Repository navigation
Why does it take so long to train a model on 10,000 samples? #493
Description
My operating environment is tensorflow 2.15.1 , tfra 0.7.0. I have one question about why is the model training so slow? Is it because dynamic embedding only supports eager mode training?
the following below is the notebook links：
https://drive.google.com/file/d/1WGPhL2MOKCVnrF9dG59kKWi2efXcTwkN/view?usp=sharing
Activity
PhyCoe commentedon Apr 22, 2025
In addition, I found that the training was very slow only in the first epoch. Is it because of the initialization parameters? The training speed of the second epoch was very fast.
MoFHeka commentedon Jun 16, 2025
dynamic embedding supports graph mode training. However, please refer to the implementation in the demo to concatenate all inputs and perform a one-time lookup; otherwise, the overhead of kernel launch will be very large.
If you want to lookup IDs without concatenate directly, please wait for the new version I developed.