clean_description2code.py transforms description2code into a more workable format and randomly selects one solution per problem to be our example code.

After running script dataset looks like:

bash

description2code_clean
├── problem_1
├── problem_2
│   ├── deccription.txt
│   ├── solution.cpp
│   ├── samples
│   │   ├── 1_input.txt
│   │   ├── 1_output.txt
│   │   ├── ...
├── ...

Please Note: CSV is in form: Name,Operations,Operands,MemoryOps,FPOps,Branch,resMII,frequentBBs,nestDepth,tripcount

A shared, working model.ipynb is shared in the 'EECS583 Project' folder on Google Drive and is here on GitHub. It's probably useful for all of us to have our own copy to run our own tests, so you can make a copy in Google Drive, for example, and open your copy in Colab to run your experiments.

To train/validate the model, you will need to drag/drop training and validation csvs into the filesystem and possibly modify the csv paths in the cell that sets up the dataloaders. By default, it is setup to look for train_norm.csv and val_norm.csv on the same level as (not inside) sample_data/ in the file system.

Two things of note are commented out by default:

• The scheduler variable, which decreases the learning rate by gamma every step_size epochs. You can pass this into the mlp.train_model() function.
• The class_weights tensor and its corresponding criterion. These were calculated according to the label distribution with the formula w_j = n_samples / (n_classes * n_samples_j) to help the model learn the minority classes.