Official implementation of GraphMask as presented in our paper Interpreting Graph Neural Networks for NLP With Differentiable Edge Masking.

Last update: Sep 2, 2022

Related tags

Overview

GraphMask

This repository contains an implementation of GraphMask, the interpretability technique for graph neural networks presented in our ICLR 2021 paper Interpreting Graph Neural Networks for NLP With Differentiable Edge Masking.

Requirements

We include a requirements.txt file for the specific environment we used to run the code. To run the code, please either set up your environment to match that, or verify that you have the following dependencies:

Python 3
PyTorch 1.8.1
PyTorch Geometric 1.7
AllenNLP 0.9.0
SpaCy 2.1.9

Running the Code

We include models and interpreters for our synthetic task, for the question answering model by De Cao et al. (2019), and for the SRL model by Marcheggiani and Titov (2017).

To train a model, use our script by replacing [configuration] in the following with the appropriate file (default is the synthetic task, configurations/star_graphs.json):

python train_model.py --configuration \[configuration\]

Once you have trained the model, train and validate GraphMask by running:

python run_analysis.py --configuration \[configuration\]

For the synthetic task, you can optionally add a comparison between the performance of GraphMask and the faithfulness gold standard as follows:

python run_analysis.py --configuration \[configuration\] --gold_standard

To experiment with other analysis techniques, you can change the analysis strategy in the configuration file.

Downloading Data

For both tasks, download the 840B Common Crawl GloVe embeddings and place the file in data/glove/. For the question answering task, download the Qangaroo dataset and place the files in data/qangaroo_v1.1/. For the SRL task, follow the instructions here to download the CoNLL-2009 dataset and generate vocabulary files. Place both dataset and vocabulary files in data/conll2009/.

Citation

Please cite our paper if you use this code in your own work:

@inproceedings{
   schlichtkrull2021interpreting,
   title={Interpreting Graph Neural Networks for {\{}NLP{\}} With Differentiable Edge Masking},
   author={Michael Sejr Schlichtkrull and Nicola De Cao and Ivan Titov},
   booktitle={International Conference on Learning Representations},
   year={2021},
   url={https://openreview.net/forum?id=WznmQa42ZAx}
}

PyTorch implementation of our Adam-NSCL algorithm from our CVPR2021 (oral) paper "Training Networks in Null Space for Continual Learning"

Adam-NSCL This is a PyTorch implementation of Adam-NSCL algorithm for continual learning from our CVPR2021 (oral) paper: Title: Training Networks in N

34 Dec 21, 2022

A weakly-supervised scene graph generation codebase. The implementation of our CVPR2021 paper ``Linguistic Structures as Weak Supervision for Visual Scene Graph Generation''

README.md shall be finished soon. WSSGG 0 Overview 1 Installation 1.1 Faster-RCNN 1.2 Language Parser 1.3 GloVe Embeddings 2 Settings 2.1 VG-GT-Graph

35 Nov 20, 2022

Official code release for ICCV 2021 paper SNARF: Differentiable Forward Skinning for Animating Non-rigid Neural Implicit Shapes.

235 Dec 26, 2022

Official implementation of the ICML2021 paper

Comments

Error for QA task.

ConfigurationError Traceback (most recent call last) in () 9 10 model_trainer = ModelTrainer(configuration, gpu=args.gpu) ---> 11 model_trainer.train()

in train(self) 19 20 def train(self): ---> 21 problem = self.experiment_utils.build_problem() 22 model = self.experiment_utils.build_model() 23

/content/GraphMask/codes/utils/experiment_utils.py in build_problem(self) 31 problem_class = QAProblem 32 ---> 33 problem = problem_class(self.configuration) 34 35 return problem

/content/GraphMask/codes/problems/qa/qa_problem.py in init(self, configuration) 44 # self.predictor = pretrained.load_predictor("coref") 45 self.predictor = Predictor.from_path( ---> 46 "https://s3-us-west-2.amazonaws.com/allennlp/models/coref-model-2018.02.05.tar.gz") 47 # self.predictor = pretrained.load_predictor("coref-spanbert") 48 # self.predictor = Predictor.from_path(

/usr/local/lib/python3.7/dist-packages/allennlp/predictors/predictor.py in from_path(cls, archive_path, predictor_name, cuda_device, dataset_reader_to_load, frozen, import_plugins, overrides)

/usr/local/lib/python3.7/dist-packages/allennlp/models/archival.py in load_archive(archive_file, cuda_device, overrides, weights_file) 203 else: 204 logger.warning(f"Archived file {replacement_filename} not found! At train time " --> 205 f"this file was located at {original_filename}. This may be " 206 "because you are loading a serialization directory. Attempting to " 207 "load the file from its train-time location.")

/usr/local/lib/python3.7/dist-packages/allennlp/models/archival.py in _load_dataset_readers(config, serialization_dir) 229 serialization_dir=serialization_dir, 230 cuda_device=cuda_device) --> 231 232 return Archive(model=model, config=config) 233

/usr/local/lib/python3.7/dist-packages/allennlp/common/from_params.py in from_params(cls, params, constructor_to_call, constructor_to_inspect, **extras)

/usr/local/lib/python3.7/dist-packages/allennlp/common/params.py in pop_choice(self, key, choices, default_to_first_choice, allow_class_names) 350 This is to allow e.g. specifying a model type as my_library.my_model.MyModel 351 and importing it on the fly. Our check for "looks like" is extremely lenient --> 352 and consists of checking that the value contains a '.'. 353 """ 354 default = choices[0] if default_to_first_choice else self.DEFAULT

ConfigurationError: coref not in acceptable choices for dataset_reader.type: ['babi', 'conll2003', 'interleaving', 'multitask', 'multitask_shim', 'sequence_tagging', 'sharded', 'text_classification_json']. You should either use the --include-package flag to make sure the correct module is loaded, or use a fully qualified class name in your config file like {"model": "my_module.models.MyModel"} to have it imported automatically.

opened by neolifer 2
WARNING:allennlp.models.model:Encountered the antecedent_indices key in the model's return dictionary which couldn't be split by the batch size. Key will be ignored.
Hi, I run your code to train the model on QA dataset using command python train_model.py --configuration configurations/qa.json. However, there is a WARNING bug when it begins as:

Building the train set (or loading from cache if possible): 4%| | 1965/43738 [00:09<04:01, WARNING:allennlp.models.model:Encountered the antecedent_indices key in the model's return dictionary which couldn't be split by the batch size. Key will be ignored.

And then, the Building the train set process have no response all the day.

Why is that? Can you help me with it?

My Environment:

Python 3

PyTorch 1.7.1

PyTorch Geometric 1.7.2

AllenNLP 0.9.0

SpaCy 2.1.9
opened by louiswng 2

Official implementation of GraphMask as presented in our paper Interpreting Graph Neural Networks for NLP With Differentiable Edge Masking.

Related tags

Overview

GraphMask

You might also like...

PyTorch implementation of our Adam-NSCL algorithm from our CVPR2021 (oral) paper "Training Networks in Null Space for Continual Learning"

A weakly-supervised scene graph generation codebase. The implementation of our CVPR2021 paper ``Linguistic Structures as Weak Supervision for Visual Scene Graph Generation''

Official code release for ICCV 2021 paper SNARF: Differentiable Forward Skinning for Animating Non-rigid Neural Implicit Shapes.

Official implementation of the ICML2021 paper "Elastic Graph Neural Networks"

The source code of the paper "Understanding Graph Neural Networks from Graph Signal Denoising Perspectives"

git《Self-Attention Attribution: Interpreting Information Interactions Inside Transformer》(AAAI 2021) GitHub:

A Closer Look at Invalid Action Masking in Policy Gradient Algorithms

[CVPR 2020] Interpreting the Latent Space of GANs for Semantic Face Editing

ALFRED - A Benchmark for Interpreting Grounded Instructions for Everyday Tasks

Comments

Error for QA task.

WARNING:allennlp.models.model:Encountered the antecedent_indices key in the model's return dictionary which couldn't be split by the batch size. Key will be ignored.

Owner

Michael Schlichtkrull

[CVPR 2021] A Peek Into the Reasoning of Neural Networks: Interpreting with Structural Visual Concepts

We have implemented shaDow-GNN as a general and powerful pipeline for graph representation learning. For more details, please find our paper titled Deep Graph Neural Networks with Shallow Subgraph Samplers, available on arXiv (https//arxiv.org/abs/2012.01380).

U-Net Implementation: Convolutional Networks for Biomedical Image Segmentation" using the Carvana Image Masking Dataset in PyTorch

Official PyTorch implementation for Generic Attention-model Explainability for Interpreting Bi-Modal and Encoder-Decoder Transformers, a novel method to visualize any Transformer-based network. Including examples for DETR, VQA.

Official implementation of the network presented in the paper "M4Depth: A motion-based approach for monocular depth estimation on video sequences"

The official codes of our CVPR2022 paper: A Differentiable Two-stage Alignment Scheme for Burst Image Reconstruction with Large Shift

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

Code & Data for the Paper "Time Masking for Temporal Language Models", WSDM 2022

Code to reproduce the results in the paper "Tensor Component Analysis for Interpreting the Latent Space of GANs".

Code for our EMNLP 2021 paper “Heterogeneous Graph Neural Networks for Keyphrase Generation”