CS685 Final Project Repo

Description / Notes

• This repo contains all of the code written for our final CS 685 project.

• All trained models and datasets used are hosted in the below drive folder.

  • Data Folder: https://drive.google.com/drive/folders/10HFYsf_Z9G6k84Cls0kp3squUr9iqfNo?usp=sharing
  • It is not reccomended to download the entire folder at once as it is very large (~100+ gigabytes) and drive often leaves out large binary files unless explicitly zipped with an external tool. Furthermore, many folders will download as seperate zip files cluttering downloads.

• All of the jupyter notebooks included here were run on either google colab or a Deep learning VM instance on GCP. For most of the notebooks, model checkpoints for the best (based on validation data) and most recent epoch get saved to drive.

Files

• Classifier_train_eval.ipynb
  • Trains binary classifier to distinguish between stylized and normal text for each of the 4 stylizing models (twitter, poetry, lyrics, formality). Evaluates on test dataset that has 5% of outputs randomly replaced with stylized phrases.
  • Training takes ~5 hours. Evaluation takes ~45 minutes. Times are averages of running on P100 or V100 GPUs
  • Notes when running
    • Logging for tensorboard isn't working.
    • File paths need to be manually changed for each model being trained / evaluated
• Style_paraphrase.ipynb
  • Praphrases IWSLT14 dataset (188,204 lines) with one of four models.
    • CDS models take 10 - 15 hours to run while formality / shakespear take 50+ on V100 GPUs hours.
    • Ouputted text is saved as a textfile to "IWSLT/2_Style_Paraphrased/". Output length does not always match the input file length and often requires manually adjusting 2-5 lines that have extra crlf .
• watermarking_approach_1.ipynb
  • Implements approach 1 watermarking as described in paper
  • Trains Victim model on original and stylized data. Output from the victim model is used to train an attacker model. Evaluation is performed using BLEU4 scores.
  • Training time is ~4 hours per model being trained and ~10 hours total per p value and replacement % combination.
• watermarking_approach_2.ipynb
  • Implements approach 2 watermarking as described in paper
  • Trains Victim model on original and stylized data. Output from the victim model is used to train an attacker model. Evaluation is performed using BLEU4 scores.
  • Training time is ~4 hours per model being trained and ~10 hours total per p value and replacement % combination.
• compute_gradient.ipynb
  • Implements the maximizes the angular deviation method in Imitation Attacks and Defenses for Black-box Machine Translation Systems from scratch with a little modification.
  • Instead of computing the whole model's gradients, we only consider the embedding layer.
  • The training data is split into batches for parallel computing. It took at least 10 days to finish the whole process (BLEU-threshold=0.8) when using a Google Colab Pro account with three running pages.
• LM.ipynb
  • Use the candidates obtained from the above approace to train a language model (LM) by leveraging fairseq.
  • Generate the alternatvie translations by considering victim model and LM through a simple linear combination.
• replac_with_syn.ipynb
  • Randomly replace words in the victim output with their synonym based on WordNet.
  • Compute the gradients again to find the best candidate.

Folders

• Preprocessing IWSLT
  • split_lines.py
    • Used to split the original data into smaller chunks to allow parallel processing when paraphrasing.
  • combine_output.py
    • Used to combine the paraphrased output of the files split using the split_line.py sc
  • match_line.py
    • Created tags for the parphrased training data.
    • Prepares data to be tokenized and properly split by fairseq's prepare-iwslt14.sh script.
  • getSize.py
    • Code snippet ot get size of all the split data for various styles.
• Graphing
  • Python code to generate pyplots based on input csv's.
  • Requires pandas, pyplot.
  • Output files are not saved but displayed as manual frame adjustments are needed to prevent overlap of axis.
• drive_download
  • Very small script to allow abitrary drive files to be downloaded to GCP based on file id.
• Latex
  • Scripts to convert csv's to latex tables for use in final report.