TPU Training Time #970
-
|
Hi @rwightman I have been recently trying to set up a training experiment using torch and tpus and I came across the bits branch in this repository. It would be really helpful if you could share the training time (at least approximate) for imagenet classification using standard training procedures for models like resnets, efficientnet, etc. Thanks |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
@MukundVarmaT something you'll need to measure as it varies based on the nework and your hparams. Typical resnet50 I manage around 2800-3000 img/sec so you can calculate for imagenet and your target # of epochs. It can go higher if I lower the precision but that isn't well supported right now due to inability to force some parts of the model to stay in float32. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for your prompt response. That really helps. |
Beta Was this translation helpful? Give feedback.
@MukundVarmaT something you'll need to measure as it varies based on the nework and your hparams. Typical resnet50 I manage around 2800-3000 img/sec so you can calculate for imagenet and your target # of epochs. It can go higher if I lower the precision but that isn't well supported right now due to inability to force some parts of the model to stay in float32.