Welcome to the Deep Learning Pipelines Python API docs!¶

Horovod Runner¶

class sparkdl.HorovodRunner(*, np, driver_log_verbosity='log_callback_only')[source]¶

Bases: object

HorovodRunner runs distributed deep learning training jobs using Horovod.

On Databricks Runtime 5.0 ML and above, it launches the Horovod job as a distributed Spark job. It makes running Horovod easy on Databricks by managing the cluster setup and integrating with Spark. Check out Databricks documentation to view end-to-end examples and performance tuning tips.

The open-source version only runs the job locally inside the same Python process, which is for local development only.

Note

Horovod is a distributed training framework developed by Uber.

Parameters

np –
number of parallel processes to use for the Horovod job. This argument only takes effect on Databricks Runtime 5.0 ML and above. It is ignored in the open-source version. On Databricks, each process will take an available task slot, which maps to a GPU on a GPU cluster or a CPU core on a CPU cluster. Accepted values are:
- If <0, this will spawn -np subprocesses on the driver node to run Horovod locally. Training stdout and stderr messages go to the notebook cell output, and are also available in driver logs in case the cell output is truncated. This is useful for debugging and we recommend testing your code under this mode first. However, be careful of heavy use of the Spark driver on a shared Databricks cluster. Note that np < -1 is only supported on Databricks Runtime 5.5 ML and above.
- If >0, this will launch a Spark job with np tasks starting all together and run the Horovod job on the task nodes. It will wait until np task slots are available to launch the job. If np is greater than the total number of task slots on the cluster, the job will fail. As of Databricks Runtime 5.4 ML, training stdout and stderr messages go to the notebook cell output. In the event that the cell output is truncated, full logs are available in stderr stream of task 0 under the 2nd spark job started by HorovodRunner, which you can find in the Spark UI.
driver_log_verbosity – driver log verbosity, “all” or “log_callback_only”(default). During training, the first worker process will collect logs from all workers. The training logs are always merged into the first Spark executors stderr logs. If driver log verbosity is “all”, HorovodRunner streams all logs to the driver and shows them in the notebook cell output. However, this might generate excessive amount of logs during distributed training. You can turn it off by setting driver log verbosity to “log_callback_only”. In this mode, HorovodRunner will only stream selected logs from remote worker. logs from remote worker can be sent by sparkdl.horovod.log_to_driver(), or use log callback in the first worker process, e.g., sparkdl.horovod.tensorflow.keras.LogCallback.

run(main, **kwargs)[source]¶

Runs a Horovod training job invoking main(**kwargs).

The open-source version only invokes main(**kwargs) inside the same Python process. On Databricks Runtime 5.0 ML and above, it will launch the Horovod job based on the documented behavior of np. Both the main function and the keyword arguments are serialized using cloudpickle and distributed to cluster workers.

Parameters

main – a Python function that contains the Horovod training code. The expected signature is def main(**kwargs) or compatible forms. Because the function gets pickled and distributed to workers, please change global states inside the function, e.g., setting logging level, and be aware of pickling limitations. Avoid referencing large objects in the function, which might result large pickled data, making the job slow to start.
kwargs – keyword arguments passed to the main function at invocation time.

Returns

return value of the main function. With np>=0, this returns the value from the rank 0 process. Note that the returned value should be serializable using cloudpickle.

sparkdl.horovod.log_to_driver(message)[source]¶: Send a log message (string type) to driver side, and driver will print log to stdout. If message length is greater than 4000, it will be truncated.

class sparkdl.horovod.tensorflow.keras.LogCallback(per_batch_log=False)[source]¶

Bases: keras.callbacks.Callback

A simple HorovodRunner log callback that streams event logs to notebook cell output.

Parameters: per_batch_log – whether to output logs per batch, default: False.

on_batch_begin(batch, logs=None)¶: A backwards compatibility alias for on_train_batch_begin.

on_batch_end(batch, logs=None)[source]¶: A backwards compatibility alias for on_train_batch_end.

on_epoch_begin(epoch, logs=None)[source]¶

Called at the start of an epoch.

Subclasses should override for any actions to run. This function should only be called during TRAIN mode.

Args:: epoch: Integer, index of epoch. logs: Dict. Currently no data is passed to this argument for this method

but that may change in the future.

on_epoch_end(epoch, logs=None)[source]¶

Called at the end of an epoch.

Subclasses should override for any actions to run. This function should only be called during TRAIN mode.

Args:

epoch: Integer, index of epoch. logs: Dict, metric results for this training epoch, and for the

validation epoch if validation is performed. Validation result keys are prefixed with val_. For training epoch, the values of the

Model’s metrics are returned. Example`{‘loss’: 0.2, ‘accuracy’:
0.7}`.

on_predict_batch_begin(batch, logs=None)¶

Called at the beginning of a batch in predict methods.