yinian.lw

> > Hi,because the lengths of datas are different,the _**collate()**_ in **_dataloader_** can't work if batchsize!=1. > > Thanks @TonyLRJ , so we can implement collate_fn of the dataloader to...

I used 4 tesla v100(32G) gpus, batch_size=14(16 is OOM) to reproduce Dog and kept the same config with your paper, but the acc is only 90.5, a big difference from...

> > I used 4 tesla v100(32G) gpus, batch_size=14(16 is OOM) to reproduce Dog and kept the same config with your paper, but the acc is only 90.5, a big...

> > Just search `SiamFCppMultiTempTracker` in project and you will get the corresponding config file, > > for example, `experiments/siamfcpp/test/vot/multi_temp/siamfcpp_googlenet-multi_temp.yaml` > > Thanks for your ripadly reply.So is our sot_video_demo.py...

I saw 'self._state['features'] = [self._state['features']] *self._hyper_params['mem_len']' in 'videoanalyst/pipeline/tracker_impl/siamfcpp_track_multi_temp.py', multi_temp only use singe temp to multiple men_len directly, and these templates did not update anymore. So these templates are the same,...

thanks a lot~ another question, if i want to track multi objects, can i put multi object templates into a template pool, then do correlation operation synchronizely in a minibatch.

同学请问解决了吗 我跟你一样 无法收敛 valid-acc 只能达到90%，而我在softmax中能达到99%以上。 test-top1只有30-40% 而softmax中能达到90%以上