ArchBERT: Bi-Modal Understanding of Neural Architectures and Natural Languages

This repo contains the implementation and demo of the paper 'ArchBERT: Bi-Modal Understanding of Neural Architectures and Natural Languages' that has been accepted to appear at the CoNLL2023 conference.

Link to arXiv paper: https://arxiv.org/abs/2310.17737

Link to Huawei's AI Gallery Notebook: https://developer.huaweicloud.com/develop/aigallery/notebook/detail?id=58b799a0-5cfc-4c2e-8b9b-440bb2315264

Download Code and Datasets

!wget https://vbdai-notebooks.obs.cn-north-4.myhuaweicloud.com/archbert/code.zip
!unzip -qo code.zip

Package Requirements

• Anaconda (version 2020.07)
• All the other requirements are listed in environment.yml file
• After installing Anaconda, use the following command to create an conda environment with the required packages:

!conda env create -f environment.yml
!conda activate archbert

AutoNet dataset generation

The code for creating AutoNet and AutoNet-AQA train/val sets (with e.g., 100 neural architectures):

!python ./data/autonet_generator.py train 100 ./data/datasets/autonet default default
!python ./data/autonet_generator.py val 100 ./data/datasets/autonet default default
!python ./data/autonet_generator.py train 100 ./data/datasets/autonet_qa qa multi
!python ./data/autonet_generator.py val 100 ./data/datasets/autonet_qa qa multi

TVHF dataset generation

Run the following command to generate the TVHF train and validation sets.

• path: the path to save the generated dataset
• num_nets: the number of architectures to be generated

!python ./data/tvhf_dataset_generator --path=./datasets/tvhf/ --num_nets=5

Architecture Reasoning (AR) on TVHF

• num_nets: the number of architectures to be evaluated

!python test_archbert.py
        --task=reasoning
        --dataset=tvhf
        --batch_size=1
        --layernorm
        --cross_encoder
        --data_dir=./data/datasets/tvhf
        --model_dir=./pretrained-models/archbert_tvhf
        --validate
        --num_nets=100

Architecture Clone Detection (ACD) on TVHF

!python test_archbert.py
        --task=na_clone_detection
        --dataset=tvhf
        --batch_size=1
        --layernorm
        --cross_encoder
        --data_dir=./data/datasets/tvhf
        --model_dir=./pretrained-models/archbert_tvhf
        --validate
        --num_nets=100

Architecture Captioning (AC) on AutoNET

!python test_archbert.py
        --task=langdec
        --dataset=autonet
        --batch_size=1
        --layernorm
        --cross_encoder
        --data_dir=./data/datasets/autonet
        --model_dir=./pretrained-models/archbert_autonet_ac
        --validate
        --num_nets=100

Video demo for different downstream tasks

Architectural Reasoning (AR)

%%HTML
<video width="1280" controls>
    <source src="https://vbdai-notebooks.obs.cn-north-4.myhuaweicloud.com/archbert/demo/AR_TVHF.mp4" type="video/mp4">
</video>

Architecture Search (AS)

%%HTML
<video width="1280" controls>
    <source src="https://vbdai-notebooks.obs.cn-north-4.myhuaweicloud.com/archbert/demo/AS_TVHF.mp4" type="video/mp4">
</video>

Architecture Clone Detection (ACD)

%%HTML
<video width="1280" controls>
    <source src="https://vbdai-notebooks.obs.cn-north-4.myhuaweicloud.com/archbert/demo/ACD_TVHF.mp4" type="video/mp4">
</video>

Bi-Modal Architecture Clone Detection (BACD)

%%HTML
<video width="1280" controls>
    <source src="https://vbdai-notebooks.obs.cn-north-4.myhuaweicloud.com/archbert/demo/BACD_TVHF.mp4" type="video/mp4">
</video>

Architecture Captioning (AC)

%%HTML
<video width="1280" controls>
    <source src="https://vbdai-notebooks.obs.cn-north-4.myhuaweicloud.com/archbert/demo/AC_AutoNet.mp4" type="video/mp4">
</video>

Architectural Question Answering (AQA)

%%HTML
<video width="1280" controls>
    <source src="https://vbdai-notebooks.obs.cn-north-4.myhuaweicloud.com/archbert/demo/AQA_AutoNet.mp4" type="video/mp4">
</video>