This project aims to predict the final accuracy of a DNN model for given hyperparameters values, without training the DNN model completely.

• dataset: this folder contains all the dataset files
• preprocess_dataset: this folder contains the preprocessed dataset files generated after running preprocessing.ipynb.
• models: this folder contains the best model generated during model training in training.ipynb.
• results: this folder contains the file containing the validation accuracies of the model in the settings asked.

• preprocessing.ipynb contains the necessary steps for preprocessing the data. The resulting dataset in stored in the preprocessed_dataset folder.
• training.ipynb

  • This contains all the necessary code for training and evaluating the model.

  • The best model is saved in models folder. This is so the results can be reproduced by loading the model and directly testing the model. To do so, you can skip the training step and run the follwing code to load the model and move on to the testing.

    model = LSTM(input_dim, hidden_dim, output_dim)
    state_dict = torch.load(f'{path}/models/final_model.pt')
    model.load_state_dict(state_dict)

• NOTE:

  • In training.ipynb, change the variable path to your folder containing all the files. For example,

    path = '/content/drive/MyDrive/Dataset_hyperEarlyStop'

  • In preprocessing.ipynb, change your working directory to your folder containing all the files. For example,

    %cd '/content/drive/MyDrive/Dataset_hyperEarlyStop/'

    And while, saving the preprocessed dataset, set the path to your folder. For example,

    train_loss_df.iloc[train_lines].to_csv("/content/drive/MyDrive/Dataset_hyperEarlyStop/preprocessed_dataset/train_loss_training.csv")