Closed xinyeCH closed 2 years ago
Sorry, but I am still confused about the speed, in model zoo , AOTS has around 40 FPS and AOTT 50 FPS in DAVIS-VOS as marked in the following picture.
@xinyeCH I update the code about result saving. At present, the result saving happens after the inference of each sequence instead of each frame.
The speed should be faster on your device.
@xinyeCH One more update. Currently, the evaluator call torch.cuda.synchronize() only once. After the inference of each sequence, I synchronize all the frame timers (start
and end
). The influence of slow CPU and storage devices should be negligible now.
I evaluated the new code on a V100 device with slow storage, maybe like your device, the speed results are slightly faster than the reported results now (AOTS: 41fps now, 33fps before).
@z-x-yang Thanks for your reply.
I updated the evaluator.py according to your newest commit and modified some hyperparameters of dataloader such as worker numbers. FPS of AOTS is about 34 now and AOTT 45.
I think the difference between my testing result and paper maybe comes from computer environment.
@xinyeCH Could try to use a single GPU in the inference?
@z-x-yang Sorry for the late reply. Yes, I use only one V100 during inference. Using 2 gpus gets worser performance.
OK. If your cpu is strong enough, using multiple gpus should not make speed slower.
Thanks for making code available! I met some questions while testing the pretrained model! It can only get a speed of near 29FPS when testing the PRE_YTB_DAV pretrained model of DAVIS2017, AOTS which should be 40FPS according to paper result. But the test J & F-mean is the same as the results posted in model_zoo which is 0.820575.
I did not modify the default test config of aots.py exclude dir such as dataset. Did I need to modify something in train_eval.sh?
My device: 2 x Tesla V100 SXM2 32GB Driver Version: 450.51.06 CUDA Version: 11.0 pytorch==1.7.0 torchvision==0.8.1 spatial-correlation-sampler == 0.3.0