search

facebookresearch / CutLER

Code release for "Cut and Learn for Unsupervised Object Detection and Instance Segmentation" and "VideoCutLER: Surprisingly Simple Unsupervised Video Instance Segmentation"
Other
932 stars 91 forks source link

set value of SAMPLING_FRAME_NUM #68

Open Lin-UH opened 3 weeks ago

Lin-UH commented 3 weeks ago

Hi author,

It seems like SAMPLING_FRAME_NUM decides the synthetic video length that fed to model. what this value shoule be during training if I want to evaluate model on 100 frames video? Do you have any suggestions? Thank you so much.

frank-xwang commented 3 weeks ago

Hey! You don't have to set cfg.INPUT.SAMPLING_FRAME_NUM to a high value during model training to evaluate videos with 100 frames. It might help the model's performance, but I recommend trying it on the testing videos first.

Hope it helps!

Lin-UH commented 2 weeks ago

Hey! You don't have to set cfg.INPUT.SAMPLING_FRAME_NUM to a high value during model training to evaluate videos with 100 frames. It might help the model's performance, but I recommend trying it on the testing videos first.

Hope it helps!

Thank you so much for your reply!

I have another question about training speed of mask2former, I prepared my dataset annotation using json format same as your imagenet(each video contain two identical images), then train mask2former. However the training speed is super slow, I have around 50k images each is 280 2803, on 8 v100 32G gpu, IMS_PER_BATCH: 40 SAMPLING_FRAME_NUM: 3, it shows almost 100 days to train. Do you know what could cause this? and how long it costs you to train on imagenet?

Thank you