Model size of DRCT
Hi,
I found that DRCT-L model only has 27M parameters, it confused me a lot that the model size is 230MB.
Also, I have tested that DRCT-L (27M parameters) inference is quite slow than HAT-L (40M parameters), It's been confusing me for days.
Looking forward to your reply, thanks a lot!
Hey
Im not exactly sure why you are focusing on model file size and on parameter count (or what you are confused by exactly).
You could have a look at the databases benchmark, maybe this helps
https://trainner-redux.readthedocs.io/en/latest/benchmarks.html
Its ordered by inference speed: drct-l is slower than hat-l, so that agrees with your experience.
And to show that parameter count does not really tell you much concerning speed: on this table if you scroll to the right the parameter count is listed (and feels unsorted, since this table is sorted after inference speed).
To give one example that parameter count != inference speed, when looking at the 4x table, you will find atd_light with 814,920 params reaching 1.01 FPS being extremely slower than rtmosr with higher 2,962,128 params reaching 264.16 FPS (assuming, of course, the data in the table being correct).
Hope this helps a bit
@Phips
Thanks so much, it helps a lot!