Feat/standard 2 layer gcn

[soma-enyi]

Closes #69

feat: Standard 2-layer Graph Convolutional Network using PyG for node classification

Problem

The existing GCN class was a configurable N-layer model that accepted a hidden_dims list. While flexible, it didn't
match the standard 2-layer GCN architecture from Kipf & Welling (2017) — the
canonical baseline for node classification benchmarks. The training script also had no validation loop, making it
impossible to track generalisation during training.

Changes

astroml/models/gcn.py

Replaced the N-layer implementation with a strict 2-layer GCN:

Input → GCNConv(input_dim, hidden_dim) → ReLU → Dropout → GCNConv(hidden_dim, output_dim) → log_softmax

• Flat, explicit constructor: input_dim, hidden_dim, output_dim, dropout
• No ambiguity about depth — exactly 2 conv layers as per the paper
• hidden_dim=16 matches the original Kipf & Welling configuration

astroml/training/train_gcn.py

• Updated to use the new flat GCN signature
• Added _accuracy() helper for clean reuse across val and test splits
• Validation accuracy logged every 20 epochs alongside training loss
• model.eval() + torch.no_grad() correctly scoped during inference

Before / After

Before:
python
GCN(input_dim=..., hidden_dims=[64], output_dim=..., dropout=0.5)

After:
python
GCN(input_dim=..., hidden_dim=16, output_dim=..., dropout=0.5)

Training output now looks like:
Epoch 20 | Loss: 1.2341 | Val Acc: 0.7100
Epoch 40 | Loss: 0.9823 | Val Acc: 0.7560
...
Test Accuracy: 0.8120

Related

Closes #[69] — Standard 2-layer Graph Convolutional Network using PyG for node classification

[drips-wave]

@soma-enyi Great news! 🎉 Based on an automated assessment of this PR, the linked Wave issue(s) no longer count against your application limits.

You can now already apply to more issues while waiting for a review of this PR. Keep up the great work! 🚀

Learn more about application limits