fpv-iplab / stillfast Goto Github PK

Why is the training of this codebase getting slower and slower? I checked wandb and found that the utilization of gpu0 continued to decrease.

Hi,
Just wanted to know why is it that the RPN module is set to be loaded twice.
As mentioned here,

The class is inherited from FasterCNN class. Is it correct that when initializing the pretrained config is set to False, and is loaded separetly from StillFast class.

Besides, I tried to run the model on v2 ego4d. I am afraid I am doing something wrong since the results are very poor. For still images, I extract the image from lmdb database, instead of loading them separately from an extracted image(as followed in code).

Can you help me here ?

Thanks :)
Sanket

Hi, I am trying to reproduce the results.

In the paper it says u trained with a batch size of 8 on four V100s. Does it mean 8 for each GPU thus 32 in total? Or does it mean 2 for each GPU and 8 in total?

Cos in the config file of this codebase you set batch_size = 14, probably meaning 3 for each GPU when using 4 GPUs, which confused me.

Hello,

Could you kindly provide the training precision/accuracy scores?
I am running the code on the Ego4D dataset but I encounter very low scores.

Kind regards,
Victoria