Open LLJING-ER opened 2 months ago
Yes it is related to the batch size. It is the computation time for processing one batch of data. So you'll have to measure the computation time in the actual hardware for that.
It is necessary to maintain consistent data format. How do you measure the computation_time of NPZ format data in UNET3D? Can you provide me with a reference code?
If possible, please also provide a copy of the test code in tfrecord format for the cosmoflow workload
I'm trying to create a config file *.yaml, but confused about the meaning of computation_time , is it related to batchsize? If I change the batchsize, do I need to run the actual workload to find out the corresponding computation_time**?