29 lines
738 B
ReStructuredText
29 lines
738 B
ReStructuredText
|
:orphan:
|
||
|
|
|
||
|
|
.. _precision_expert:
|
||
|
|
|
||
|
|
########################
|
||
|
|
N-Bit Precision (Expert)
|
||
|
|
########################
|
||
|
|
**Audience:** Researchers looking to integrate their new precision techniques into Lightning.
|
||
|
|
|
||
|
|
|
||
|
|
*****************
|
||
|
|
Precision Plugins
|
||
|
|
*****************
|
||
|
|
|
||
|
|
You can also customize and pass your own Precision Plugin by subclassing the :class:`~pytorch_lightning.plugins.precision.precision_plugin.PrecisionPlugin` class.
|
||
|
|
|
||
|
|
- Perform pre and post backward/optimizer step operations such as scaling gradients.
|
||
|
|
- Provide context managers for forward, training_step, etc.
|
||
|
|
|
||
|
|
.. code-block:: python
|
||
|
|
|
||
|
|
class CustomPrecisionPlugin(PrecisionPlugin):
|
||
|
|
precision = 16
|
||
|
|
|
||
|
|
...
|
||
|
|
|
||
|
|
|
||
|
|
trainer = Trainer(plugins=[CustomPrecisionPlugin()])
|