State sharing between graphs

[dmlc]

State sharing between graphs

[zzhang-cn]

I am not familiar enough with this problem, but like the way networkx follows Python's style so that users, if they are confused, have some "prior" suspect. For the "independent shallow" that networkx provides, does it raises issue with our current models, or does it work okay and we just need to warn the user about the potential problems?

And, the worst is to ask users to do deep copy and maintain the states themselves. Correct?

[GaiYu0]

After an "independent shallow" copy, a graph and its duplicate will share representations. However, if we update the representations of either the graph or its duplicate with current API, they will not share representations any more.

# Assume that g.nodes[n]['x'] == 0 for every node n
h = g.copy()
h.register_message_func(lambda src, trg, _: None)
h.register_update_func(lambda node_reprs, edge_reprs: 1)
h.update_all()

Now, g.nodes[n]['x'] == 0 and h.nodes[n]['x'] == 1 for every node n. To sync between g and h, we have to write

for n, d in g.nodes.items():
  d['x'] = h.nodes[n]['x']

In this case sync is not tricky. However, sync between graphs may become a burden for user, e.g. line 153 - 156 and line 164 - 167 in line_graph/gnn.py. Syncing by iterating through all nodes is not very efficient either. So the question is can we help users to sync between graphs automatically?

I hope this makes more sense.

[zzhang-cn]

I played with networkx's copy a little and understand what it is now:

For the time being, I'd suggest simply add a sync_state(g_src, g_tgt, name="all") to make it easy for the user, and deal with performance issue later.

[GaiYu0]

I would suggest the following (implemented by overriding networkx methods):

# Assume that isinstance(g, DGLGraph), g.nodes[0]['h'] == 12, g.edges[0, 1]['h'] == 12
g_c = g.copy(share=['h'])
g_c.nodes[0]['h'] = 13 # Now g.nodes[0]['h'] == 13

g_c = DGLGraph()
g_c.add_nodes_from(g.nodes, share=['x'])
g_c.nodes[0]['h'] = 13 # Now g.nodes[0]['h'] == 13

g_c = DGLGraph()
g_c.add_edges_from(g.edges, share=['x'])
g_c.edges[0, 1]['h'] = 13 # Now g.edges[0, 1]['h'] == 13

This interface is less flexible compared to sync_state because states are always synced after copy, add_nodes_from and add_edges_from. However, it provides us with more information about which attributes will be synced and thus may benefit optimization.

[zzhang-cn]

Nice. Code link?

…

On Mon, Jun 25, 2018 at 6:13 PM, GaiYu0 ***@***.***> wrote: I would suggest the following (implemented by overriding networkx methods): # Assume that isinstance(g, DGLGraph), g.nodes[0]['h'] == 12, g.edges[0, 1]['h'] == 12 g_c = g.copy(share=['h']) g_c.nodes[0]['h'] = 13 # Now g.nodes[0]['h'] == 13 g_c = DGLGraph() g_c.add_nodes_from(g.nodes, share=['x']) g_c.nodes[0]['h'] = 13 # Now g.nodes[0]['h'] == 13 g_c = DGLGraph() g_c.add_edges_from(g.edges, share=['x']) g_c.edges[0, 1]['h'] = 13 # Now g.edges[0, 1]['h'] == 13 This interface is less flexible compared to sync_state because states are always synced after copy, add_nodes_from and add_edges_from. However, it provides us with more information about which attributes will be synced and thus may benefit optimization. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#13 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIKVUVCcEqlw7LBP3EZuRDDsVG942mQuks5uALfNgaJpZM4U06mo> .

-- *Zheng Zhang <https://shanghai.nyu.edu/academics/faculty/zheng-zhang> 张峥* Professor of Computer Science, NYU Shanghai 上海纽约大学计算机系教授 Room 1118, 1555 Century Avenue Pudong New District, Shanghai 200122, China **上海市浦东新区世纪大道1555号1118室, 200122 T Shanghai office 上海办公室 [phone] T China mobile **手机 [phone] E 电子邮件 [email protected] <[email protected]>

[jermainewang]

I suggest writing some examples such as the line-graph-graph, the multi-radius-graph.

…

On Mon, Jun 25, 2018, 8:51 PM Zheng Zhang ***@***.***> wrote: Nice. Code link? On Mon, Jun 25, 2018 at 6:13 PM, GaiYu0 ***@***.***> wrote: > I would suggest the following (implemented by overriding networkx methods): > > # Assume that isinstance(g, DGLGraph), g.nodes[0]['h'] == 12, g.edges[0, 1]['h'] == 12 > g_c = g.copy(share=['h']) > g_c.nodes[0]['h'] = 13 # Now g.nodes[0]['h'] == 13 > > g_c = DGLGraph() > g_c.add_nodes_from(g.nodes, share=['x']) > g_c.nodes[0]['h'] = 13 # Now g.nodes[0]['h'] == 13 > > g_c = DGLGraph() > g_c.add_edges_from(g.edges, share=['x']) > g_c.edges[0, 1]['h'] = 13 # Now g.edges[0, 1]['h'] == 13 > > This interface is less flexible compared to sync_state because states are > always synced after copy, add_nodes_from and add_edges_from. However, it > provides us with more information about which attributes will be synced and > thus may benefit optimization. > > — > You are receiving this because you commented. > Reply to this email directly, view it on GitHub > <#13 (comment)>, or mute > the thread > < https://github.com/notifications/unsubscribe-auth/AIKVUVCcEqlw7LBP3EZuRDDsVG942mQuks5uALfNgaJpZM4U06mo > > . > -- *Zheng Zhang <https://shanghai.nyu.edu/academics/faculty/zheng-zhang> 张峥* Professor of Computer Science, NYU Shanghai 上海纽约大学计算机系教授 Room 1118, 1555 Century Avenue <https://maps.google.com/?q=1118,+1555+Century+Avenue+%0D%0A+%0D%0APudong+New+District,+Shanghai+200122&entry=gmail&source=g> <https://maps.google.com/?q=1118,+1555+Century+Avenue+%0D%0A+%0D%0APudong+New+District,+Shanghai+200122&entry=gmail&source=g> Pudong New District, Shanghai 200122 <https://maps.google.com/?q=1118,+1555+Century+Avenue+%0D%0A+%0D%0APudong+New+District,+Shanghai+200122&entry=gmail&source=g>, China **上海市浦东新区世纪大道1555号1118室 <https://maps.google.com/?q=%E4%B8%AD%E5%9B%BD%E4%B8%8A%E6%B5%B7%E5%B8%82%E6%B5%A6%E4%B8%9C%E6%96%B0%E5%8C%BA%E4%B8%96%E7%BA%AA%E5%A4%A7%E9%81%931555%E5%8F%B71118%E5%AE%A4&entry=gmail&source=g>, 200122 T Shanghai office 上海办公室 [phone] T China mobile **手机 [phone] E 电子邮件 ***@***.*** ***@***.***> — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#13 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AD3qZahos7rFZvxgzp0NVGdR3QBHlLIIks5uAM7KgaJpZM4U06mo> .

[GaiYu0]

Does the example need to be runnable?

[jermainewang]

Let's see how the APIs look like first.

…

On Mon, Jun 25, 2018, 11:02 PM GaiYu0 ***@***.***> wrote: Does the example need to be runnable? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#13 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AD3qZTmBUEIsYlZVu2890efzCSeASMSJks5uAO2KgaJpZM4U06mo> .

[jermainewang]

Current solution for state sharing is using FrameRef (PR #39 ). More examples are coming on this syntax.