Modelpublic

abalhomaid/disaster-uda-models

sourceHugging Facemitupdated 6mo agoView on Hugging Face

0likes

Model Card

UDA Disaster Damage Assessment - Model Weights

Pre-trained model weights for reproducing the experiments in:

Unsupervised Domain Adaptation for Rapid Disaster Damage Assessment

These checkpoints reproduce Tables 1-6 in the paper.

Models Included

Model	Count	Description
ResNet50 (source-only)	12	State dict files for all 12 source-target domain pairs
CDAN	36	12 tasks x 3 seeds (SWD=0)
CORAL	36	12 tasks x 3 seeds (SWD=0)
MMD	36	12 tasks x 3 seeds (SWD=0)
DANN	36	12 tasks x 3 seeds (SWD=0)
AWDAN (best)	36	12 tasks x 3 seeds (best SWD config per task, selected by max F1)
Total	192 files	~19.8 GB

Domains

Domain	Code	Event	Images
Ecuador Earthquake	E	2016	1,724
Nepal Earthquake	N	2015	19,102
Hurricane Matthew	M	2016	333
Typhoon Ruby	R	2014	833

All 12 source-to-target combinations (E2M, E2N, E2R, M2E, M2N, M2R, N2E, N2M, N2R, R2E, R2M, R2N) are evaluated with seeds {2, 32, 128}.

Download

bash

pip install huggingface_hub
huggingface-cli download abalhomaid/disaster-uda-models --local-dir .

This places files into the correct directory structure expected by the evaluation scripts.

File Structure

models/resnet50/c4/
    {S}_{T}_model_epoch_best_statedict.pth    # Source-only baseline (12 files)

train_jobs/cdan/logs/{model}/seed_{seed}/
    Damage_{S}2{T}_SWD_{swd}_trade_offs_{to}/
        checkpoints/best.pth                   # DA model checkpoint

Where {model} is one of: cdan, coral, mmd, dann.

Usage

Clone the reproduction repository and download the weights:

bash

git clone https://github.com/abalhomaid/disaster-assesment.git
cd disaster-assesment
git checkout reproducibility

# Download model weights
huggingface-cli download abalhomaid/disaster-uda-models --local-dir .

# Set up evaluation environment
conda env create -f tllib_metric.yaml
conda activate tllib_metric
pip install -e .

# Evaluate a model (e.g., DANN E->M, seed=32)
PYTHONPATH="$PWD:$PWD/examples/domain_adaptation/image_classification" \
python examples/domain_adaptation/image_classification/dann.py \
  data/damage -d Damage -s E -t M -a resnet50 --seed 32 --scratch \
  --log train_jobs/cdan/logs/dann/seed_32/Damage_E2M_SWD_0_trade_offs_1 \
  --phase test

See the repository README for full reproduction instructions.

Expected Results (Table 5 Averages)

Model	Accuracy	Precision	Recall	F1
Source-only	75.2	84.1	70.0	75.6
CDAN	81.0	83.6	83.1	83.2
CORAL	79.5	82.2	81.7	81.8
MMD	81.1	83.8	83.0	83.2
DANN	80.6	83.9	81.9	82.7
AWDAN	81.9	83.9	85.0	84.2

Architecture

All models use ResNet50 backbone with 2-class output (damage / no damage), input size 224x224.

—Source-only: Standard ResNet50 fine-tuned on source domain only
—DANN/CDAN: ResNet50 + domain discriminator (adversarial DA)
—CORAL/MMD: ResNet50 + divergence minimization (domain generalization)
—AWDAN: DANN + Sliced Wasserstein Distance regularization in label space