/ULTRA

A foundation model for knowledge graph reasoning

Primary LanguagePythonMIT LicenseMIT

ULTRA: Towards Foundation Models for Knowledge Graph Reasoning

pytorch pyg ULTRA arxiv UltraQuery arxiv HuggingFace Hub license

ULTRA

PyG implementation of ULTRA and UltraQuery, a foundation model for KG reasoning. Authored by Michael Galkin, Zhaocheng Zhu, and Xinyu Yuan. Logo generated by DALL·E 3.

Overview

ULTRA is a foundation model for knowledge graph (KG) reasoning. A single pre-trained ULTRA model performs link prediction tasks on any multi-relational graph with any entity / relation vocabulary. Performance-wise averaged on 50+ KGs, a single pre-trained ULTRA model is better in the 0-shot inference mode than many SOTA models trained specifically on each graph. Following the pretrain-finetune paradigm of foundation models, you can run a pre-trained ULTRA checkpoint immediately in the zero-shot manner on any graph as well as use more fine-tuning.

ULTRA provides unified, learnable, transferable representations for any KG. Under the hood, ULTRA employs graph neural networks and modified versions of NBFNet. ULTRA does not learn any entity and relation embeddings specific to a downstream graph but instead obtains relative relation representations based on interactions between relations.

The original implementation with the TorchDrug framework is available here for reproduction purposes.

This repository is based on PyTorch 2.1 and PyTorch-Geometric 2.4.

Your superpowers ⚡️:

Table of contents:

Updates

  • Oct 1st, 2024: UltraQuery got accepted at NeurIPS 2024!
  • Apr 23rd, 2024: Release of UltraQuery for complex multi-hop logical query answering on any KG (with new checkpoint and 23 datasets).
  • Jan 15th, 2024: Accepted at ICLR 2024!
  • Dec 4th, 2023: Added a new ULTRA checkpoint ultra_50g pre-trained on 50 graphs. Averaged over 16 larger transductive graphs, it delivers 0.389 MRR / 0.549 Hits@10 compared to 0.329 MRR / 0.479 Hits@10 of the ultra_3g checkpoint. The inductive performance is still as good! Use this checkpoint for inference on larger graphs.
  • Dec 4th, 2023: Pre-trained ULTRA models (3g, 4g, 50g) are now also available on the HuggingFace Hub!

Installation

You may install the dependencies via either conda or pip. Ultra PyG is implemented with Python 3.9, PyTorch 2.1 and PyG 2.4 (CUDA 11.8 or later when running on GPUs). If you are on a Mac, you may omit the CUDA toolkit requirements.

From Conda

conda install pytorch=2.1.0 pytorch-cuda=11.8 cudatoolkit=11.8 pytorch-scatter=2.1.2 pyg=2.4.0 -c pytorch -c nvidia -c pyg -c conda-forge
conda install ninja easydict pyyaml -c conda-forge

From Pip

pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cu118
pip install torch-scatter==2.1.2 torch-sparse==0.6.18 torch-geometric==2.4.0 -f https://data.pyg.org/whl/torch-2.1.0+cu118.html
pip install ninja easydict pyyaml
Compilation of the `rspmm` kernel

To make relational message passing iteration O(V) instead of O(E) we ship a custom rspmm kernel that will be compiled automatically upon the first launch. The rspmm kernel supports transe and distmult message functions, others like rotate will resort to full edge materialization and O(E) complexity.

The kernel can be compiled on both CPUs (including M1/M2 on Macs) and GPUs (it is done only once and then cached). For GPUs, you need a CUDA 11.8+ toolkit with the nvcc compiler. If you are deploying this in a Docker container, make sure to start from the devel images that contain nvcc in addition to plain CUDA runtime.

Make sure your CUDA_HOME variable is set properly to avoid potential compilation errors, eg

export CUDA_HOME=/usr/local/cuda-11.8/

Checkpoints

We provide two pre-trained ULTRA checkpoints in the /ckpts folder of the same model size (6-layer GNNs per relation and entity graphs, 64d, 168k total parameters) trained on 4 x A100 GPUs with this codebase:

  • ultra_3g.pth: trained on FB15k237, WN18RR, CoDExMedium for 800,000 steps, config is in /config/transductive/pretrain_3g.yaml
  • ultra_4g.pth: trained on FB15k237, WN18RR, CoDExMedium, NELL995 for 400,000 steps, config is in /config/transductive/pretrain_4g.yaml

You can use those checkpoints for zero-shot inference on any graph (including your own) or use it as a backbone for fine-tuning. Both checkpoints are rather small (2 MB each).

Zero-shot performance of the checkpoints compared to the paper version (PyG experiments were run on a single RTX 3090, PyTorch 2.1, PyG 2.4, CUDA 11.8 using the run_many.py script in this repo):

Model Inductive (e) (18 graphs) Inductive (e,r) (23 graphs)
MRR Hits@10 MRR Hits@10
ULTRA (3g) Paper 0.430 0.566 0.345 0.512
ULTRA (4g) Paper 0.439 0.580 0.352 0.518
ULTRA (3g) PyG 0.420 0.562 0.344 0.511
ULTRA (4g) PyG 0.444 0.588 0.344 0.513

Run Inference and Fine-tuning

The /scripts folder contains 3 executable files:

  • run.py - run an experiment on a single dataset
  • run_many.py - run experiments on several datasets sequentially and dump results into a CSV file
  • pretrain.py - a script for pre-training ULTRA on several graphs

The yaml configs in the config folder are provided for both transductive and inductive datasets.

Run a single experiment

The run.py command requires the following arguments:

  • -c <yaml config>: a path to the yaml config
  • --dataset: dataset name (from the list of datasets)
  • --version: a version of the inductive dataset (see all in datasets), not needed for transductive graphs. For example, --dataset FB15k237Inductive --version v1 will load one of the GraIL inductive datasets.
  • --epochs: number of epochs to train, --epochs 0 means running zero-shot inference.
  • --bpe: batches per epoch (replaces the length of the dataloader as default value). --bpe 100 --epochs 10 means that each epoch consists of 100 batches, and overall training is 1000 batches. Set --bpe null to use the full length dataloader or comment the bpe line in the yaml configs.
  • --gpus: number of gpu devices, set to --gpus null when running on CPUs, --gpus [0] for a single GPU, or otherwise set the number of GPUs for a distributed setup
  • --ckpt: full path to the one of the ULTRA checkpoints to use (you can use those provided in the repo ot trained on your own). Use --ckpt null to start training from scratch (or run zero-shot inference on a randomly initialized model, it still might surprise you and demonstrate non-zero performance).

Zero-shot inference setup is --epochs 0 with a given checkpoint ckpt.

Fine-tuning of a checkpoint is when epochs > 0 with a given checkpoint.

An example command for an inductive dataset to run on a CPU:

python script/run.py -c config/inductive/inference.yaml --dataset FB15k237Inductive --version v1 --epochs 0 --bpe null --gpus null --ckpt /path/to/ultra/ckpts/ultra_4g.pth

An example command for a transductive dataset to run on a GPU:

python script/run.py -c config/transductive/inference.yaml --dataset CoDExSmall --epochs 0 --bpe null --gpus [0] --ckpt /path/to/ultra/ckpts/ultra_4g.pth

Run on many datasets

The run_many.py script is a convenient way to run evaluation (0-shot inference and fine-tuning) on several datasets sequentially. Upon completion, the script will generate a csv file ultra_results_<timestamp> with the test set results and chosen metrics. Using the same config files, you only need to specify:

  • -c <yaml config>: use the full path to the yaml config because workdir will be reset after each dataset;
  • -d, --datasets: a comma-separated list of datasets to run, inductive datasets use the name:version convention. For example, -d ILPC2022:small,ILPC2022:large;
  • --ckpt: ULTRA checkpoint to run the experiments on, use the full path to the file;
  • --gpus: the same as in run single;
  • -reps (optional): number of repeats with different seeds, set by default to 1 for zero-shot inference;
  • -ft, --finetune (optional): use the finetuning configs of ULTRA (default_finetuning_config) to fine-tune a given checkpoint for specified epochs and bpe;
  • -tr, --train (optional): train ULTRA from scratch on the target dataset taking epochs and bpe parameters from another pre-defined config (default_train_config);
  • --epochs and --bpe will be set according to a configuration, by default they are set for a 0-shot inference.

An example command to run 0-shot inference evaluation of an ULTRA checkpoint on 4 FB GraIL datasets:

python script/run_many.py -c /path/to/config/inductive/inference.yaml --gpus [0] --ckpt /path/to/ultra/ckpts/ultra_4g.pth -d FB15k237Inductive:v1,FB15k237Inductive:v2,FB15k237Inductive:v3,FB15k237Inductive:v4

An example command to run fine-tuning on 4 FB GraIL datasets with 5 different seeds:

python script/run_many.py -c /path/to/config/inductive/inference.yaml --gpus [0] --ckpt /path/to/ultra/ckpts/ultra_4g.pth --finetune --reps 5 -d FB15k237Inductive:v1,FB15k237Inductive:v2,FB15k237Inductive:v3,FB15k237Inductive:v4

Pretraining

Run the pre-training script pretrain.py with the config/transductive/pretrain_<ngraphs>.yaml config file.

graphs in the config specify the pre-training mixture. pretrain_3g.yaml uses FB15k237, WN18RR, CoDExMedium; pretrain_4g.yaml adds NELL995 to those three. By default, we use the training option fast_test: 500 to run faster evaluation on a random subset of 500 triples (that approximates full validation performance) of each validation set of the pre-training mixture. You can change the pre-training length by varying batches per epoch batch_per_epoch and epochs hyperparameters.

On the training graph mixture

Right now, 10 transductive datasets are supported for the pre-training mixture in the JointDataset:

  • FB15k237
  • WN18RR
  • CoDExSmall
  • CoDExMedium
  • CoDExLarge
  • NELL995
  • YAGO310
  • ConceptNet100k
  • DBpedia100k
  • AristoV4

You can add more datasets (from all 57 implemented as well as your custom ones) by modifying the datasets_map in datasets.py. By adding inductive datasets you'd need to add proper filtering datasets (similar to that in test() function in run.py) to have a consistent evaluation protocol.

An example command to start pre-training on 3 graphs:

python script/pretrain.py -c /path/to/config/transductive/pretrain_3g.yaml --gpus [0]

Pre-training can be computationally heavy, you might need to decrease the batch size for smaller GPU RAM. The two provided checkpoints were trained on 4 x A100 (40 GB).

Distributed setup

To run ULTRA with multiple GPUs, use the following commands (eg, 4 GPUs per node)

python -m torch.distributed.launch --nproc_per_node=4 script/pretrain.py -c config/transductive/pretrain.yaml --gpus [0,1,2,3]

Multi-node setup might work as well(not tested):

python -m torch.distributed.launch --nnodes=4 --nproc_per_node=4 script/pretrain.py -c config/transductive/pretrain.yaml --gpus [0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]

Datasets

The repo packs 57 different KG datasets of sizes from 1K-120K nodes and 1K-2M edges. Inductive datasets have splits of different version and a common notation is dataset:version, eg ILPC2022:small

Transductive datasets (16)
  • FB15k237, WN18RR, NELL995, YAGO310, CoDExSmall, CoDExMedium, CoDExLarge, Hetionet, ConceptNet100k, DBpedia100k, AristoV4 - full head/tail evaluation
  • WDsinger, NELL23k, FB15k237_10, FB15k237_20, FB15k237_50- only tail evaluation
Inductive (entity) datasets (18) - new nodes but same relations at inference time
  • 12 GraIL datasets (FB / WN / NELL) x (V1 / V2 / V3 / V4)
  • 2 ILPC 2022 datasets
  • 4 datasets from INDIGO
Dataset Versions
FB15k237Inductive v1, v2, v3, v4
WN18RRInductive v1, v2, v3, v4
NELLInductive v1, v2, v3, v4
ILPC2022 small, large
HM 1k, 3k, 5k, indigo
Inductive (entity, relation) datasets (23) - both new nodes and relations at inference time
  • 13 Ingram datasets (FB / WK / NL) x (25 / 50 / 75 / 100)
  • 10 MTDEA datasets
Dataset Versions
FBIngram 25, 50, 75, 100
WKIngram 25, 50, 75, 100
NLIngram 0, 25, 50, 75, 100
WikiTopicsMT1 tax, health
WikiTopicsMT2 org, sci
WikiTopicsMT3 art, infra
WikiTopicsMT4 sci, health
Metafam single version
FBNELL single version

All the datasets will be automatically downloaded upon the first run. It is recommended to first download pre-training datasets on single GPU experiments rather than immediately start multi-GPU training to prevent racing conditions.

Adding your own graph

We provide two base classes in datasets.py (based on InMemoryDataset of PyG) that you can inherit from:

  • TransductiveDataset requires 3 links in the urls field by convention urls = ["train_set_link", "valid_set_link", "test_set_link"] and name.
Code example
class CustomDataset(TransductiveDataset):

    urls = [
        "link/to/train.txt",
        "link/to/valid.txt",
        "link/to/test.txt",
        ]
    name = "custom_data"
  • InductiveDataset requires 4 links in the urls field by convention urls = ["transductive_train_set_link", "inference_graph_link", "inference_valid_set_link", "inference_test_set_link"] and name. By default, we assume that validation and test edges are based on inference_graph (but you can modify the loaders to account for different combinations).
Code example
class CustomDataset(InductiveDataset):

    urls = [
        "link/to/train.txt",
        "link/to/inference_graph.txt",
        "link/to/inference_valid.txt",
        "link/to/inference_test.txt",
        ]
    name = "custom_data"

TSV / CSV files are supported by setting a delimiter (eg, delimiter = "\t") in the class definition. After adding your own dataset, you can immediately run 0-shot inference or fine-tuning of any ULTRA checkpoint.

UltraQuery

You can now run complex logical queries on any KG with UltraQuery, an inductive query answering approach that uses any Ultra checkpoint with non-parametric fuzzy logic operators. Read more in the new preprint.

Similar to Ultra, UltraQuery transfers to any KG in the zero-shot fashion and sets a few SOTA results on a variety of query answering benchmarks.

Checkpoint

Any existing ULTRA checkpoint is compatible with UltraQuery but we also ship a newly trained ultraquery.pth checkpoint in the ckpts folder.

  • A new ultraquery.pth checkpoint trained on complex queries from the FB15k237LogicalQuery dataset for 40,000 steps, the config is in config/ultraquery/pretrain.yaml - the same ULTRA architecture but tuned for the multi-source propagation needed in complex queries (no need for score thresholding)
  • You can use any existing ULTRA checkpoint (3g / 4g / 50g) for starters - don't forget to set the --threshold argument to 0.8 or higher (depending on the dataset). Score thresholding is required because those models were trained on simple one-hop link prediction and there are certain issues (namely, the multi-source propagation issue, read Section 4.1 in the new preprint for more details)

Performance

The numbers reported in the preprint were obtained with a model trained with TorchDrug. In this PyG implementation, we managed to get even better performance across the board with the ultraquery.pth checkpoint.

EPFO is the averaged performance over 9 queries with relation projection, intersection, and union. Neg is the averaged performance over 5 queries with negation.

Model Total Average (23 datasets) Transductive (3 datasets) Inductive (e) (9 graphs) Inductive (e,r) (11 graphs)
EPFO MRR EPFO Hits@10 Neg MRR Neg Hits@10 EPFO MRR EPFO Hits@10 Neg MRR Neg Hits@10 EPFO MRR EPFO Hits@10 Neg MRR Neg Hits@10 EPFO MRR EPFO Hits@10 Neg MRR Neg Hits@10
UltraQuery Paper 0.301 0.428 0.152 0.264 0.335 0.467 0.132 0.260 0.321 0.479 0.156 0.291 0.275 0.375 0.153 0.242
UltraQuery PyG 0.309 0.432 0.178 0.286 0.411 0.518 0.240 0.352 0.312 0.468 0.139 0.262 0.280 0.380 0.193 0.288

In particular, we reach SOTA on FB15k queries (0.764 MRR & 0.834 Hits@10 on EPFO; 0.567 MRR & 0.725 Hits@10 on negation) compared to much larger and heavier baselines (such as QTO).

Run Inference

The running format is similar to the KG completion pipeline - use run_query.py and run_query_many for running a single expriment on one dataset or on a sequence of datasets. Due to the size of the datasets and query complexity, it is recommended to run inference on a GPU.

An example command for running transductive inference with UltraQuery on FB15k237 queries

python script/run_query.py -c config/ultraquery/transductive.yaml --dataset FB15k237LogicalQuery --epochs 0 --bpe null --gpus [0] --bs 32 --threshold 0.0 --ultra_ckpt null --qe_ckpt /path/to/ultra/ckpts/ultraquery.pth

An example command for running transductive inference with a vanilla Ultra 4g on FB15k237 queries with scores thresholding

python script/run_query.py -c config/ultraquery/transductive.yaml --dataset FB15k237LogicalQuery --epochs 0 --bpe null --gpus [0] --bs 32 --threshold 0.8 --ultra_ckpt /path/to/ultra/ckpts/ultra_4g.pth --qe_ckpt null

An example command for running inductive inference with UltraQuery on InductiveFB15k237Query:550 queries

python script/run_query.py -c config/ultraquery/inductive.yaml --dataset InductiveFB15k237Query --version 550 --epochs 0 --bpe null --gpus [0] --bs 32 --threshold 0.0 --ultra_ckpt null --qe_ckpt /path/to/ultra/ckpts/ultraquery.pth

New arguments for _query scripts:

  • --threshold: set to 0.0 when using the main UltraQuery checkpoint ultraquery.pth or 0.8 (and higher) when using vanilla Ultra checkpoints
  • --qe_ckpt: path to the UltraQuery checkpoint, set to null if you want to run vanilla Ultra checkpoints
  • --ultra_ckpt: path to the original Ultra checkpoints, set to null if you want to run the UltraQuery checkpoint

Datasets

23 new datasets available in datasets_query.py that will be automatically downloaded upon the first launch. All datasets include 14 standard query types (1p, 2p, 3p, 2i, 3i, ip, pi, 2u-DNF, up-DNF, 2in, 3in,inp, pin, pni).

The standard protocol is training on 10 patterns without unions and ip,pi queries (1p, 2p, 3p, 2i, 3i, 2in, 3in,inp, pin, pni) and running evaluation on all 14 patterns including 2u, up, ip, pi.

Transductive query datasets (3)

All are the BetaE versions of the datasets including queries with negation and limiting the max number of answers to 100

  • FB15k237LogicalQuery, FB15kLogicalQuery, NELL995LogicalQuery
Inductive (e) query datasets (9)

9 inductive datasets extracted from FB15k237 - first proposed in Inductive Logical Query Answering in Knowledge Graphs (NeurIPS 2022)

InductiveFB15k237Query with 9 versions where the number shows the how large is the inference graph compared to the train graph (in the number of nodes):

  • 550, 300, 217, 175, 150, 134, 122, 113, 106

In addition, we include the InductiveFB15k237QueryExtendedEval dataset with the same versions. Those are supposed to be inference-only datasets that measure the faithfulness of complex query answering approaches. In each split, as validation and test graphs extend the train graphs with more nodes and edges, training queries now have more true answers achievable by simple edge traversal (no missing link prediction required) - the task is to measure how well CLQA models can retrieve new easy answers on training queries but on larger unseen graphs.

Inductive (e,r) query datasets (11)

11 new inductive query datasets (WikiTopics-CLQA) that we built specifically for testing UltraQuery. The queries were sampled from the WikiTopics splits proposed in Double Equivariance for Inductive Link Prediction for Both New Nodes and New Relation Types

WikiTopicsQuery with 11 versions

  • art, award, edu, health, infra, loc, org, people, sci, sport, tax

Metrics

New metrics include auroc, spearmanr, mape. We don't support Mean Rank mr in complex queries. If you ever see nan in one of those metrics, consider reducing the batch size as those metrics are computed with the variadic functions that might be numerically unstable on large batches.

Citation

If you find this codebase useful in your research, please cite the original papers.

The main ULTRA paper:

@inproceedings{galkin2023ultra,
    title={Towards Foundation Models for Knowledge Graph Reasoning},
    author={Mikhail Galkin and Xinyu Yuan and Hesham Mostafa and Jian Tang and Zhaocheng Zhu},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=jVEoydFOl9}
}

UltraQuery:

@inproceedings{galkin2024ultraquery,
  title={A Foundation Model for Zero-shot Logical Query Reasoning},
  author={Mikhail Galkin and Jincheng Zhou and Bruno Ribeiro and Jian Tang and Zhaocheng Zhu},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year={2024},
  url={https://openreview.net/forum?id=JRSyMBBJi6}
}