Skip to content

Latest commit

 

History

History
133 lines (81 loc) · 6.34 KB

miner.md

File metadata and controls

133 lines (81 loc) · 6.34 KB
Raw

Miner

Miners train locally and periodically publish their best model to 🤗 Hugging Face and commit the metadata for that model to the Bittensor chain.

Miners can only have one model associated with them on the chain for evaluation by validators at a time.

The communication between a miner and a validator happens asynchronously chain and therefore Miners do not need to be running continuously. Validators will use whichever metadata was most recently published by the miner to know which model to download from 🤗 Hugging Face.

System Requirements

Miners will need enough disk space to store their model as they work on. Each uploaded model (As of Jun 15th, 2024) may not be more than 15 GB. It is recommended to have at least 50 GB of disk space.

Miners will need enough processing power to train their model. The device the model is trained on is recommended to be a large GPU with at least 48 GB of VRAM. To be competitive you will likely need clusters of GPUs.

Getting started

Prerequisites

  1. Get a Hugging Face Account:

Miners and validators use 🤗 Hugging Face in order to share model state information. Miners will be uploading to 🤗 Hugging Face and therefore must attain a account from 🤗 Hugging Face along with a user access token which can be found by following the instructions here.

Make sure that any repo you create for uploading is public so that the validators can download from it for evaluation.

  1. Get a Wandb Account: Miners and validators use Wandb to download data from subnet 1. Wandb accounts can be obtained at https://wandb.ai/ and the user access token can be found at https://wandb.ai/authorize once logged in.

  2. Clone the repo

git clone https://github.com/macrocosm-os/finetuning.git

  1. Setup your python virtual environment or Conda environment.

  2. Install the requirements. From your virtual environment, run

cd finetuning
python -m pip install -e .

Note: We require a python version of at least 3.9.

  1. Make sure you've created a Wallet and registered a hotkey.

  2. (Optional) Run a Subtensor instance:

Your node will run better if you are connecting to a local Bittensor chain entrypoint node rather than using Opentensor's. We recommend running a local node as follows and passing the --subtensor.network local flag to your running miners/validators. To install and run a local subtensor node follow the commands below with Docker and Docker-Compose previously installed.

git clone https://github.com/opentensor/subtensor.git
cd subtensor
docker compose up --detach

Running the Miner

The mining script has a shell that does some initial setup but is not fully implemented. You will need to implement the specific training logic yourself prior to running it.

As of Oct 1st, 2024 for the current competition the subnet works with models matching the subnet 9 outputs and evaluates them against synthetic data from subnet 1.

The specific requirements for each competition can be found here.

The finetune/mining.py file has several methods that you may find useful. See the examples Jupyter notebook for some example uses.

See Validator Psuedocode for more information on how the evaluation occurs.

Env File

The Miner requires a .env file with your 🤗 Hugging Face access token in order to upload models and a Wandb access token in order to download training data from subnet 1.

Create a .env file in the finetuning directory and add the following to it:

HF_ACCESS_TOKEN="YOUR_HF_ACCESS_TOKEN"
WANDB_ACCESS_TOKEN="YOUR_WANDB_ACCESS_TOKEN"

Starting the Miner

To start your miner the most basic command is

python neurons/miner.py --wallet.name coldkey --wallet.hotkey hotkey --hf_repo_id my-username/my-project --avg_loss_upload_threshold YOUR_THRESHOLD

  • --wallet.name: should be the name of the coldkey that contains the hotkey your miner is registered with.

  • --wallet.hotkey: should be the name of the hotkey that your miner is registered with.

  • --hf_repo_id: should be the namespace/model_name that matches the hugging face repo you want to upload to. Must be public so that the validators can download from it.

  • --avg_loss_upload_threshold: should be the minimum average loss or deviation before you want your miner to upload the model.

  • --competition_id: competition you wish to mine for; run --list_competitions to get a list of available competitions

Flags

The Miner offers some flags to customize properties, such as how to train the model and which hugging face repo to upload to.

You can view the full set of flags by running

python ./neurons/miner.py -h

Some flags you may find useful:

  • --offline: when set you can run the miner without being registered and it will not attempt to upload the model.

  • --wandb_entity + --wandb_project: when both flags are set the miner will log its training to the provided wandb project.

  • --device: by default the miner will use your gpu but you can specify with this flag if you have multiple.

Training from pre-existing models

  • --load_best: when set you will download and train the model from the current best miner on the network.
  • --load_uid: when passing a uid you will download and train the model from the matching miner on the network.
  • --load_model_dir: the path to a local model directory [saved via Hugging Face API].

Manually uploading a model

In some cases you may have failed to upload a model or wish to upload a model without further training.

Due to rate limiting by the Bittensor chain you may only upload a model every 20 minutes.

You can manually upload with the following command:

python scripts/upload_model.py --load_model_dir <path to model> --hf_repo_id my-username/my-project --wallet.name coldkey --wallet.hotkey hotkey