Bittensor version control

[steffencruz]

@joeylegere proposes we use submodules 🤮

[steffencruz]

Related to #13

[steffencruz]

@ifrit98 @p-ferreira @joeylegere :The current plan is to merge a modified version of neuron into the text_prompting branch in bittensor, and then pin this repo to that specific commit. We can then include the commit of bittensor that we are experimenting with in our wandb config tracking.

Requirements for new neuron class

Key requirements are:

• Should be possible to disable downloading models and checkpoints when creating an instance
• Should be possible to avoid wallet registering and subtensor connection
• Should be possible to specify number of train steps

Requirements for new GatingModel class:

• Uses sentence embeddings as base model to extract representations (we will develop this locally #79)
• Loaded model is not trainable. Only last linear layer is. Currently all layers in the gating model are trainable

Please remind me if I forgot anything: otherwise it will become abundantly clear soon when nothing is working.. :)

When these features are available in bittensor we will remove the sources directory from mirror_neuron 👻