vit-mushroom-classifier
Fine-tuned google/vit-base-patch16-224 image classifier for 16 mushroom species.
This model was trained for an educational mushroom image classification project. It predicts the most likely species among the 16 trained classes and can be used for candidate ranking, demos, and model comparison. It is not suitable for real-world foraging, edibility, or toxicity decisions.
Model Details
- Base model:
google/vit-base-patch16-224
- Architecture:
ViTForImageClassification
- Task: single-label image classification
- Image size:
224 x 224
- Classes: 16 mushroom species
- Final model directory:
models/vit_mushroom_classifier
- Final evaluation date: 2026-06-03
- Transformers version in config:
4.57.6
Classes
1Amanita_muscaria
2Amanita_phalloides
3Armillaria_mellea
4Cerioporus_squamosus
5Chlorophyllum_brunneum
6Clitocybe_nuda
7Coprinellus_micaceus
8Coprinus_comatus
9Flammulina_velutipes
10Gliophorus_psittacinus
11Hygrophoropsis_aurantiaca
12Hypholoma_lateritium
13Stereum_hirsutum
14Suillus_luteus
15Tricholomopsis_rutilans
16Tylopilus_felleus
Dataset
The image dataset combines a Kaggle mushroom image dataset with additional GBIF image records for the selected species.
Final image scope:
- Total images: 11,200
- Classes: 16
- Images per class: 700
- Train images: 7,839
- Validation images: 1,680
- Test images: 1,681
- Split: stratified 70% / 15% / 15%
Training Procedure
The original ImageNet classifier head was replaced with a new 16-class head. Training used two phases:
- Train the new classifier head.
- Fine-tune the last 4 ViT encoder blocks.
Main configuration:
- Batch size: 16
- Head training epochs: 1
- Fine-tuning epochs: 4
- Head learning rate:
3e-4
- Fine-tuning learning rate:
2e-5
- Weight decay:
0.01
- Random crop scale:
0.85-1.0
- Color jitter:
0.1
- Best model metric: macro F1
- Early stopping: enabled, patience 2
Validation improved during fine-tuning:
| Fine-tune epoch | Validation loss | Accuracy | Macro F1 | Top-3 accuracy |
|---|
| 1 | 0.3553 | 0.8994 | 0.8986 | 0.9762 |
| 2 | 0.2701 | 0.9196 | 0.9195 | 0.9804 |
| 3 | 0.2437 | 0.9250 | 0.9251 | 0.9833 |
| 4 | 0.2325 | 0.9304 | 0.9303 | 0.9845 |
Test Results
Final test set results:
| Metric | Score |
|---|
| Accuracy | 0.9161 |
| Balanced accuracy | 0.9161 |
| Macro F1 | 0.9162 |
| Top-3 accuracy | 0.9780 |
Bootstrap 95% confidence intervals:
| Metric | Mean | 95% CI |
|---|
| Accuracy | 0.9165 | 0.9036-0.9286 |
| Macro F1 | 0.9162 | 0.9035-0.9282 |
| Top-3 accuracy | 0.9780 | 0.9714-0.9845 |
Lowest per-class F1 scores:
| Class | Precision | Recall | F1 |
|---|
| Flammulina_velutipes | 0.8302 | 0.8381 | 0.8341 |
| Armillaria_mellea | 0.8762 | 0.8762 | 0.8762 |
| Suillus_luteus | 0.9020 | 0.8762 | 0.8889 |
| Gliophorus_psittacinus | 0.8649 | 0.9143 | 0.8889 |
| Hygrophoropsis_aurantiaca | 0.8727 | 0.9143 | 0.8930 |
Common confusion patterns included:
Suillus_luteus predicted as Tylopilus_felleusCoprinus_comatus predicted as Chlorophyllum_brunneum- Several confusions involving
Flammulina_velutipes, Gliophorus_psittacinus, and Hygrophoropsis_aurantiaca
Known Limitations
- The model only predicts among the 16 trained species.
- It does not know whether an image belongs to an unseen species.
- It may make confident mistakes on visually similar species.
- Some exact duplicate image hashes were found across splits: 55 duplicate rows in 27 hash groups, with 19 groups crossing train/validation/test splits. This may slightly overestimate test performance.
- The dataset combines images from different sources, so source-specific visual patterns may influence predictions.
- The model must not be used for mushroom consumption, toxicity, or foraging decisions.
Intended Use
Suitable uses:
- Educational image classification demos
- Candidate species ranking among the 16 known classes
- Comparison against structured ecological/context models
- Prototype user interfaces for mushroom image classification
Unsuitable uses:
- Real-world mushroom identification without expert review
- Edibility or toxicity classification
- Safety-critical decisions
- Open-world species recognition beyond the 16 labels
Summary
The final ViT model achieved strong and stable test performance: about 91.6% macro F1 and 97.8% top-3 accuracy on a balanced 16-class test set. The result is promising for educational candidate prediction, but a deduplicated re-split and re-evaluation would be recommended before treating the score as final.
