In this assginment, we’ll be moving on from traditional n-gram based language models to more advanced forms of language modeling using neural networks. Specifically, we’ll be setting up a character-level recurrent neural network, known as a char-rnn for short.
Andrej Karpathy, previously a researcher at OpenAI, has written an excellent blog post about using RNNs for language models, which you should read before beginning this assignment. The title of his blog post is The Unreasonable Effectiveness of Recurrent Neural Networks.
Karpathy shows how char-rnns can be used to generate texts for several fun domains:
In this assignment you will follow a PyTorch tutorial to implement your own char-rnn, and then test it on a dataset of your choice. You will also train on our provided training set, and submit to the leaderboard.
Here are the materials that you should download for this assignment:
Please look at the FAQ section before you start working. Before submitting, please make sure to review the checklist on Piazza!
PyTorch is one of the most popular deep learning frameworks in both industry and academia, and learning its use will be invaluable should you choose a career in deep learning. You will be using PyTorch for this assignment, we ask you to build off a couple PyTorch tutorials.
PyTorch abstracts the back-propogation process from us, allowing us to define neural network structures and use a generic .backward()
function to compute the gradients that are later used in gradient descent (PyTorch also implements such optimization algorithms for us).
PyTorch does all of this for us by maintaining a computational graph, which allows differentiation to happen automatically! Don’t worry if you don’t remember your chain rules from MATH 114. Another nice thing about PyTorch is that it makes strong use of both object-oriented and functional programming paradigms, which makes reading and writing PyTorch code very accessible to previous programmers.
GPU
under Change Runtime Type
in the Runtime
menu.Read through the tutorial here that builds a char-rnn that is used to classify baby names by their country of origin. It is recommended that you can reproduce the tutorial’s results on the provided name dataset before moving on, since the neural network architectures remain largely the same. Make sure you try your best to understand the dimensions of each layer (e.g. which ones can stay the same, and which are hyperparameters for us to tweak).
Download the city names dataset:
Modify the tutorial code to instead read from the city names dataset that we used in the previous assignment. The tutorial code problematically used the same text file for both training and evaluation. We learned in class about how this is not a great idea. For the city names dataset we provide you separate train and validation sets, as well as a test file for the leaderboard.
All training should be done on the train set and all evaluation (including confusion matrices and accuracy reports) on the validation set. You will need to change the data processing code to get this working. In addition, to handle unicode, you might need to replace calls to open
with calls to codecs.open(filename, "r",encoding='utf-8', errors='ignore')
.
Warning: you’ll want to lower the learning rating to 0.002 or less or you might get NaNs when training.
Experimentation and Analysis
Complete the following analysis on the city names dataset, and include your finding in the report.
Leaderboard
Write code to make predictions on the provided test set. The test set has one unlabeled city name per line. Your code should output a file labels.txt
with one two-letter country code per line. Extra credit will be given to the top 5 leaderboard submissions. Here are some ideas for improving your leaderboard performance:
In your report, describe your final model and training parameters.
Bookkeeping
Another tip for experimenting with neural network hyperparameters is to maintain notes (e.g. a spreadsheet or text-file) with different parameters and their resulting accuray. As you can imagine, there is a combinatorial explosion of the possible hyperparameter space so navigating it efficiently is best done by remembering your past experiments. Feel free to include this in your report as well.
In this section, you will be following more PyTorch tutorial code in order to reproduce Karpathy’s text generation results. Read through the tutorial here to get an understanding for the neural architecture, and then download this iPython notebook to base your own code on (it’s a bit easier to follow than the former).
You will notice that the code is quite similar to that of the classification problem. The biggest difference is in the loss function. For classification, we run the entire sequence through the RNN and then impose a loss only on the final class prediction. For the text generation task, we impose a loss at each step of the RNN on the predicted character. The classes in this second task are the possible characters to predict.
Be creative! Pick some dataset that interests you. Here are some ideas:
Potential extra credit will be given for creative and impressive methods of curating a text dataset for generation!
Include a sample of the text generated by your model, and give a qualitative discussion of the results. Where does it do well? Where does it seem to fail? Report perplexity on a couple validation texts that are similar and different to the training data. Compare your model’s results to that of an n-gram language model.
Here are the deliverables that you will need to submit:
labels.txt
predictions for leaderboard.Before submitting, please make sure to review the checklist on Piazza!
Autograders can be finicky, and Gradescope doesn’t let us change visibility of the error log. If you follow all instructions, you shouldn’t have an issue. Most of the issues come down to one of the following:
models.py
, main_classify.py
, main_generate.py
__init__
in CharRNNClassify
(IMPORTANT!)CharRNNClassify()
without any arguments.To debug, print out the shapes of your tensors! This usually is a good sanity check that your architecture is correct and that you are performing the right computations.
Use the command below. Please ensure that your model can be used for inference.
torch.save(model.state_dict(), PATH)
Use the command below.
model = CharRNNClassify()
model.load_state_dict(torch.load(PATH))
model.eval() #To predict
If you are new to the paradigm of computational graphs and functional programming, please have a look at this tutorial before getting started.
jupyter nbconvert --to script notebook.ipynb
The TA’s model, which passed all the testcases, had the following configuration:
torch.optim.SGD(model.parameters(), lr=0.005)
nn.NLLLoss()
Send the model and the input, output tensors to the GPU using .to(device)
. Refer the PyTorch docs for further information.
Noisy data is common when data is harvested automatically like the cities dataset. The onus is on the data scientist to ensure that their data is clean. However, for this assignment, you are not required to clean the dataset.
Neural Nets and Neural Language Models. Dan Jurafsky and James H. Martin. Speech and Language Processing (3rd edition draft) . |
The Unreasonable Effectiveness of Recurrent Neural Networks. Andrej Karpathy. Blog post. 2015. |
A Neural Probabilistic Language Model (longer JMLR version). Yoshua Bengio, Réjean Ducharme, Pascal Vincent and Christian Jauvin. Journal of Machine Learning Research 2003. |