Implment get_as_torch_geometric
#1200
Conversation
mewim
force-pushed
the
get-as-pyg
branch
4 times, most recently
from
252e441
to
d086ece
Compare
mewim
force-pushed
the
get-as-pyg
branch
6 times, most recently
from
800e65a
to
a4b6c62
Compare
| LIST = "LIST" | ||
| NODE = "NODE" | ||
| REL = "REL" | ||
| NODE_ID = "NODE_ID" |
There was a problem hiding this comment.
Update NODE_ID once Ziyi's PR is in.
|
|
||
| PRIMARY_KEY_SYMBOL = "(PRIMARY KEY)" | ||
| LIST_SYMBOL = "[]" | ||
| result_str = self.query_result.connection._connection.get_node_property_names( |
There was a problem hiding this comment.
This comment is beyond this PR, but let me document it here:
The parsing here makes me think that our cpp interface string Connection::getNodePropertyNames(const string& tableName) is incorrect. Instead of returning a string, it's better to return vector<string>, each inside which is a property name. The caller (shell) shall be responsible to organize them in a printable way.
| continue | ||
|
|
||
| # Ignore primary key | ||
| if node_property_names[prop_name]["is_primary_key"]: |
There was a problem hiding this comment.
Why not put primary_key prop_name also into the ignored properties?
There was a problem hiding this comment.
Ignored property is not read at all, but for primary key, it still needs to be read as we need to maintain the pos_to_primary_key_dict. I have changed the comment to Read primary key but do not add it to the node properties.
| from enum import Enum | ||
|
|
||
|
|
||
| class Type(Enum): |
There was a problem hiding this comment.
Is there a way that we don't need to maintain a separate mapping in Python?
There was a problem hiding this comment.
As discussed in #1189, pybind11 has a memory leakage bug when binding enum class. We can change this Enum to binding after the bug is fixed.
ray6080
added
the
feature
This PR implements a function
get_as_torch_geometric()which converts the query result into atorch_geometric.dataobject.Principles:
NODEandRELs are converted. Selected properties, such asa.isWorkeris ignored.INT64andBOOLtypes. Other types are ignored. Multi-dimensional lists are supported as long as the dimension is consistent and there is no sub-list of different length.torch.Tensor.