Baseline CNN

SimpleCNN is the from-scratch reference: if a pretrained model cannot beat it, transfer learning is not helping.

Key results

59.5%

Balanced accuracy

test set, 2,660 images

67.2%

Accuracy

test set

651,946

Trainable parameters

four conv blocks + MLP head

Provenance. Model scores in this section are copied from the output cells of the AstroVision notebooks (and, for DINOv2 per-class recall, read from an exported figure). The documentation data pack contains no metrics file for them, so they cannot be re-derived from the pack; the cell of origin is listed for each value in _facts/astrovision/FACTS_INDEX.md §3 and shown when hovering over a number.

Architecture

Input (3, 256, 256)
  conv_block(3→32)    2×[Conv3×3 · BN · ReLU] · MaxPool · Dropout2d(0.1)   → 128×128
  conv_block(32→64)   idem                                                  →  64×64
  conv_block(64→128)  idem, Dropout2d(0.2)                                  →  32×32
  Conv3×3(128→256) · BN · ReLU · MaxPool                                    →  16×16
  AdaptiveAvgPool → Linear(256→256) · ReLU · Dropout(0.4) → Linear(256→10)

The Grad-CAM target layer (gradcam_target) is the 128→256 convolution before global pooling.

Training

Loss: cross-entropy with label smoothing 0.1; optimiser AdamW (weight decay 1e-4) with a cosine schedule (trainer.train). The notebook trains for 30 epochs at a learning rate of 1e-3; the recorded output is a final 5-epoch continuation at 1e-4 from an epoch-25 checkpoint, whose best validation balanced accuracy is 0.610. The best-validation checkpoint is kept.

Behaviour by class

SimpleCNN recall per class (test)

0 · Disturbed11%

1 · Merging55%

2 · Round Smooth90%

3 · In-between Round Smooth86%

4 · Cigar Shaped Smooth12%

5 · Barred Spiral68%

6 · Unbarred Tight Spiral63%

7 · Unbarred Loose Spiral61%

8 · Edge-on without Bulge75%

9 · Edge-on with Bulge76%

Smooth and edge-on classes are recognised well, but the model almost ignores two classes: Disturbed (recall 0.11) and Cigar Shaped Smooth (0.12) — the rarest classes. With 651k parameters trained on about 12,400 images, the network cannot learn the minority classes without class re-weighting.