How to re-initlaize a custom accelerator inside the training loop ?

[shamanez]

In the code I use, there is a custom accelerator (DDPAccelerator) that loads an embedding similarity search model as a separate process. Now I need to re-initialize this similar method after certain training steps. In order to do this, I need to get access to the accelerator callback mentioned in this line.

Is this possible?

@williamFalcon

[carmocca]

The link you provided isn't working. Can you share the correct link?

[shamanez]

I have updated the link.

[carmocca]

Thanks. I didn't dive deep into the code but looks to me like you actually want use a Plugin (https://pytorch-lightning.readthedocs.io/en/stable/plugins.html) not an Accelerator. You could subclass the DDPPlugin to properly initialize your embeddings. That could be combined with a Callback to accomplish whatever logic you need during training using whatever state was set by the plugin.

Not sure if this helps. Would probably need more details.