Support for tensorflow gpu about ranking HOT 8 CLOSED

See #16

from ranking.

Hi Dave,

We have run using GPU and CPU versions, but haven't done the kind of benchmarking you have pointed out, which is definitely interesting to know.

The dependency on tensorflow vs tensorflow-gpu occurs in the setup.py file.
The way to support both GPU and CPU versions would be to have tensorflow_ranking and tensorflow_ranking-gpu packages installable via pip, each having a separate setup.py.

from ranking.

Could you clarify what changes we should do in tensorflow_ranking/tools/pip_package/build_pip_package.sh file for GPU support, along with changes in setup.py?

from ranking.

@NarasimhaM3101 just change this line https://github.com/tensorflow/ranking/blob/master/tensorflow_ranking/tools/pip_package/setup.py#L35 to tensorflow-gpu >= 1.12.0

from ranking.

Thanks for the fix, Ishant! We are planning to remove the dependency on tensorflow in the package, and ask the user to explicitly install tensorflow (or tensorflow-gpu). This will allow us to maintain one package on PyPI.

from ranking.

The user is expected to install tensorflow or tensorflow-gpu manually before installing tensorflow-ranking.The README is updated to reflect this. Closing this for now. Please feel free to reopen if they are any issues.

from ranking.

I installed tensorflow-gpu(v1.13.1) and tensorflow ranking by pip command. But when I run the training process, the GPU is actually not used. The code I used is as following, is there any problem?

config_proto = tf.ConfigProto(log_device_placement=False,
allow_soft_placement=False)
config_proto.gpu_options.per_process_gpu_memory_fraction = 0.6
run_config = tf.estimator.RunConfig(model_dir=config.model_dir,
save_summary_steps=config.log_steps,
save_checkpoints_steps=config.checkpoints_steps,
keep_checkpoint_max=config.keep_checkpoint_max,
session_config=config_proto,
log_step_count_steps=config.log_steps
)
return tf.estimator.Estimator(
model_fn=tfr.model.make_groupwise_ranking_fn(
group_score_fn=make_score_fn(),
group_size=config.group_size,
transform_fn=None,
ranking_head=ranking_head),
config=run_config,
params={'hidden_units': hidden_units_list,
'drop_out': config.drop_out,
'drop_rates': drop_rates_list,
'group_size': config.group_size,
'num_features': config.num_features,
'l2_reg': config.l2_reg,
'batch_norm': config.batch_norm}
)

from ranking.

hi @iamhere1 , did you get any solution to this problem? I have tried to install the library by all means but still, my nvidia-smi command does not list my process id

from ranking.