Client Flow
1. Claim Outstanding Rewards:
Automatically retrieve and claim any uncollected rewards for previous epochs, enhancing user incentive alignment.
2. Role Assignment Wait Period:
Pause the workflow and wait for the assignment of a specific role to the user within the federated learning process.
3. Role-Based Actions:
The client executes different actions based on the assigned role:
i. Proposer:
Download global model parameters:
Retrieve the latest global model parameters from the blockchain.
Retrain model with dataset:
Use the input dataset to fine-tune the global model parameters locally.
Upload updated model parameters:
Store the updated parameters on IPFS for permissionless.
Submit contribution hash:
Register the contribution with the FlockManager to provide a transparent record of contributions.
ii. Voter:
Download global model & calculate initial accuracy:
Retrieve latest parameters and assess initial accuracy on the user's test dataset partition.
Retrieve current epoch's contributions:
Gather the contribution hashes for the current learning epoch.
Download contributions from IPFS:
Retrieve the localised insights for aggregation.
Aggregate contributions via averaging:
Compile individual contributions to update the global model.
Upload aggregated model to IPFS:
Store the updated model securely and transparently.
Push aggregated model hash to blockchain:
Register the aggregated model's hash for validation.
Calculate aggregated model accuracy:
Test the updated model on the user's data partition to assess performance.
Determine output accuracy:
Compute the final accuracy metric.
Submit vote score:
Cast a vote based on the accuracy delta (step 8 vs step 1), contributing to the consensus process.
Training Completion Check:
Continue the loop, executing the appropriate steps for the assigned role, until the federated learning process reaches its defined completion criteria.
The loop ensures a seamless and continuous engagement with the FLock ecosystem, allowing participants to contribute to the learning process and aligning their incentives
Last updated