mirror of
https://github.com/clearml/clearml-docs
synced 2025-02-24 21:14:37 +00:00
81 lines
3.0 KiB
Markdown
81 lines
3.0 KiB
Markdown
---
|
|
title: Remote Execution
|
|
---
|
|
|
|
The [execute_remotely_example](https://github.com/allegroai/clearml/blob/master/examples/advanced/execute_remotely_example.py)
|
|
script demonstrates the use of the [`Task.execute_remotely()`](../../references/sdk/task.md#execute_remotely) method.
|
|
|
|
:::note
|
|
Make sure to have at least one [ClearML Agent](../../clearml_agent.md) running and assigned to listen to the `default` queue:
|
|
```
|
|
clearml-agent daemon --queue default
|
|
```
|
|
:::
|
|
|
|
## Execution Flow
|
|
|
|
The script trains a simple deep neural network on the PyTorch built-in MNIST dataset. The following describes the code's
|
|
execution flow:
|
|
1. The training runs for one epoch.
|
|
1. The code uses [`Task.execute_remotely()`](../../references/sdk/task.md#execute_remotely), which terminates the local execution of the code and enqueues the task
|
|
to the `default` queue, as specified in the `queue_name` parameter.
|
|
1. An agent listening to the queue fetches the task and restarts task execution remotely. When the agent executes the task,
|
|
the `execute_remotely` is considered no-op.
|
|
|
|
An execution flow that uses `execute_remotely` is especially helpful when running code on a development machine for a few iterations
|
|
to debug and to make sure the code doesn't crash, or to set up an environment. After that, the training can be
|
|
moved to be executed by a stronger machine.
|
|
|
|
During the execution of the example script, the code does the following:
|
|
* Uses ClearML's automatic and explicit logging.
|
|
* Creates an experiment named `Remote_execution PyTorch MNIST train` in the `examples` project.
|
|
|
|
|
|
## Scalars
|
|
|
|
In the example script's `train` function, the following code explicitly reports scalars to ClearML:
|
|
|
|
```python
|
|
Logger.current_logger().report_scalar(
|
|
title="train",
|
|
series="loss",
|
|
iteration=(epoch * len(train_loader) + batch_idx),
|
|
value=loss.item()
|
|
)
|
|
```
|
|
|
|
In the script's `test` function, the code explicitly reports `loss` and `accuracy` scalars.
|
|
|
|
```python
|
|
Logger.current_logger().report_scalar(
|
|
title="test", series="loss", iteration=epoch, value=test_loss
|
|
)
|
|
Logger.current_logger().report_scalar(
|
|
title="test", series="accuracy", iteration=epoch, value=(correct / len(test_loader.dataset))
|
|
)
|
|
```
|
|
|
|
These scalars can be visualized in plots, which appear in the ClearML web UI, in the experiment's **SCALARS** tab.
|
|
|
|
|
|
|
|
## Hyperparameters
|
|
|
|
ClearML automatically logs command line options defined with `argparse`. They appear in **CONFIGURATION** **>** **HYPERPARAMETERS** **>** **Args**.
|
|
|
|
|
|
|
|
## Console
|
|
|
|
Text printed to the console for training progress, as well as all other console output, appear in **CONSOLE**.
|
|
|
|
|
|
|
|
## Artifacts
|
|
|
|
Models created by the experiment appear in the experiment's **ARTIFACTS** tab. ClearML automatically logs and tracks models
|
|
and any snapshots created using PyTorch.
|
|
|
|
|
|
|