Expected behavior of `example_unsupervised.sh` about graphsage HOT 4 CLOSED

Hi,
There's a flag (validate_iter) that sets the number of iterations between two validation runs. It's set to 5000 by default for efficiency reasons when running on large dataset. For this small dataset, you could set a smaller number.

from graphsage.

Ah great, thanks -- on the example data, it looks like the default value is so high it never gets triggered.

Related question: I'm trying to reproduce the unsupervised results on the reddit and ppi datasets. If I run:

python -m graphsage.unsupervised_train \
    --train_prefix ./data/reddit/reddit \
    --model graphsage_mean \
    --learning_rate 0.001

python ./eval_scripts/reddit_eval.py

should I get something close?

from graphsage.

Just pushed a change to have "--validate_iter 10" for the unsupervised example.

And yeah, those settings should get you in the ballpark. I recommend starting with a learning rate of 0.0001 though (this was usually optimal iirc), and the "big" models will probably perform a bit better (though with longer runtime).

from graphsage.

Cool thanks -- will close

from graphsage.