Yihao Zhao

@juncgu Hi, I use n32g4 cluster setup, check the 60_job trace provided in this repo, and find no extra large job. Each job is set to use 5GB memory per...

Thanks for your interest. I think there are two steps. First, change the input_dim / output_dim in config file. Second, modify the architecture of the network in networks.py.

In our experiment, we leave these two parameters as False, which is default in PyTorch. When set `track_running_states=True`, it will track the mean and variance. When False, it will only...

Thanks for your suggestions! We did not test for images with h!=w, and thus there may be some errors. I think center crop should work.

Thanks to your timely reply. Does byteps have scheduling at the worker side by default? Is there any parameter to control this feature?

Thanks a lot. Is the priority used for worker-side scheduling the index of the layers? What is the priority used for server-side scheduling?

Hi, @ymjiang ,I tried to set these parameters to control scheduling behavior, which is successful while using MxNet BytePS. However, when I set `byteps_server_enable_schedule=1` and `byteps_scheduling_credit=4`, the timeline looks like...

I'm a little confused. First of all, in MxNet, when I use these parameters, the former layers are selected for communication, and the latter ones that are not sent while...

I think there is a problem with the for loop in GetTask function in https://github.com/bytedance/byteps/blob/249006c9105d7b4fd09962eb133c3e76de1c8656/byteps/common/scheduled_queue.cc. For example, when a large tensor with high priority is not chosen because of credit...

Thanks to your explanation! Could you please tell me what else makes the priority unguaranteed?