Batch Prediction from Trained deepFM model is very slow

[BhaveshBhansali]

I trained deepFM model using approx. 4.5 million rows consisting of approx. 1.5 million unique users and 900k unique items.

While batch predicting for approx. 20 users at a time, my jupyter kernel gets killed. I am using 64GM machine.

res = model.recommend_user(user=list_of_users, n_rec=20)

Could you please give tip to make batch inference faster?

Many thanks in advance.

[massquantity]

The model will make prediction for every item in recommend_user, which means it will compute scores for 900k items in one recommendation. A batch of 20 users will compute 18 million items at once, no wonder your jupyter kernel gets killed.

Faster? Maybe using a better GPU?

[BhaveshBhansali]

Thank you for your answer.

Is there explicit parameter to utilize GPU while training and inference?

[massquantity]

TensorFlow will use GPU automatically if you have one. Make sure you have enough GPU memory to compute 900k items.

[BhaveshBhansali]

Thank you for your answer.