TransE error: "ValueError: One of the provided node embedding computed with the TransE method contains NaN values."

[realmarcin]

When generating embeddings for KG-Microbe (KGX edge file from KG-Hub) using TransE, the following error was observed:

ValueError Traceback (most recent call last)
in
----> 1 embedding = model.fit_transform(kg)

~/Library/Python/3.7/lib/python/site-packages/cache_decorator/cache.py in wrapped(*args, **kwargs)
595 if not cache_enabled:
596 self.logger.info("The cache is disabled")
--> 597 result = function(*args, **kwargs)
598 self._check_return_type_compatability(result, self.cache_path)
599 return result

~/Library/Python/3.7/lib/python/site-packages/embiggen/utils/abstract_models/abstract_embedding_model.py in fit_transform(self, graph, return_dataframe, verbose)
164 graph=graph,
165 return_dataframe=return_dataframe,
--> 166 verbose=verbose
167 )
168

~/Library/Python/3.7/lib/python/site-packages/embiggen/embedders/ensmallen_embedders/transe.py in _fit_transform(self, graph, return_dataframe, verbose)
112 embedding_method_name=self.model_name(),
113 node_embeddings= node_embedding,
--> 114 edge_type_embeddings= edge_type_embedding,
115 )
116

~/Library/Python/3.7/lib/python/site-packages/embiggen/utils/abstract_models/embedding_result.py in init(self, embedding_method_name, node_embeddings, edge_embeddings, node_type_embeddings, edge_type_embeddings)
76 if np.isnan(numpy_embedding).any():
77 raise ValueError(
---> 78 f"One of the provided {embedding_list_name} "
79 f"computed with the {embedding_method_name} method "
80 "contains NaN values."

ValueError: One of the provided node embedding computed with the TransE method contains NaN values.

I am attaching a jupyter notebook to reproduce the problem.
load_graph_and.ipynb.zip

The input edge file is here: https://kg-hub.berkeleybop.io/kg-microbe/current/kg-microbe.tar.gz

[LucaCappelletti94]

Hello Marcin, in the provided Jupyter you are loading the edge list using:

kg = Graph.from_csv(
    edge_path="./merged-kg_edges.tsv",
   sources_column_number=0,
   edge_list_edge_types_column_number=1,
   destinations_column_number=2,
   directed=False,
   name="kg-microbe")

but this will load the id column as source nodes, since the file is not a triples file like the other one.

If you load the graph from the automatic retrieval (which points to the same edge list) you should not encounter any issue:

from grape.datasets.kghub import KGMicrobe
kg = KGMicrobe()

Nonetheless, it is interesting that this causes this peculiar issue, I will look into it.

[sanyabt]

Hi @LucaCappelletti94, I ran into the same issue after running the embeddings on my graph - TransE model run after ntriples file loaded. Here is a screenshot of the graph loading and the error.

ValueError Traceback (most recent call last)
Input In [17], in <cell line: 1>()
----> 1 embedding = model.fit_transform(npkg)

File ~/.conda/envs/faers-embed/lib/python3.8/site-packages/cache_decorator/cache.py:597, in Cache._decorate_function..wrapped(*args, **kwargs)
595 if not cache_enabled:
596 self.logger.info("The cache is disabled")
--> 597 result = function(*args, **kwargs)
598 self._check_return_type_compatability(result, self.cache_path)
599 return result

File ~/.conda/envs/faers-embed/lib/python3.8/site-packages/embiggen/utils/abstract_models/abstract_embedding_model.py:163, in AbstractEmbeddingModel.fit_transform(self, graph, return_dataframe, verbose)
149 if graph.has_disconnected_nodes():
150 warnings.warn(
151 (
152 f"Please be advised that the {graph.get_name()} graph "
(...)
160 )
161 )
--> 163 result = self._fit_transform(
164 graph=graph,
165 return_dataframe=return_dataframe,
166 verbose=verbose
167 )
169 if not isinstance(result, EmbeddingResult):
170 raise NotImplementedError(
171 f"The embedding result produced by the {self.model_name()} method "
172 f"from the library {self.library_name()} implemented in the class "
173 f"called {self.class.name} does not return an Embeddingresult "
174 f"but returns an object of type {type(result)}."
175 )

File ~/.conda/envs/faers-embed/lib/python3.8/site-packages/embiggen/embedders/ensmallen_embedders/transe.py:111, in TransEEnsmallen._fit_transform(self, graph, return_dataframe, verbose)
102 node_embedding = pd.DataFrame(
103 node_embedding,
104 index=graph.get_node_names()
105 )
106 edge_type_embedding = pd.DataFrame(
107 edge_type_embedding,
108 index=graph.get_unique_edge_type_names()
109 )
--> 111 return EmbeddingResult(
112 embedding_method_name=self.model_name(),
113 node_embeddings= node_embedding,
114 edge_type_embeddings= edge_type_embedding,
115 )

File ~/.conda/envs/faers-embed/lib
/python3.8/site-packages/embiggen/utils/abstract_models/embedding_result.py:77, in EmbeddingResult.init(self, embedding_method_name, node_embeddings, edge_embeddings, node_type_embeddings, edge_type_embeddings)
74 numpy_embedding = embedding
76 if np.isnan(numpy_embedding).any():
---> 77 raise ValueError(
78 f"One of the provided {embedding_list_name} "
79 f"computed with the {embedding_method_name} method "
80 "contains NaN values."
81 )
83 self._embedding_method_name = embedding_method_name
84 self._node_embeddings = node_embeddings

ValueError: One of the provided node embedding computed with the TransE method contains NaN values.

[LucaCappelletti94]

Hello @sanyabt! Fortunately, most likely your error is only caused by the fact that the graph is loaded as direct and there may be trap nodes there. Could you try to run kg.get_trap_nodes_number()? If there are any, that is the cause and I have fixed it yesterday (I had forgotten about this corner case).

[LucaCappelletti94]

Resolved also the corner case presented in the other peculiar undirected graph topology.

[sanyabt]

Thank you! Do we need to update or reinstall grape for the fix?

[LucaCappelletti94]

It will be necessary, but currently, @zommiommy is working on @pnrobinson Printer issue. As soon as that is fixed, we will run the build procedure and deploy the updated version on PyPI. I will notify you here when we do so.

We have added in the READMEs links to the telegram, discord and Twitter accounts to easily reach us.

[LucaCappelletti94]

Deployed updated versions on Pypi, GraPE version 0.1.3.