X2T: Training an X-to-Text Typing Interface with Online Learning from User Feedback

This repository contains code for X2T applied to the gaze and handwriting domains described in the paper. This includes both a real gaze interface that can be operated on a computer with an available webcam, simulated offline gaze (by replaying data collected in the gaze user study for the paper), and simulated handwriting.

Setup Instructions

1. Run conda create -n YOUR_ENV_NAME python=3.7 to create a new conda env for this project and activate it.
2. Install the package requirements using pip install -r requirements.txt.
3. Install the local bci-typing package pip install -e ..
4. Install the gym_bci_typing environment using pip install -e gym-bci-typing/.
5. Make sure that CUDA 10.0 and cuDNN 7.6.5 for CUDA 10.0 are both installed if you want GPU support.
6. If running into issues with dlib (needed for the real gaze interface with a webcam), you may need to install dependencies. Instructions can be found here.

The following steps are needed to use the included Transformer-XL language model (for use with the simulated handwriting domain).

1. Run transformer_xl/tf/sota/download.sh to download the pretrained Transformer-XL model on One Billion Words.
2. Run transformer_xl/tf/fix_cached_data.py to fix the pretrained model's data cache to work with the project's directory structure. Alternatively, after running the above download script, replace the folder transformer_xl/pretrained_xl/tf_lm1b/data with the included transformer_xl/pretrained_xl/tf_lm1b/data_fixed folder (renaming the folder to data as well).

Running Experiments

The train.py script is used for all experiment types. Run python train.py --help for information on all command line arguments.

Each experiment is logged to experiments/USER_TYPE/SAVE_NAME, where USER_TYPE is the user type of the experiment (e.g. real_gaze) and SAVE_NAME is specified by the flag -s SAVE_NAME. By default, SAVE_NAME is the date and time of the experiment (e.g. 2020-08-21--18-20-27). This folder may contain each of the following, when appropriate for the parameters of the experiment:

• log_1 contains the TensorBoard training log for adaptive sessions.
• offline_log_1 contains the TensorBoard offline training log.
• data.hdf5 is the data for each step recorded during the session (can be used by -dp or -gp for future sessions). It contains 4 keys ('obses', 'actions', 'rewards', 'targets').
  • 'obses' is a (timesteps, n_samples, input dim) shaped array containing the inputs provided by the user. In the data from the gaze study, n_samples is the default value of 10. n_samples is always 1 in the simulated handwriting domain.
  • 'actions' is a (timesteps,) shaped array, containing the action selected by the interface at each timestep.
  • 'rewards' is a (timesteps,) shaped array, containing 0 if the user backspaced the action for that timestep, and 1 otherwise.
  • 'targets' is a (timesteps,) shaped array, containing the true action indices that the user intended to select each timestep.
• calibration_data.hdf5 is the data collected by real gaze users to calibrate default interfaces (can be used by -gp for future sessions). It contains 2 keys ('signals', 'gaze_labels').
  • 'signals' is a (dataset size, input dim) shaped array containing the gaze feature inputs provided by the user during the calibration phase.
  • 'gaze_labels' is a (dataset size, 2) shaped array containing the normalized 2D coordinates of the
    locations that the user was directed to look at during the calibration phase. It is aligned with 'signals', such that each index in the first dimension represents an (input, coordinate) pair for supervised learning.
  • This file also contains an attributes object (.attrs), which contains the 'cam_coord', 'height', 'width', 'window_height', and 'window_width' of the user's setup during the session calibrated by this data.
• final_model.zip contains the final model weights (can be passed to -l for future sessions).
• offline_model.zip contains the model weights after offline pretraining. (can be passed to -l for future sessions).
• metrics.txt contains the overall prediction accuracy of the session.
• params.json contains the parameters of the session.
• baseline_estimates.pkl contains the 2D normalized gaze position estimates of the baseline default interface. It is a n_actions length list (default 8), and each element is a list of normalized 2D estimates that the default interface predicted for that action.

Also, if running in baseline mode, the data.hdf5 file will be copied to experiments/USER_TYPE/recent/baseline_data.hdf5, which can be loaded from in the future using -ur.

Replicating X2T Paper Experiments

Real Webcam Gaze

1. Run python train.py -u rg -m b -n N_STEPS to run the default interface that uses a calibrated nearest-action baseline for the desired number of steps. The interface will first run through a calibration procedure where you are directed to look at various positions on the screen. Then the typing interface will start. This command was used in gaze user study for both collecting data to pretrain the learned reward model in the adaptive sessions, and running default interface sessions to compare with the adaptive sessions.

2. Run python train.py -u rg -ur -n N_STEPS to pretrain the learned reward model of the interface using the offline default interface data collected in the previous step, and then start an adaptive session for the desired number of steps, using the mixed learned + default policy. The interface will first run through a calibration procedure where you are directed to look at various positions on the screen. Then the typing interface will start.

Additional Flags

For using the calibrated default baseline with real gaze (any -m flag except l), the following arguments should also be set according to the user's webcam and screen setup:

• -cx CAM_X the horizontal displacement (cm) of the webcam from the left end of the interface window. Positive values indicate rightward displacement.
• -cy CAM_Y the vertical displacement (cm) of the webcam from the top end of the interface window. Positive values indicate downward displacement.
• -ww WINDOW_WIDTH The width (cm) of the interface window.
• -wh WINDOW_HEIGHT The height (cm) of the interface window.

The default values are for a 13-inch MacBook Pro in fullscreen.

How to Use the Interface:

If using in fullscreen mode (default), make sure your computer only has 1 monitor. Try to keep your face in the relative center of your webcam, as if your face is outside of the center square region of your webcam (the webcam recording is horizontally truncated to match the height of the recording), it may not be detected. After the interface has loaded, there will be n_actions different numbers, arranged in a circle. Press SPACE to begin the calibration phase. The number 0 will be colored orange, and you should stare at it until the next number is colored orange. This process will repeat for all numbers, for a total of two complete cycles. If the orange number is not changing after a short period of time, it is because your face is not being detected, as the interface will only continue to the next number once enough data for that number is recorded.

Once calibration is complete, you will be presented with a sentence to type. The top of the screen will display this sentence, as well what has been typed so far, and the next word in the sentence to type. There will be n_actions different words arranged in a circle, in the same positions as the calibration points. One of these words will always be the next word to type. Once you have located this word, you should stare at it, then press SPACE to begin recording. After a short period of time, the interface will predict which word you intended to type, and add it to what has been typed so far. If the interface made an error, the newly typed word will appear red, and you should press BACKSPACE to undo this word and try again. Otherwise, you should continue to try typing the next word in the same manner as before. After typing the number of words in the target sentence, the interface will reset with a new sentence. The interface will close automatically when N_STEPS timesteps have passed (each word prediction by the model is a timestep).

You can only undo the most recent word, i.e. you cannot undo then undo again without trying to type another word. You can only undo a word before you press SPACE to start typing the next word. You cannot undo the last word in a sentence (it will always move on to the next sentence, regardless if it was correct or not). Press the ESC key at any time to terminate the program early. Press the p key during a step to pause the program. Press p while paused to unpause the program. After unpausing, any input collected during the previous interrupted step is discarded. No other keys are used.

Simulated Offline Gaze

All data from the gaze user study described in the paper is found in the folder experiments/gaze_study. The folder offline contains data used to pretrain the adaptive interfaces. The folder default contains data for the default interface sessions. The folder x2t contains data for the adaptive interface sessions. Each of these folders contains folders numbered 0-11, one for each of the 12 participants in the study. Each of these numbered folders contains the calibration_data.hdf5 and data.hdf5 associated with those sessions. The x2t folders also contain offline_model.zip, which are the weights for the models that were pretrained and used to initialize the adaptive sessions in the study and baseline_estimates.pkl, which contains the default interface baseline estimates on the data collected during the adaptive sessions in the study.

The -gp flag should always be set to a folder containing a data.hdf5 file produced by real gaze session. For using the calibrated default baseline with simulated gaze (any -m flag except l), the folder should also contain a calibration_data.hdf5 file produced by a real gaze session. Here is an example of a command to run a simulated offline gaze default interface session, using data from participant 0 during their adaptive session:

python train.py -u og -m b -gp experiments/gaze_study/x2t/0/

For simulated adaptive sessions, you can pretrain the learned reward model by setting the -dp flag to a data.hdf5 file produced by a session. For example, to pretrain using the data from participant 0 that was collected for this purpose in the study, you should add the flag -dp experiments/gaze_study/offline/0/data.hdf5. You can also initialize the learned reward model weights directly using the -l flag. For example, to initialize using the same pretrained initialization used for the adaptive session of participant 0, you should add the flag -l experiments/gaze_study/x2t/0/offline_model.zip.

Coadaptation Experiments

To recreate the coadaptation experiments in the appendix of the paper, run ./scripts/coadapt.sh. This script initializes learned reward models using corresponding pretrained initializations from the adaptive sessions in the study. The results will be saved in experiments/offline_gaze/x2t_on_default_data/USER_INDEX and experiments/offline_gaze/default_on_x2t_data/USER_INDEX, where USER_INDEX is a number 0-11, representing which user from the study the data came from.

Reward Noise Experiments

To recreate the reward noise sensitivity experiments in the appendix of the paper, run ./scripts/rew_noise.sh. This script initializes learned reward models using corresponding pretrained initializations from the adaptive sessions in the study. The results will be saved in experiments/offline_gaze/rew_noise/NOISE_LEVEL/USER_INDEX, where NOISE_LEVEL is the fraction the user's rewards that are incorrect (e.g. 0.1), and USER_INDEX is a number 0-11, representing which user from the study the data came from.

Simulated Handwriting

Ablation Experiments

To recreate the full simulated handwriting ablation experiments in the paper, run ./scripts/ablation.sh. The results will be saved to experiments/sim_uji/ablation/CONDITION/USER_INDEX, where CONDITION indicates the ablation condition (e.g. default, x2t, x2t_no_pretrain), and USER_INDEX is a number 0-59 representing which writer in the UJI Pen Characters v2 dataset was used.

Personalization Experiments

To recreate the full simulated handwriting personalization experiments in the paper, first recreate the ablation experiments as above, then run ./scripts/personalization.sh. The results will be saved to experiments/sim_uji/personalization/USER_TRAIN/USER_EVAL, where USER_TRAIN is the writer in the UJI Pen Characters v2 dataset that the model was trained on, and USER_EVAL is the writer that the model was evaluated on.

Handwriting Data Format

The file online_handwriting.pkl contains data representing pen strokes from the UJI Pen Characters v2 dataset. It is a dictionary with 27 keys, one for each lower case letter and space, e.g. 'a', ' '. Each key indexes into a length 120 list, 2 elements for each writer in the dataset. They are ordered as (writer 0, writer 0, writer 1, writer 1, ...). Each element of this list is a drawing of a character, and is represented list of arrays. Each of these arrays represents a pen stroke, and is of the shape (None, 2), where the first dimension represents the different positions in the stroke (in order), and the second dimension is the coordinate of each position. Each of the positions in a stroke should be connected in order by lines, and all the strokes together form a character. There should not be additional lines connecting the strokes in a character.

Hyperparameters

The default hyperparameters in train.py, which is used by all the above scripts, are those specified in the X2T paper. However, it is recommended to instead use a learning rate of 5e-4 in the gaze domain (both real and simulated) to have more stable performance.

Notebooks

See notebooks/plots.ipynb to create the plots in the paper, and notebooks/anova.ipynb to obtain the ANOVA results in the paper.

Citation

If you find this repository useful for your work, please cite:

@inproceedings{
gao2021xt,
title={X2T: Training an X-to-Text Typing Interface with Online Learning from User Feedback},
author={Jensen Gao and Siddharth Reddy and Glen Berseth and Nicholas Hardy and Nikhilesh Natraj and Karunesh Ganguly and Anca Dragan and Sergey Levine},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=LiX3ECzDPHZ}
}

Acknowledgements

Contains code, models, and data from the PyTorch implementation of "Eye Tracking for Everyone," found at https://github.com/CSAILVision/GazeCapture/tree/master/pytorch.

Kyle Krafka, Aditya Khosla, Petr Kellnhofer, Harini Kannan, Suchi Bhandarkar, Wojciech Matusik and Antonio Torralba. “Eye Tracking for Everyone”. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.

Contains code and models from "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context," found at https://github.com/kimiyoung/transformer-xl.

Dai, Zihang, et al. "Transformer-xl: Attentive language models beyond a fixed-length context." arXiv preprint arXiv:1901.02860 (2019).

Contact

For any questions, bugs, or suggestions, please feel free to contact jenseng@berkeley.edu.