search

xiangwang1223 / neural_graph_collaborative_filtering

Neural Graph Collaborative Filtering, SIGIR2019
MIT License
793 stars 262 forks source link

how to test the performence in the test dataset #23

Open namenotexist opened 5 years ago

namenotexist commented 5 years ago

why the training process so slow

wanjunhong0 commented 5 years ago

In my experience, u need a GPU. Tested on 2080ti, 20s for 1 epoch while without GPU, it took 800s 1 epoch. And by the way, 99% the training time is used in sample process.

namenotexist commented 5 years ago

why did i often occur "loss is nan"

xiangwang1223 commented 5 years ago

You can try the following code to modify the BPR loss, in order to avoid NAN:

loss = tf.reduce_sum(tf.nn.softplus(-(pos_result - neg_result)))

namenotexist commented 5 years ago

you mean tf.reduce_mean?

namenotexist commented 5 years ago

why did i can only use 149M GPU memory, which has a full memory of 10GB

xiangwang1223 commented 5 years ago

I have no idea... Have you installed the tensorflow-gpu successfully?

namenotexist commented 5 years ago

I have no idea... Have you installed the tensorflow-gpu successfully? i fix it . the reason is that i install tensorflow-gpu 1.14.0,when i change it to tensorflow-gpu 1.8.0, it works