 0., 1., 1., 0., 1., 1., 0., 1., 1., 1., 0.])],
      dtype=object))

My first thought is that this error is not Spektral, but something I'm doing wrong. Is there any way to avoid it or is there a way to cast it (either A or y or both) as a tensor so before fitting?

[danielegrattarola]

Hi,

the first entry in y has 14 entries while the second entry has 15.
BatchLoader only supports graph-level labels (meaning that labels do not get zero-padded -- that would not make sense) so all labels should have the same shape.

In this specific case, the error is because the Loader tries to create a jagged array and Numpy automatically converts that array to an object dtype, which is not supported by tensorflow.
The solution is to create labels that have all the same shape.

I also see another problem with your code: GraphSage does not support batch mode.
It seems that you're trying to do node-level prediction anyway, so BatchLoader will not be usable in your case. You should try to convert your code yo use a DisjointLoader (and set node_level=True when you instantiate the loader).
This should solve your current problem AND the one you would have after solving the current one.

Cheers,
Daniele

[herman-nside]

Thank you very much! Yes, most of these problems stem from me changing from the Disjoint loader to the Batch loader. My concern with the disjoint loader is that I have networks of different sizes. Having a fixed step-size over the resultant disjoint matrix will cut off some matrices halfway, and in other cycles will combine more than one matrix (of more than one graph). Is that a valid concern? Is there something else I can do to mitigate this effect?

[danielegrattarola]

It would be a valid concern if you first create a big disjoint matrix and then cut it up to create mini-batches, but that is the "magic" of the DisjointLoader. It will create a disjoint graph on the fly for each mini-batch, so there should be no problem.

You only need to change this line:

batch_loader = spektral.data.loaders.BatchLoader(dataset, batch_size=2)

to

disjoint_loader = spektral.data.loaders.DisjointLoader(dataset, batch_size=2, node_level=True)

Cheers

[herman-nside]

Thank you very much! I understand that better now.