Graph-Topological-Data-Analysis

This repo contains code for Topological structure of complex predictions.

packages requirement

• pytorch (with cuda)
• torchvision
• scikit-learn
• numpy
• scipy
• timm
• tqdm
• torch_geometric
• networkx
• seaborn
• rembg

install using conda

The easiest way to setup the environment required by GTDA is to use the included GTDA.yml file. Run this file with the following command:

conda env create --file GTDA.yml

which will create a virtual environment under the current working directory. The installation can take up to one or two hours depending on network speed. Then, the code can be run after activating the conda environment using the following command:

conda activate GTDA

create a Reeb network

from GTDA.GTDA_utils import compute_reeb, NN_model
from GTDA.GTDA import GTDA

nn_model = NN_model()
nn_model.A = G # graph to analyze, in scipy csr format
nn_model.preds = preds # prediction matrix, samples-by-class
nn_model.labels = labels # integer class assignments starting from 0
nn_model.train_mask = train_mask # training mask in bool
nn_model.val_mask = val_mask # validation mask in bool
nn_model.test_mask = test_mask # testing mask in bool
smallest_component = 20
overlap = 0.1
labels_to_eval = list(range(preds.shape[1]))
GTDA_record = compute_reeb(GTDA,nn_model,labels_to_eval,smallest_component,overlap,
    node_size_thd=5,reeb_component_thd=5,nprocs=10,device='cuda')

analyze a Reeb network

g_reeb = GTDA_record['g_reeb'] # Reeb network in csr format
gtda = GTDA_record['gtda'] # an instance of GTDA class
gtda.final_components_filtered[gtda.filtered_nodes[reeb_node_index]] # map a reeb node back to the original component
gtda.A_reeb # projected Reeb network with the same set of nodes as the original graph
gtda.sample_colors_mixing # GTDA estimated errors for each sample

Swiss Roll experiment (demo)

Prerequisites:

None, self contained

Files:

• train_swiss_roll.py generates the dataset and trains the model

• analyze_swiss_roll.ipynb creates and analyzes Reeb network

• dataset/precomputed/swiss_roll contains precomputed lens and graph to run analyze_swiss_roll.ipynb directly

Expected running time:

• A few minutes.

Amazon Electronics experiment

Prerequisites:

Download All products under 'Electronics' from the 2014 version of Amazon reviews data from http://jmcauley.ucsd.edu/data/amazon/index_2014.html and put under dataset/electronics folder

Files:

• train_electronics.py generates the dataset and trains the model

• analyze_electronics.ipynb creates and analyzes Reeb network

• dataset/precomputed/electronics contains precomputed lens and graph to run analyze_electronics.ipynb directly

Imagenette experiment

Prerequisites:

Download from https://github.com/fastai/imagenette and put under dataset/imagenette folder

Files:

• train_imagenette.py generates the dataset and trains the model

• analyze_imagenette.ipynb creates and analyzes Reeb network

• dataset/precomputed/imagenette contains precomputed lens and graph to run analyze_imagenette.ipynb directly

Gene mutation experiment

Prerequisites:

Download variant_summary.txt from https://ftp.ncbi.nlm.nih.gov/pub/clinvar/tab_delimited/ and put under dataset/variants folder

Download human reference genome hg19.fa, hg38.fa and put under dataset/variants folder

Install tensorflow and setup Enformer based on instructions of https://github.com/deepmind/deepmind-research/tree/master/enformer. GPU with cuda support is highly recommended.

Files:

• mutation_dataset.py preprocesses downloaded files, tensorflow is required

• train_gene_mutation.py generates the dataset and trains the model

• analyze_gene_mutation.ipynb creates and analyzes Reeb network

• dataset/precomputed/variants contains precomputed lens and graph to run analyze_gene_mutation.ipynb directly

Additional experiments in supplement

Prerequisites:

To run the CNN model comparison experiment, all training and validation images of ImageNet-1k dataset must be downloaded from https://www.image-net.org/ and put under dataset/imagenet_1k. The folder should be organized as:

dataset/imagenet_1k/
    train/
        class1/
            img1
            img2
            ...
        class2/
        ...
    val/
        class1/
            img1
            img2
            ...
        class2/
        ...

It also requires to install timm package to get the pretrained VOLO model.

To run the chest X-ray experiment, all X-ray images and expert labels should be downloaded from https://cloud.google.com/healthcare-api/docs/resources/public-datasets/nih-chest. The images should be put under dataset/chest_xray/images and expert labels under dataset/chest_xray. We also use the implementation from https://github.com/zoogzog/chexnet to train a DenseNet-121 model.

Files:

• train_analyze_imagenet_1k.py preprocesses downloaded files and build the Reeb network
• train_chest_xray.py preprocesses downloaded files, train a DenseNet-121 model from scratch, build the Reeb network and analyze_chest_xray.ipynb uses precomputed data to find potential labeling errors for images with expert labels

Interactive web interface for better exploration

Once we have computed the Reeb net, other than examing the results in traditional figures, we have also created an interactive web interface using D3 library to explore the results in a web browser. To transform the GTDA results into the format compatible with the web interface, the following function needs to be called:

from GTDA.GTDA_utils import save_to_json
savepath = "web/"
save_to_json(GTDA_record, nn_model, savepath)

This function will save the results to web/reeb_net.js. Currently, this interface doesn't support another filename as reeb_net.js is hard coded in the code. This is to facilitate running the interface locally. Once the function is called, we can open web/index.html to explore the results.

Demos:

https://mengliupurdue.github.io/Graph-Topological-Data-Analysis/

Supported operations:

• Hover over a node to get more information, show or hide legend

• Use mouse wheel or buttons to zoom, move and fit to window, click and hold a node to drag to tweak layout

• Drag while pressing SHIFT to pan components, use buttons to change node size

• Check estimated errors or true errors, filter components by class or id

• Expand to samples or collapse to Reeb net nodes, understand estimated errors by highlighting training/validation