GalaxyClassifier

Deep learning ensemble for galaxy morphology classification on the Galaxy10 DECaLS dataset.

87.92% macro-average test accuracy (TTA) on 10 galaxy morphology classes.

---

Architecture

3-model weighted soft-voting ensemble trained with PyTorch:

Model	Params	Test Accuracy (TTA)
ConvNeXt-Tiny	28M	87.97%
ResNeXt-50 (32x4d)	23M	86.40%
DenseNet-161	27M	86.28%

Each model trained independently with:

• 224x224 input, ImageNet pretrained weights
• WeightedRandomSampler for class imbalance
• CosineAnnealingWarmRestarts (T0=20, T_mult=2)
• Progressive stem unfreezing at epoch 20
• Mixed precision (AMP) + gradient accumulation (effective batch 64)
• 8-view Test-Time Augmentation (flips + rotations)

---

System Architecture

graph LR
    subgraph Input
        A[Galaxy Image from User]
    end

    subgraph Preparation
        B[Resize and Normalize Image]
        C[Create 8 Flipped and Rotated Copies]
    end

    subgraph Three Trained Models
        D[Model 1 - ConvNeXt]
        E[Model 2 - ResNeXt]
        F[Model 3 - DenseNet]
    end

    subgraph Combining Results
        G[Merge All Predictions with Weighted Voting]
        H[Pick the Galaxy Type with Highest Confidence]
    end

    A --> B
    B --> C
    C --> D
    C --> E
    C --> F
    D --> G
    E --> G
    F --> G
    G --> H

Loading

How It Works

Step	What Happens
1	You upload a photo of a galaxy
2	The image is resized and its colors are adjusted so all three models can read it consistently
3	Eight slightly different versions of the image are created using flips and rotations, giving each model more angles to consider
4	All three models independently look at every version of the image and guess what kind of galaxy it is
5	The three sets of guesses are combined — models that have historically been more accurate get a bigger say in the final decision
6	The galaxy type with the highest combined confidence score is returned as the final answer

---

Dataset

Galaxy10 DECaLS — 17,736 images, resized to 224x224, 10 classes:

Class	Test Acc (TTA)
Disturbed Galaxies	65.4%
Merging Galaxies	93.2%
Round Smooth Galaxies	95.7%
In-between Round Smooth Galaxies	91.8%
Cigar Shaped Smooth Galaxies	94.0%
Barred Spiral Galaxies	84.4%
Unbarred Tight Spiral Galaxies	87.2%
Unbarred Loose Spiral Galaxies	75.4%
Edge-on Galaxies without Bulge	96.7%
Edge-on Galaxies with Bulge	95.4%

---

Setup

pip install -r requirements.txt

Download Galaxy10_DECals.h5 from AstroNN and place it in data/.

---

Training

# Train a model  (convnext | resnext50 | densenet161)
python src/train_optimized_v3.py --model convnext --size 224 --epochs 80

# Resume from checkpoint
python src/train_optimized_v3.py --model convnext --resume models/checkpoint_v3_convnext_*.pth

Weights save to models/best_v3_{model}_{size}_{timestamp}.pth.

---

Running

API server (port 8080):

python src/inference_server.py

Loads all 3 ensemble models automatically from models/.

Endpoint	Method	Description
/api/predict	POST	Classify uploaded image (multipart image)
/api/gradcam	POST	Grad-CAM visualization
/api/health	GET	Model load status
/api/gpu-stats	GET	GPU memory usage

Web UI (port 3000):

cd frontend && npm install && npm run dev

Open http://localhost:3000.

---

Project Structure

src/
  inference_server.py     # Flask inference server (port 8080)
  train_optimized_v3.py   # Training (all 3 models)
  load_data.py            # HDF5 loading, uint8 RAM caching
frontend/
  App.tsx                 # React classifier UI (port 3000)
  components/             # ModelCard, EnsembleFeed
models/
  best_v3_convnext_224_*.pth
  best_v3_resnext50_224_*.pth
  best_v3_densenet161_224_*.pth
tests/
  test_api.py
  test_classifier.py

---

Hardware

Trained on RTX 3050 Laptop GPU (4GB VRAM). Peak VRAM ~1.2GB per model at batch=16.