This is a PoC for HackWek! - a Diligence internal 5-day Hackathon 🥷⚔️.
TLDR; My plan was to have fun with tensorflow, RNN's, text-prediction, and connect this to solidity smart contracts 🙌. This is an excerpt of my journey.
😵💫 A Recurrent Neural Network (RNN) hallucinating solidity source code.
- We train our model on samples from the https://github.com/tintinweb/smart-contract-sanctuary
- And then "hallucinate" new contracts
Note: train the model on https://colab.research.google.com/ as it is much faster than doing this locally.
Copy the python notebook to your own collab/google drive and runit.
Hint: Google Collab → Runtime → Change Runtime Type: GPU
| Folder | Description |
|---|---|
| solidity_model_text | contains a pre-trained model trained on 15mb solidity input, naive character based training, with sampling sequence length of 250 chars. The model has an embedding_dimension of 256 with 1024 rnn_units. It was trained for 15 epochs on google collab (hw-accelleration: GPU) which took somewhere between 1-1.5 hrs. |
| Tutorial 2: load & hallucinate | loads the pre-trained model from ./solidity_model_text/ and hallucinates more solidity. |
| Tutorial 1: train & hallucinate | is the code that downloads samples from https://github.com/tintinweb/smart-contract-sanctuary, creates the model, trains it, hallucinates some text, and then continues to show how to save/restore/re-train the model. |
- Note: The model can be exported for use with tensorflow.js so that it can be used with any javascript/web-front/backend. See Tutorial 1 for how to do this.
- Note: The model can also be used for non-solidity code. Just make sure to write your own
SolidityTrainerclass 🙌.
Of course, there's no way to explore everything in this 5-day HackWek period, but, here're a couple of thoughts on what to improve:
- vocabulary should be based on tokentype_text instead of chars. E.g. use
pygmentsto lexsolidityand map this as the vocabulary. This should give way higher quality output and allows the model to learn the source structure more efficiently. - input cleanup should reliably remove all comments/pragmas/etc.
- loss function should reinforce training towards fuzzy-parseable code
- shuffle before downloading contract sources
- continuous learning. re-train with more sources (not only 15mb 😂)
- the pre-trained
solidity_model_textis pretty shitty an will generate a lot of garbage. Obviously, 15 epochs is not enough and the text based shuffling approach makes no sense. But at least it is generating something 😂
Copy the two tutorials to your google drive and run them.
Input:
>>> print(trainingData.predict(['contract '], 3000))Output:
contract Ownable {
address public owner;
event OwnershipTransferred(address indexed previousOwner, address indexed newOwner);
function Ownable() public {
owner = msg.sender;
}
modifier onlyOwner() {
require(msg.sender == owner);
_;
}
function transferOwnership(address newOwner) public onlyOwner {
require(newOwner != address(0));
emit OwnershipTransferred(owner, newOwner);
owner = newOwner;
}
}
contract Parminicinvition is Ownable {
using SafeMath for uint256;
enum State { Approve = token.totalSupply();
require(tokens >= summaryTokens.add(bonus));
totalDailydested = totalEthInWei + msg.value;
totalSoldTokens = token.totalSupply();
emit Transfer(address(0), 0xCf49B9298aC4d4933a7D6984d89A49aDc84A6CA602BA513D872C3,21f36325D28718](0));
totalSupply = totalSupply.mul(totalValue.add(soldSignedMap[tokensBough.mul(1)));
restributedPluyRates[msg.sender] = true;
nonStokenSupplyFinallow
}
if(opits[msg.sender].amount <= totalSupply)) ether;
}
assignOpe(address(this).balance, weiAmount);
require(canTra_secrecover(_approved) >= rNo(_reward, _weight, _amount);
totalAmount = totalAmount.add(_amount);
Transfer(_addr, msg.sender, amount);
}
...Based on the TensorFlow Text Generation Tutorial