mirror of
https://github.com/clearml/clearml-docs
synced 2025-01-31 14:37:18 +00:00
71 lines
2.9 KiB
Markdown
71 lines
2.9 KiB
Markdown
---
|
||
title: MMEngine
|
||
---
|
||
|
||
:::tip
|
||
If you are not already using ClearML, see [Getting Started](../getting_started/ds/ds_first_steps.md) for setup
|
||
instructions.
|
||
:::
|
||
|
||
[MMEngine](https://github.com/open-mmlab/mmengine) is a library for training deep learning models based on PyTorch.
|
||
MMEngine supports ClearML through a builtin logger: It automatically logs experiment environment information, such as
|
||
required packages and uncommitted changes, and supports reporting scalars, parameters, and debug samples.
|
||
|
||
Integrate ClearML with the following steps:
|
||
1. Instantiate a [`ClearMLVisBackend`](https://mmengine.readthedocs.io/en/latest/api/generated/mmengine.visualization.ClearMLVisBackend.html#mmengine.visualization.ClearMLVisBackend)
|
||
object. This creates a ClearML Task that logs the experiment’s environment information.
|
||
|
||
```python
|
||
from mmengine.visualization import ClearMLVisBackend
|
||
vis_backend = ClearMLVisBackend(
|
||
artifact_suffix=('.py', 'pth'),
|
||
init_kwargs=dict(
|
||
project_name='examples',
|
||
task_name='OpenMMLab cifar10',
|
||
output_uri=True
|
||
)
|
||
)
|
||
```
|
||
|
||
You can specify the following parameters:
|
||
|
||
* `init_kwargs` – A dictionary that contains the arguments to pass to ClearML's [`Task.init()`](../references/sdk/task.md#taskinit).
|
||
* `artifact_suffix` – At the end of training, artifacts with these suffixes will be uploaded to the task's `output_uri`.
|
||
Defaults to (`.py`, `pth`).
|
||
|
||
2. Log experiment parameters using `ClearMLVisBackend.add_config()`. Under the `config` parameter, input a dictionary of parameter key-value pairs.
|
||
|
||
```python
|
||
cfg = Config(dict(a=1, b=dict(b1=[0, 1])))
|
||
vis_backend.add_config(config=cfg)
|
||
```
|
||
|
||
The parameters will be displayed in the ClearML WebApp, under the experiment’s Hyperparameters
|
||
|
||
3. Log your experiment’s scalars using either `ClearMLVisBackend.add_scalar()` for single values or `ClearMLVisBackend.add_scalars()`
|
||
for multiple values:
|
||
|
||
```python
|
||
vis_backend.add_scalar(name='mAP', value=0.6, step=1)
|
||
vis_backend.add_scalars(scalar_dict={'loss': 0.1,'acc':0.8}, step=1)
|
||
```
|
||
|
||
The scalars are displayed in the experiment's Scalars tab.
|
||
|
||
5. Report images to your experiment using `ClearMLVisBackend.add_image()`. Under the `image` parameter, input the image
|
||
to be reported as an `np.ndarray` in RGB format:
|
||
|
||
```python
|
||
img = np.random.randint(0, 256, size=(10, 10, 3))
|
||
vis_backend.add_image(name='img.png', image=img, step=1)
|
||
```
|
||
The images will be displayed in the experiment's Debug Samples
|
||
|
||
5. Once you've finished training, make sure to run `ClearMLVisBackend.close()` so that ClearML can mark the task as
|
||
completed. This will also scan the directory for relevant artifacts with the suffixes input when instantiating
|
||
`ClearMLVisBackend` and log the artifacts to your experiment.
|
||
|
||
```python
|
||
vis_backend.close()
|
||
```
|