[distGB] graphbolt graph edge's mask will be filled with 0 if these edges have no mask initial #7846
Conversation
|
@thvasilo ，this pr can fix the graphbolt's mask issue in graphstorm. |
There was a problem hiding this comment.
I think the code looks good now. One suggestion on the naming, padding implies we pad an existing tensor with additional elements until it reaches a specifc length, see https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html
What we are doing here is filling a tensor with a specifc value, which used to be 0 and is now 1. So I would suggest we change the name of the argument and variables to fill instead padding.
| gb_padding : int, optional | ||
| The padding value for GraphBolt partitions' new edge_attributes if the attributes in DistGraph are None. | ||
| e.g. prob/mask-based sampling. | ||
| Only when the mask of one edge is set as 1, the edge will be sampled. |
There was a problem hiding this comment.
more details about why we set default value as 1 while set 0 for node counterpart
There was a problem hiding this comment.
I have set the gb_padding as 1 in NodeCollator and changed the comment correspondingly.
python/dgl/distributed/dist_graph.py
Outdated
| attr_data = torch.zeros(num_edges, dtype=data_type) | ||
|
|
||
| # Padding is used here to fill missing edge attributes (e.g., 'prob' or 'mask') for certain edge types. | ||
| # In DGLGraph, some edges may not have attributes or their values could be None. |
There was a problem hiding this comment.
not clear enough. we should make it clear that the None edges will be sampled in DGL, not sampled while in GB.
There was a problem hiding this comment.
Ok, I have enriched the comment and explained the None edges will be sampled in DGLGraph.
| gb_padding : int, optional | ||
| The padding value for GraphBolt partitions' new edge_attributes. | ||
| e.g. some edges of specific types have no mask, the mask will be set as gb_padding. | ||
| the edge will not be sampled if the mask is 0. |
There was a problem hiding this comment.
| the edge will not be sampled if the mask is 0. | |
| An edge will not be sampled if the mask is 0. |
There was a problem hiding this comment.
OK, I have changed the comment.
|
图机器学习算法库作业已收到！
|
Add a new parameter "padding" to add_edge_attribute, the edge's mask will be filled with padding rather than 0.
When use DistEdgeDataloader the padding will be set to 1 to fully sample the edge with no mask.