RNN Acceptors and BiRNN Transducers

This project contains the solutions and explanations for the tasks described in Exercise3.pdf. These tasks mainly revolve around building and manipulating Recurrent Neural Networks (RNNs), including RNN Acceptors and Bi-directional RNN (BiRNN) Transducers.

Installation

To install the necessary packages for this project, please run:

!pip install pandas matplotlib tensorboard numpy sympy scikit-learn ordered-set -q

Contents

Exercise Definitions and Tasks

The definitions and tasks for this project are described in detail in the Exercise3.pdf file. You can find this file in the root directory of the project.

Configurations

The project uses a configuration file named config.py for tweaking the parameters and specifying directories. Feel free to modify this file according to your requirements.

Code

The actual implementations of the tasks can be found under the code directory. This includes the generation of data, creation of the RNN acceptor network, acceptor capabilities, and the BiLSTM tagger. Each of these parts is discussed in detail below.

Part 1: Generating the Data

For Assignment 1.2, run gen_examples.py. This will generate positive and negative examples, with 500 samples each. These examples will be stored in the data/1/samples directory. The corresponding train and test sets are also provided in the same directory.

Part 2: Writing the RNN Acceptor Network

For Assignment 1.3, run experiment.py. This will train an RNN acceptor on the train set and validate it on the test set.

Part 3: Acceptor Capabilities

For this part, try to challenge the model and make it fail. The file part2.py can be run to train and validate the model on the PRIME_1 problem as described in the PDF file.

Part 4: BiLSTM Tagger

To train a BiLSTM model, use the following command:

bilstmTrain.py repr trainFile modelFile

Here, repr is one of the options: a, b, c, d; trainFile is the input file for training (expected to have a space delimiter); and modelFile is the directory to save the trained model. This process also creates dataset wrappers in the directory defined in config.py.

To predict using the trained BiLSTM model, use the following command:

bilstmPredict.py repr modelFile inputFile

Here, modelFile is the path to the trained model, and inputFile is the file on which to make predictions. The predictions will be saved to the prediction file in the data folder.

Notebooks

For a more interactive experience, you can use the provided notebooks in the notebooks directory. These notebooks guide you through each step of the project, using the models and functions described above.