What if our model generates code for us ?? Isn't that interesting ?
Yes ! We gonna build a LSTM model to generate python code for us !
- Importing the 'txt' source code file into the environment
- Reading each line in the file and converting them into lowercase
- Tokenizing them into characters and finding the total number of characters and number of unique characters
- Mapping the unique characters to integers, since we cant feed the characters to our model
- Creating semi-redundant sequences and splitting them into X and Y(labels)
- One-hot encoding the Y(labels)
- Building a LSTM model with 2-LSTM layers and a fully-connected layer
- Training and predicting to generate source code.
Prediction results after running 10 Epochs :
As you can see, our model predicted a "function" starting with "def", and with correct indendations. We can improvise this by building a bigger model with 4 LSTM layers containing 500 or 700 neurons in each layer, but this could be computationally expensive and time-consuming if trained on CPU.