llama / data /xtuner /docs /zh_cn /training /visualization.rst
kai119's picture
Upload folder using huggingface_hub
22fb4ec verified
==============
可视化训练过程
==============
XTuner 支持通过 `MMEngine <https://github.com/open-mmlab/mmengine>`__
使用 `TensorBoard <https://www.tensorflow.org/tensorboard?hl=zh-cn>`__
和 `Weights & Biases (WandB) <https://docs.wandb.ai/>`__
实验管理工具,只需在 config 中添加一行代码,就可以跟踪和可视化损失、显存占用等指标。
TensorBoard
============
1. 设置 config 中的 ``visualizer`` 字段,并将 ``vis_backends`` 设置为 `TensorboardVisBackend <https://github.com/open-mmlab/mmengine/blob/2c4516c62294964065d058d98799402f50afdef6/mmengine/visualization/vis_backend.py#L514>`__\ :
.. code:: diff
# set visualizer
- visualizer = None
+ from mmengine.visualization import Visualizer, TensorboardVisBackend
+ visualizer = dict(type=Visualizer, vis_backends=[dict(type=TensorboardVisBackend)])
2. 启动实验后,tensorboard 产生的相关文件会存在 ``vis_data`` 中,通过 tensorboard 命令可以启动进行实时可视化:
|image1|
.. code::
tensorboard --logdir=$PATH_TO_VIS_DATA
WandB
======
1. 使用 WandB 前需安装依赖库 ``wandb`` 并登录至 wandb。
.. code:: console
$ pip install wandb
$ wandb login
2. 设置 config 中的 ``visualizer`` 字段,并将 ``vis_backends`` 设置为 `WandbVisBackend <https://github.com/open-mmlab/mmengine/blob/2c4516c62294964065d058d98799402f50afdef6/mmengine/visualization/vis_backend.py#L330>`__\ :
.. code:: diff
# set visualizer
+ from mmengine.visualization import Visualizer, WandbVisBackend
- visualizer = None
+ visualizer = dict(type=Visualizer, vis_backends=[dict(type=WandbVisBackend)])
.. tip::
可以点击 `WandbVisBackend
API <https://github.com/open-mmlab/mmengine/blob/2c4516c62294964065d058d98799402f50afdef6/mmengine/visualization/vis_backend.py#L330>`__
查看 ``WandbVisBackend`` 可配置的参数。例如
``init_kwargs``\ ,该参数会传给
`wandb.init <https://docs.wandb.ai/ref/python/init>`__ 方法。
.. code:: diff
# set visualizer
- visualizer = None
+ from mmengine.visualization import Visualizer, WandbVisBackend
+ visualizer = dict(
+ type=Visualizer,
+ vis_backends=[
+ dict(type=WandbVisBackend, init_kwargs=dict(project='toy-example'))])
3. 启动实验后,可在 wandb 网页端 ``https://wandb.ai`` 上查看可视化结果:
|image2|
.. |image1| image:: https://github.com/InternLM/xtuner/assets/67539920/abacb28f-5afd-46d0-91b2-acdd20887969
.. |image2| image:: https://github.com/InternLM/xtuner/assets/41630003/fc16387a-3c83-4015-9235-8ec811077953