Quantized Transformers with Graph Neural Networks

This repository contains the implementation of "Quantized Transformers with Graph Neural Networks" about to be submitted to IEEE Transactions on Signal and Information Processing over Networks (TSIPN).

Introduction

In this work, we propose a method to improve the performance of transformers quantized to low-bit precision. Our main method, BitTransGNN, trains a quantized transformer jointly with a Graph Neural Network (GNN) to improve quantized model performance. We additionally propose two extensions that eliminate the shortcoming of BitTransGNN by seperating the quantized transformer from GNNs after training. These extensions allow to use the quantized transformer solitarily. BitTransGNN methods significantly improve quantized transformer performance while maintaining the efficiency advantages in inference time.

Installation

1. Cloning the repository

   git clone https://github.com/EralpK/BitTransGNN.git
   

2. Installing the required dependencies

   We recommend to create a Conda environment or a virtual environment and install the dependencies to the environment.

   pip install -r requirements.txt

   Alternatively, the dependencies can also be installed from the .yml file:

   conda env create -f environment.yml

Running the Code

The main folder contains bittransgnn and bittransgnn_calcs folders. bittransgnn folder contains the scripts written to implement the methods introduced in the manuscript. bittransgnn_calcs contains sample notebooks that calculate the theoretical memory and energy consumption of different quantized methods. To run the method, first navigate to the bittransgnn directory:

cd bittransgnn

Notebook files that investigate the memory and energy efficiency of different transformer models can be inspected within the bittransgnn_calcs folder.

Dataset Preprocessing

For 20NG, MR, R8, R52 and Ohsumed datasets, we first clean and tokenize the raw text data.

python remove_words.py

To construct the raw text files for the GLUE datasets, prepare_data_glue.py script can be used.

python prepare_data_glue.py

After the preprocessing, the text graphs for each dataset are constructed:

python build_graph.py

The text graphs constructed through build_graph.py contain all train, test, and validation nodes. In case one wants to operate methods only over the training nodes, the following command should be executed:

python build_graph_inductive.py

Methods

Our work is a composition of four methods. The four methods are contained within the /methods directory. We list the methods in the table below:

Method	Description
bittransformer	Fine-tunes full-precision and quantized transformer models before integrating them into the main architecture.
bittransgnn	Trains a quantized transformer jointly with a GNN in different configurations.
direct-seperation	Seperates the quantized transformer from the joint architecture after training for inference.
knowledge-distillation	Compresses the knowledge of the joint architecture into a quantized transformer model through knowledge distillation.

Each of the bittransformer, bittransgnn and knowledge-distillation methods contain train.py, run_sweep.py and inference.py scripts. direct-seperation involves no additional training and only applies inference. For that reason, it only contains the inference.py script.

train.py script is used to train the method based on the configurations defined in configs/train_config.py. sweep.py conducts training over the ranges of hyperparameters defined in configs/sweep_config.yaml. The trained models can be used for inference using the script in inference.py based on the configuration in configs/sweep_config.yaml.

To use each script,

python -m methods.{framework}.{mode}

where {framework} refers to one of the four methods being used and {mode} defines whether the script is used for training, testing, or sweeping through hyperparameters.

Fine-Tuning Transformer Models with Quantization

We fine-tune quantized transformer models before integrating them into our architecture. The fine-tuning is conducted through the following code:

python -m methods.bittransformer.train

The save_ckpt variable should be set to True in the configs/train_config.yaml file if one wants to save model checkpoint for later use in the joint architecture.

BitTransGNN: Training Quantized Transformers with GNNs

We combine quantized transformers with GNNs during training to improve their performance and enhance their predictions. We use a Graph Convolutional Network (GCN) as the GNN companion. We interpolate the outputs of the quantized transformer and the GNN during training.

We propose Static and Dynamic variants of BitTransGNN methods. Static methods use quantized transformer models only for inference, while Dynamic methods jointly train the two models. The joint_training variable should be set to True to train the Dynamic variant, and it should be set to False to train a Static BitTransGNN method. The following line of code should be run to train BitTransGNN methods:

python -m methods.bittransgnn.train

Note that the parameters of a fine-tuned quantized transformer should be saved either within the local machine or an external source to be able to operate the joint architecture.

We combine quantization and fine-tuning in two different stages. Quantization-Aware Fine-Tuning (QAFT) first quantizes the transformer parameters and then fine-tunes the model over the task. Post Fine-Tuning Quantization (PFTQ) conducts quantization after fine-tuning is completed.

We use both QAFT and PFTQ while integrating quantized transformer models into Dynamic training. We only integrate QAFT methods into Static BitTransGNN training. The fine-tuning method can be chosen by changing the bert_quant_type file in the configuration file.

Different $\lambda$ values can be searched using the run_sweep.py script.

For use in direct-seperation and knowledge-disillation methods, the save_ckpt variable should be set to True.

To initialize the node embeddings of the teacher BitTransGNN model with the latest state of the BitTransformer model when using the KD method and when employing the trained BitTransGNN method for inference, save_logits should be set to True. In that case, ext_cls_feats are used to initialize the document node embeddings.

Direct Seperation (DS)

After BitTransGNN training, DS seperates the quantized transformer and its classifier from the GCN after training and uses them for inference, without GCN's assistance during inference time.

To use the quantized transformer that is jointly trained with the GNN solitarily in inference time:

python -m methods.direct-seperation.inference

Induction Through Knowledge Distillation (KD)

KD compresses the knowledge of the BitTransGNN model into a solitary quantized transformer through distillation. A BitTransGNN model is used as the teacher and a BitTransformer model is used as the student model.

To train the student BitTransformer:

python -m methods.knowledge_distillation.train

We conduct offline distillation. The distillation type can be modified from the configuration file. The sweeping script can be used to search over the distillation parameters $\alpha_d$ and $T$:

python -m methods.knowledge_distillation.run_sweep

Calculations on Memory and Energy Efficiency

The exact values of parameter counts and the number of arithmetic operations can be found in the scripts within the bittransgnn_calcs/ folder.

For sample calculations on memory and energy efficiency and to reproduce the efficiency plots displayed in the manuscript, you can look at the bittransgnn_calcs/bittransgnn_efficiency_notebook.ipynb notebook.

Visualization

The script plot_lmbd_vs_acc_static.py can be used to plot the accuracy of a model in different quantization rates for varying values of $\lambda$. The script is originally written to focus on Static BitTransGNN variants, but it can be very simply adapted to the Dynamic variant as well.

cd visualization
python plot_lmbd_vs_acc_static.py

To conduct the other visualization methods, including dimensionality reduction techniques such as UMAP and T-SNE, and PCA analysis applied on the similarity profiles of the final logits of BitTransformer and GNN models, can be found in visualization.ipynb.

Results

Results will be made public upon the completion of the reviewing stage.

Acknowledgements

• The data preprocessing and graph construction scripts are adapted from TextGCN and BertGCN repositories.
• We acknowledge the guidance we received from the implementation of BitNet during the construction of our quantization module.