Code for the paper "Implicit Representations of Meaning in Neural Language Models"

Belinda Li

Last update: Nov 3, 2022

Related tags

Deep Learning state-probes

Overview

Implicit Representations of Meaning in Neural Language Models

Preliminaries

Create and set up a conda environment as follows:

conda create -n state-probes python=3.7
conda activate state-probes
pip install -r requirements.txt

Install the appropriate torch 1.7.0 for your cuda version:

conda install pytorch==1.7.0 cudatoolkit=<cuda_version> -c pytorch

Before running any command below, run

export PYTHONPATH=.
export TOKENIZERS_PARALLELISM=true

Data

The Alchemy data is downloaded from their website.

wget https://nlp.stanford.edu/projects/scone/scone.zip
unzip scone.zip

The synthetic version of alchemy was generated by running:

echo 0 > id #the code requires a file called id with a number in it ...
python alchemy_artificial_generator.py --num_scenarios 3600 --output synth_alchemy_train
python alchemy_artificial_generator.py --num_scenarios 500 --output synth_alchemy_dev
python alchemy_artificial_generator.py --num_scenarios 900 --output synth_alchemy_test

You can also just download our generated data through:

wget http://web.mit.edu/bzl/www/synth_alchemy.tar.gz
tar -xzvf synth_alchemy.tar.gz

The Textworld data is under

wget http://web.mit.edu/bzl/www/tw_data.tar.gz
tar -xzvf tw_data.tar.gz

LM Training

To train a BART or T5 model on Alchemy data

python scripts/train_alchemy.py \
    --arch [t5|bart] [--no_pretrain] \
    [--synthetic] --encode_init_state NL

Saves model checkpoints under sconeModels/*.

To train a BART or T5 model on Textworld data

python scripts/train_textworld.py \
    --arch [t5/bart] [--no_pretrain] \
    --data tw_data/simple_traces --gamefile tw_games/simple_games

Saves model checkpoints nder twModels/*.

Probe Training & Evaluation

Alchemy

The main probe command is as follows:

python scripts/probe_alchemy.py \
    --arch [bart|t5] --lm_save_path <path_to_lm_checkpoint> [--no_pretrain] \
    --encode_init_state NL --nonsynthetic \
    --probe_target single_beaker_final.NL --localizer_type single_beaker_init_full \
    --probe_type linear --probe_agg_method avg \
    --encode_tgt_state NL.[bart|t5] --tgt_agg_method avg \
    --batchsize 128 --eval_batchsize 1024 --lr 1e-4

For evaluation, add --eval_only --probe_save_path <path_to_probe_checkpoint>. This will save model predictions to a .jsonl file under the same directory as the probe checkpoint.

Add --control_input for No LM experiment.

Change --probe_target to single_beaker_init.NL to decode initial state.

For localization experiments, set --localizer_type single_beaker_init_{$i}.offset{$off} for some token i in {article, pos.[R0|R1|R2], beaker.[R0|R1], verb, amount, color, end_punct} and some integer offset off between 0 and 6.

Saves probe checkpoints under probe_models_alchemy/*.

Intervention experiment results follow from running the script:

python scripts/intervention.py \
    --arch [bart|t5] \
    --encode_init_state NL \
    --create_type drain_1 \
    --lm_save_path <path_to_lm_checkpoint>

which creates two contexts and replaces a select few encoded tokens to modify the underlying belief state.

Textworld

Begin by creating the full set of encoded proposition representations

python scripts/get_all_tw_facts.py \
    --data tw_data/simple_traces --gamefile tw_data/simple_games \
    --state_model_arch [bart|t5] \
    --probe_target belief_facts_pair \
    --state_model_path [None|pretrain|<path_to_lm_checkpoint>] \
    --out_file <path_to_prop_encodings>

Run the probe with

python scripts/probe_textworld.py \
    --arch [bart|t5] --data tw_data/simple_traces --gamefile tw_data/simple_games \
    --probe_target final.full_belief_facts_pair --encode_tgt_state NL.[bart|t5] \
    --localizer_type belief_facts_pair_[first|last|all] --probe_type 3linear_classify \
    --probe_agg_method avg --tgt_agg_method avg \
    --lm_save_path <path_to_lm_checkpoint> [--no_pretrain] \
    --ents_to_states_file <path_to_prop_encodings> \
    --eval_batchsize 256 --batchsize 32

For evaluation, add --eval_only --probe_save_path <path_to_probe_checkpoint>. This will save model predictions to a .jsonl file under the same directory as the probe checkpoint.

Add --control_input for No LM experiment.

Change --probe_target to init.full_belief_facts_pair to decode initial state.

For remap experiments, change --probe_target to final.full_belief_facts_pair.control_with_rooms.

For decoding from just one side of propositions, replace any instance of belief_facts_pair (in --probe_target and --localizer_type) with belief_facts_single and rerun both commands (first get the full proposition encodings, then run the probe).

Saves probe checkpoints under probe_models_textworld/*.

Print Metrics

Print full metrics (state EM, entity EM, subdivided by relations vs. propositions, etc.) using scripts/print_metrics.py.

python scripts/print_metrics.py \
    --arch [bart|t5] --domain [alchemy|textworld] \
    --pred_files <path_to_model_predictions_1>,<path_to_model_predictions_2>,<path_to_model_predictions_3>,... \
    [--use_remap_domain --remap_fn <path_to_remap_model_predictions>] \
    [--single_side_probe]

Code for the paper Learning the Predictability of the Future

Learning the Predictability of the Future Code from the paper Learning the Predictability of the Future. Website of the project in hyperfuture.cs.colu

Computer Vision Lab at Columbia University

139 Nov 18, 2022

PyTorch code for the paper: FeatMatch: Feature-Based Augmentation for Semi-Supervised Learning

FeatMatch: Feature-Based Augmentation for Semi-Supervised Learning This is the PyTorch implementation of our paper: FeatMatch: Feature-Based Augmentat

43 Nov 19, 2022

Code for the paper A Theoretical Analysis of the Repetition Problem in Text Generation

A Theoretical Analysis of the Repetition Problem in Text Generation This repository share the code for the paper "A Theoretical Analysis of the Repeti

37 Nov 21, 2022

Code for our ICASSP 2021 paper: SA-Net: Shuffle Attention for Deep Convolutional Neural Networks

SA-Net: Shuffle Attention for Deep Convolutional Neural Networks (paper) By Qing-Long Zhang and Yu-Bin Yang [State Key Laboratory for Novel Software T

199 Jan 8, 2023

Open source repository for the code accompanying the paper 'Non-Rigid Neural Radiance Fields Reconstruction and Novel View Synthesis of a Deforming Scene from Monocular Video'.

Non-Rigid Neural Radiance Fields This is the official repository for the project "Non-Rigid Neural Radiance Fields: Reconstruction and Novel View Synt

296 Dec 29, 2022

Comments

Question about interpreting probe metrics
Hi Belinda!

I am a student from Princeton, and a couple friends and I are trying to reproduce results of your 2021 paper as part of a class project. We encountered two questions regarding the output of scripts/print_metrics.py, and would really appreciate some clarifications:

When we printed evaluation metrics of the probes specified in Table 1, the output showed two kinds of metrics: "All queried results" and "All full results" (see screenshot below, showing results my team got from evaluating a t5 probe finetuned on textWorld and localized to first). Do you mind clarifying what's the difference between how State EM was calculated in the two kinds of results?

Line 175 of scripts/print_metrics.py mentions the directory state-probes-TW/tw_games. After following instructions on the readme file to download data files, I couldn't find such a folder (the closest I could find is tw_data). Could you clarify what the former directory contains, and if the content is the same as the directory tw_data? When I ran scripts/print_metrics.py, I had to change the path state-probes-TW/tw_games to tw_data to get the program to run, and I wonder if this revision contributed to us getting zero for the State EM under "All full results".

Thank you in advance for your help!

Best, Tony
opened by junzeye 2
Undefined variables

Hi! Could you please share the definitions of state_keys and tgt_state_keys variables here? Any description of the else condition and of the task being achieved in this condition would also really help.

opened by sm354 1

Code for the paper "Implicit Representations of Meaning in Neural Language Models"

Related tags

Overview

Implicit Representations of Meaning in Neural Language Models

Preliminaries

Data

LM Training

Probe Training & Evaluation

Alchemy

Textworld

Print Metrics

You might also like...

Code for the paper Learning the Predictability of the Future

PyTorch code for the paper: FeatMatch: Feature-Based Augmentation for Semi-Supervised Learning

Code for the paper A Theoretical Analysis of the Repetition Problem in Text Generation

Code for our ICASSP 2021 paper: SA-Net: Shuffle Attention for Deep Convolutional Neural Networks

Open source repository for the code accompanying the paper 'Non-Rigid Neural Radiance Fields Reconstruction and Novel View Synthesis of a Deforming Scene from Monocular Video'.

Code for the Shortformer model, from the paper by Ofir Press, Noah A. Smith and Mike Lewis.

PyTorch code for ICLR 2021 paper Unbiased Teacher for Semi-Supervised Object Detection

Official code for paper "Optimization for Oriented Object Detection via Representation Invariance Loss".

Code for our CVPR 2021 paper "MetaCam+DSCE"

Comments

Question about interpreting probe metrics

Undefined variables

Owner

Belinda Li

Inference code for "StylePeople: A Generative Model of Fullbody Human Avatars" paper. This code is for the part of the paper describing video-based avatars.

This is the official source code for SLATE. We provide the code for the model, the training code, and a dataset loader for the 3D Shapes dataset. This code is implemented in Pytorch.

Code for paper ECCV 2020 paper: Who Left the Dogs Out? 3D Animal Reconstruction with Expectation Maximization in the Loop.

TensorFlow code for the neural network presented in the paper: "Structural Language Models of Code" (ICML'2020)

Code for the prototype tool in our paper "CoProtector: Protect Open-Source Code against Unauthorized Training Usage with Data Poisoning".

Code to use Augmented Shapiro Wilks Stopping, as well as code for the paper "Statistically Signifigant Stopping of Neural Network Training"

Code for our method RePRI for Few-Shot Segmentation. Paper at http://arxiv.org/abs/2012.06166

Code for ACM MM 2020 paper "NOH-NMS: Improving Pedestrian Detection by Nearby Objects Hallucination"

Official TensorFlow code for the forthcoming paper

This is the code for the paper "Contrastive Clustering" (AAAI 2021)