ZeroDivsion error?

[ziweiyang]

Hi developer,

Thanks for sharing this work. I am trying to reproduce the MeshGraphNet developed by Deepminds in 2021, by using the module.EncodeProcessDecode and module.InteractionNetwroks.

I am using demo.tf2 as prototype and replace self._core with module.InteractionNetwork. I am only using test.tfrecord in original cylinder_flow case for easy testing the model on colab. (There is no global features, and num_nodes/num_edges equal to num_vertices/num_edges in that moment of mesh-space). I define each individual timestep as a GraphTuple. Everything works fine until I started the actual training loop, then I got this error message:

StagingError: in user code:

File "<ipython-input-32-abdb2890eda7>", line 19, in update_step  *
    pred = model(inputs_tr, stats_list[0], stats_list[1], stats_list[2], stats_list[3], params.message_passing_steps)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "<ipython-input-15-c64a739fd414>", line 98, in __call__  *
    latent = self._encoder(inputs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "<ipython-input-15-c64a739fd414>", line 31, in __call__  *
    return self._network(inputs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/graph_nets/_base.py", line 78, in __call__  *
    return self._build(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/graph_nets/modules.py", line 368, in _build  *
    globals=self._global_model(graph.globals))
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/graph_nets/_base.py", line 78, in __call__  *
    return self._build(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/graph_nets/_base.py", line 112, in _build  *
    return self._model(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/sequential.py", line 68, in __call__  *
    outputs = mod(outputs, *args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/nets/mlp.py", line 97, in __call__  *
    inputs = layer(inputs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/linear.py", line 83, in __call__  *
    self._initialize(inputs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/once.py", line 89, in wrapper  *
    _check_no_output(wrapped(*args, **kwargs))
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/utils.py", line 85, in _decorate_unbound_method  *
    return decorator_fn(bound_method, self, args, kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/base.py", line 262, in wrap_with_name_scope  *
    return method(*args, **kwargs)
File "/usr/local/lib/python3.9/dist-packages/sonnet/src/linear.py", line 71, in _initialize  *
    stddev = 1 / math.sqrt(self.input_size)

ZeroDivisionError: float division by zero

I am not sure where I should look at to solve this. Any ideas?

Best, ZY

[alvarosg]

This sounds like one of the vectors of features you are building (e.g. node features or edge features), have zero dimensionality. So this is trying to instantiate a weight of size [0, N], and it is failing when running the initialization for that weight.

[ziweiyang]

Hi @alvarosg , thanks for your comment. I figured it out like what you mentioned here. Thanks for sharing this library, which saves my lots of work.

Best, ZY