nasyxx / MDF-HS

Readme

Introduction

For paper “Multimodal Deep Fusion in Hyperbolic Space for Mild Cognitive Impairment Study”.

We have a split JAX hyperbolic library here: https://github.com/nasyxx/jaxrie

Requirements

dependencies = [
    "jax>=0.4.3",
    "jaxlib>=0.4.3",
    "jaxopt>=0.6",
    "optax>=0.1.4",
    "dm-haiku>=0.0.9",
    "tqdm>=4.64.1",
    "smile-config[rich]>=0.11.0",
    "scikit-learn>=1.2.1",
    "jaxtyping>=0.2.12",
    "chex>=0.1.6",
    "matplotlib>=3.7.1",
]

Installation

pip install pdm
pdm install

Usage

> python -m train --help

Usage: config.py [-h] [--data_path str] [--model_dir str] [--seed int] [--reload | --no-reload] [--norm | --no-norm] [--val | --no-val]
                 [--split | --no-split] [--group | --no-group] [--batch_size int] [--epochs int] [--chkt int] [--lr float] [--k int] [--optim str]
                 [--model str] [--manifold Annotated] [--c float] [--bt | --no-bt] [--net.out_size str [str ...]] [--net.activation str]
                 [--net.leaky_relu_slope float] [--net.dropedge | --no-net.dropedge] [--net.dropedge_rate float] [--net.dropout | --no-net.dropout]
                 [--net.dropout_rate float] [--net.use_att | --no-net.use_att] [--net.heads int] [--net.concat | --no-net.concat]
                 [--net.pearson | --no-net.pearson] [--net.trainc | --no-net.trainc]

Configuration for brain.

Options:
  -h, --help            show this help message and exit
  --data_path str       - (default: data/Data_MIA/Data_MIA_deepFusion)
  --model_dir str       - (default: results)
  --seed int            - (default: 7)
  --reload, --no-reload
                        - (default: False)
  --norm, --no-norm     - (default: True)
  --val, --no-val       - (default: True)
  --split, --no-split   - (default: True)
  --group, --no-group   - (default: True)
  --batch_size int      - (default: 1000)
  --epochs int          - (default: 200)
  --chkt int            - (default: 50)
  --lr float            - (default: 0.001)
  --k int               - (default: 5)
  --optim str           - (default: adam)
  --model str           - (default: bgcn)
  --manifold Annotated  (e)uclidean|(p)oincare (default: euclidean)
  --c float             - (default: 1.0)
  --bt, --no-bt         - (default: False)

Net:
  --net.out_size str [str ...]
                        - (default: (148, 296))
  --net.activation str  - (default: leaky_relu)
  --net.leaky_relu_slope float
                        - (default: 0.2)
  --net.dropedge, --no-net.dropedge
                        - (default: True)
  --net.dropedge_rate float
                        - (default: 0.8)
  --net.dropout, --no-net.dropout
                        - (default: True)
  --net.dropout_rate float
                        - (default: 0.6)
  --net.use_att, --no-net.use_att
                        - (default: False)
  --net.heads int       - (default: 4)
  --net.concat, --no-net.concat
                        - (default: False)
  --net.pearson, --no-net.pearson
                        - (default: True)
  --net.trainc, --no-net.trainc
                        - (default: False)

Example

Train brain GCN in Euclidean space

python -m train --model "bgcn" --manifold "e" --norm --val --seed 10

Train brain GCN with Poincaré embeddings

python -m train --model "bgcn" --manifold "p" --norm --val --seed 10

Train brain Hyperbolic GCN

python -m train --model "bhgcn" --manifold "p" --norm --val --seed 10