search

RomeroBarata / skeleton_based_anomaly_detection

Code for the CVPR'19 paper "Learning Regularity in Skeleton Trajectories for Anomaly Detection in Videos"
132 stars 49 forks source link

GPUs are configured but not used #37

Open Sara-Esam opened 8 months ago

Sara-Esam commented 8 months ago

Hello,

I would like to thank you first for this nice work. I have encountered an issue while training, where I find the gpus are configured, e.g. memory is reserved but they're not used during either training or evaluation. Is this behavior expected?

A follow up question: was this code used to train a large set of trajectories? or what portions of trajectories were used during training? according to the behavior above, the code does not seem to handle large sets of trajectories, e.g. 6000+.

Thanks in advance.

RomeroBarata commented 8 months ago

Hi Sara-Esam,

The gpus should be used by the model during training and evaluation. How are you checking that the gpus are being used?

I don't remember right now the total size of the training data (you might be able to double check by adding some print statements to the code), but if I remember correctly the code used about 2Gbs of gpu memory (for a batch size of 256). If you batch size is too big (6,000 as you mentioned) you may have memory issues depending on your gpu.

Romero