Inference time

[palffybalazs]

Hello!

I am reading the article belonging to this repository.
Regarding the inference time, in Sec. 7.2: " A single forward pass for TULIP-L takes 1.33s."
Does this sentence refer to one sample passing through the model, and getting out the prediction for it?

I am curious about wheter the model could be used in a real-time system as an experiment, however this inference time suggests that it is not feasible.

Thank your for your answear in advance.
Best regards!

[binyang97]

Hi!

sorry for the late reply. The number in the supplementary regarding inference time is not correct. Here are the speed test of one iteration (passing a batch of samples into model and getting predictions) with a single RTX 2080TI:

batch = 8, 19.7 ms
batch = 1, 18.6 ms

The number is actually tested on TULIP instead of TULIP-L, but the numbers should be comparable. I hope this resolves your concern.

Best regards,
Bin