Richard Liaw comments

Results 122 comments of


                                            Richard Liaw

[Train] LightningTrainer supports `pl.Trainer.test()` for batch inference example

Let's get the next level of detail - what would the API look like exactly?

[Train] LightningTrainer supports `pl.Trainer.test()` for batch inference example

What are the downsides of using Ray Data for batch prediction? Assume we have something like: ```python ds = ray.data.from_ptl_datamodule(datamodule) predictor = trainer.to_predictor() ds.map_predictor(predictor) ```

redo connection error message

@jjyao can you help review this quick PR?

``requests`` model not listed in environment depencies

Would you be willing to add a patch for this in the environment file?

Run multiple trials on a single GPU

Can you post some logs?

Tune exercise 1 crashes on invoking tune.run

Ah yes, can you try `pip uninstall typing` just to see if that works?

Tune exercise 1 crashes on invoking tune.run

Thanks for reporting this - what version of Ray are you using? On Wed, Jul 17, 2019 at 3:20 AM nesou2 wrote: > I still get an error: > https://gist.github.com/nesou2/4e7fa950f48c127d2ba17c34f099b24e...

Tune exercise 1 crashes on invoking tune.run

Should be fixed with #135

simple multiprocess enhancement for speedup

@superaja Maybe instead of the default HyperOpt parallelism mechanism (or dask), try using HyperOpt [with Tune](https://ray.readthedocs.io/en/latest/tune.html)? Tune can use [HyperOpt to search](https://ray.readthedocs.io/en/latest/tune-searchalg.html#hyperopt-search-tree-structured-parzen-estimators) but manages the asynchronous parallel execution for you...

[Train] TorchTrainer does not free all GPUs on shutdown

@scv119 @matthewdeng @cadedaniel Any ideas here?