Saving GraphSAGE model -- ValueError: GraphDef cannot be larger than 2GB about graphsage HOT 8 CLOSED

I was able to save the models to disk once I realized the graph adjacency and the features were also being written, as they are defined via tf.Constant. This was pushing the size of the graph over the limit. To avoid saving these constants, I used tf.placeholder in their place instead.

The change for supervised models can be seen in this commit in my fork. Happy to put in a pull request if you think there are no obvious downsides.

from graphsage.

Hi, sorry for the late reply since we are at conference.

What you said is indeed the case. Using placeholder for adj will solve the problem. However, we were concerned that there would be significant overhead when feeding in the same np array at each iteration.
In fact when I was using GraphSAGE for extremely large datasets (10^9 scale), I set the adj as a variable, and feed it in only at the start of training/validation/test.

But of course ideally we would want tensorflow to avoid the overhead when feeding the same array without modifications, so that placeholder would be the most elegant solution.

Rex

from graphsage.

@RexYing what is the final word on this? I suppose our current implementation in master branch is the simplest, but has the issue that the adjacency matrix is treated as part of the graph. I don't think having the adjacency as a placeholder is best because of the data I/O overhead, but maybe we should have a helper function or something that makes it easier to save the model?

from graphsage.

Shouldn't it be enough to assign the adjacency matrix placeholder once, before starting the training loop? Further I/O could be avoided also for validation/test sets if another instance of the learned model is created, and only the learned weights are copied to this instance when validation/testing is to be performed. Copying the weights should be orders of magnitude cheaper than swapping the adjacency matrix.

EDIT: I guess a clarification is in order. As far as I can tell from observed behaviour, I think if you do

adjacency_ph = tf.placeholder(tf.int32, shape=(size,), name='adjacency_ph')
adjacency = tf.Variable(adjacency_ph, trainable=False, name='adjacency')

and then outside the training loop run the assignment, e.g.: sess.run(tf.global_variables_initializer(), feed_dict={adjacency_ph: adj}) you are 1) only assigning to adjacency once, not on every iteration of the training loop and 2) still preventing the value of adjacency being written to the GraphDef by using the placeholder.

from graphsage.

@RexYing do you have thoughts on @laurimi's solution above? You have more experience dealing with this I/O overhead issue, but her solution seems to check out and it would be nice to not have the adjacency as part of the GraphDef.

(Btw, I am traveling and won't be able to test and push anything to master for at least a few days if this solution should be merged in).

from graphsage.

@laurimi I tested and it does work. Thanks for pointing it out. Would be happy to take the pull request.
Or I could do it myself if you are busy.

Thanks
Rex

from graphsage.

@RexYing Pull request sent. I decided to only apply the placeholder it for the adjacency matrix. It should be possible to do the same for the features as well, but I did not want to mess around with the model files since I don't feel so confident for now making modifications there.

from graphsage.

Thank very much for the contribution!

from graphsage.