Some API corner cases

[dmlc]

Some API corner cases

[GaiYu0]

(1) I agree that updating without message can be completed outside DGLGraph. Thus DGLGraph can refuse to update node without message.
(2) Maybe we can let user decide? Basically there is a tradeoff between storage and re-computation.
(3) I think we need a special object for this, e.g. None. The tricky case is when we tensorize reduce/update function. Currently I use a Python loop to filter out nodes without neighbor, which is very inefficient.

[zzhang-cn]

There’s a case of topdown.py where the root simply observe the entire canvas. I forgot what exactly we did— can Andy comment?

…

Sent from my iPhone On Jul 8, 2018, at 9:32 AM, GaiYu0 <[email protected]> wrote: (1) I agree that updating without message can be completed outside DGLGraph. Thus DGLGraph can refuse to update node without message. (2) Maybe we can let user decide? Basically there is a tradeoff between storage and re-computation. (3) I think we need a special object for this, e.g. None. The tricky case is when we tensorize reduce/update function. Currently I use a Python loop to filter out nodes without neighbor, which is very inefficient. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#20 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIKVUTj7aePS9qwpO5t8CKu9DmPS1eO_ks5uEjQEgaJpZM4VGx7O> .

[BarclayII]

Requested by @zzhang-cn:
In the current implementation my UpdateModule treats the "no message" case as receiving a zero vector. This only happens on the root node of a tree since it is the starting point of the whole message passing process. The restriction that message phase is required for an update should have minimal impact on our model since I can move the root node computation out from the send/recv framework and initialize the states by set_repr().

For the three questions:
(1) What are the downsides (either from our perspective or the user's experience) if we do allow updates with no messages? I think allowing no-message-updates can let the users write more compact code if the nodes do update itself without any incoming messages.
(2) In socket programming, calling sendto would put the message into a buffer while recvfrom clears it. I suggest we stick to this convention as well (i.e. calling a recvfrom will clear the message from the edges). The users can stash the message by themselves if they hope to reuse it. The storage is not an issue here since the underlying tensor libraries already take care of keeping or freeing the message tensors.
(3) If we allow no-message-updates then I feel that this is not a problem.

[jermainewang]

Another corner case:
(4) If there is only one predecessor, do we allow optional reduce phase?

[jermainewang]

Done by PR #25