This is a submission for the ETHToronto hackathon (2023).
Please see the notebooks directory for the Python code of synthetic dataset generation as well as generation of dataset in Cairo.
The process of solving the logistic regression problem is present as well.
The src directory contains the Cairo code. We make use of the Orion framework for tensor manipulation.
The train.cairo file contains the functions to run gradient descent in Cairo. This can be used as a scaffold for verifiable deep learning.
We use Scarb for managing the installation of Cairo.
To run the tests:
scarb test
The output of the test should correspond to the same result in the Jupyter notebooks.
I set the number of iterations to be 100 to keep the script consise. Running it for more iterations would lead to a further decrease in loss and better results.
Learning rate alpha and number of iterations are parameters that can be specified by the user.
Deployment steps 🚀
scarb build
starkli declare target/dev/logistic_regression.sierra \
--compiler-version 2.2.0 \
--rpc http://0.0.0.0:5050 \
--account ~/.starkli-wallets/deployer/account0_account.json \
--keystore ~/.starkli-wallets/deployer/account0_keystore.json
Deploy on Katana
starkli deploy 0x056486e36366049e8475e58e309668d0d829f8e7a1303737f1e6804844adfd09 42
--rpc http://0.0.0.0:5050 \
--account ~/.starkli-wallets/deployer/account0_account.json \
--keystore ~/.starkli-wallets/deployer/account0_keystore.json
Call smart contract ☎️
starkli call 0x07ba115d69cb54cd1ab4559cfd2386ff7fa44303896507f6c610d005dd52a10e get \
--rpc http://0.0.0.0:5050
Invoke function smart contract ☎️
starkli invoke 0x07ba115d69cb54cd1ab4559cfd2386ff7fa44303896507f6c610d005dd52a10e set 5 \
--rpc http://0.0.0.0:5050 --account ~/.starkli-wallets/deployer/account0_account.json \
--keystore ~/.starkli-wallets/deployer/account0_keystore.json
TODO: [ ] wrap ml code in https://book.cairo-lang.org/ch99-00-starknet-smart-contracts.html #[contract]