Is it actually using GPU? about fast_tffm HOT 4 CLOSED

@daxiongshu That's weird. Actually I developed this software on machines without GPU card, and everything goes well. Moreover, the core operators I implemented in "cc/" folder is cpu version only. Could you please double-check the gpu usage actually comes from fast_tffm? It will be more helpful if you could identify which operators the gpu usage comes from. I will further investigate it. Thanks!

from fast_tffm.

@daxiongshu FYI, the tensorflow version I installed is "cpu only" version. I haven't run the code on the "gpu" version tensorflow yet. Is it possible that the "gpu" version tensorflow will consume some gpu no matter what mode is chosen?

from fast_tffm.

Thank you for the reply! Yeah, it seems CPU does all heavy lifting. The GPU process only uses 64 MB GPU memory. Previously i thought maybe that is due to the small batch size but when i increase batch size 10x the GPU process still uses 64MB. I installed the GPU version tensorflow and the code runs very well. Thank you!

I saw the nested loops in cc/fm_grad_op.cc so yeah it is CPU-like implementation. Do you happen to know where i can find some guides for implementing GPU ops? I'd like to contribute to this repo with a GPU version.

Thank you very much!

from fast_tffm.

Welcome to contribute :) There are numerous operator implementations in https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/kernels . You can check how they work.

However, when I print the timeline for each operators, it seems that currently the most time consuming operators is the parse operator; therefore, the gpu implementation might not bring any difference. You can verify that. Thanks.

from fast_tffm.