EfficientNet-B0

EfficientNetGalaxy wraps torchvision's EfficientNet-B0 (ImageNet weights) and replaces the classifier with Dropout(0.3) → Linear(1280→10). The Grad-CAM target is the last block of features.

Key results

82.7%

Balanced accuracy

fine-tuned checkpoint · test set

84.5%

Accuracy

test set, 2,660 images

84.5%

Best validation

phase 2 (full fine-tuning)

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.

Two-phase transfer learning

Phase	Trainable	Epochs	Learning rate	Best validation balanced acc.
1 — warm-up	classifier head only (`freeze_backbone=True`)	5	1e-3	33.0%
2 — fine-tuning	whole network	25	1e-4	84.5%

The warm-up checkpoint (efficientnet_galaxy10_warmup_best.pt) only initialises phase 2 and is not the reference model. The reference is efficientnet_galaxy10_finetune_best.pt, evaluated in the results notebook. Both phases use the shared trainer.train loop (label smoothing 0.1, AdamW, cosine schedule).

Per class

Class (recall, test)	SimpleCNN	EfficientNet-B0
Disturbed	0.11	0.44
Cigar Shaped Smooth	0.12	0.83
Unbarred Loose Spiral	0.61	0.76
Edge-on without Bulge	0.75	0.96

Full per-class table: Morphology Classes page.

Pretraining on ImageNet roughly closes the gap on the rare classes: Cigar Shaped Smooth goes from 0.12 to 0.83. Disturbed remains the weakest class at 0.44.