vit-base-rocks

This model is a fine-tuned version of google/vit-base-patch16-224-in21k on the rocks dataset. It achieves the following results on the evaluation set:

• Loss: 0.7099
• Accuracy: 0.7778

Model description

This model is a fine-tuned version of Google's vit-base-patch16-224-in21k designed to identify geological hand samples.

Intended uses & limitations

Currently the VIT is fine-tuned on 10 classes:

['Andesite', 'Basalt', 'Chalk', 'Dolomite', 'Flint', 'Gneiss', 'Granite', 'Limestone', 'Sandstone', 'Slate']

Future iteartions of the model will feature an expanded breadth of rock categories.

Training and evaluation data

The model performs relatively well on 10 classes of rocks - with some confusion between limestone and other carbonates.

Training procedure

495 images of geological hand samples were selected with an 80:20 train-test/validation split.

Classes were roughly equally represented across the 495 samples.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 64
• eval_batch_size: 8
• seed: 42
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 25

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
2.0408	1.4286	10	1.7371	0.6111
1.4489	2.8571	20	1.3254	0.7407
0.9469	4.2857	30	1.0768	0.7407
0.586	5.7143	40	0.9118	0.7778
0.3757	7.1429	50	0.9902	0.6852
0.2798	8.5714	60	0.8498	0.7778
0.2087	10.0	70	0.7939	0.7407
0.176	11.4286	80	0.8220	0.7222
0.1613	12.8571	90	0.7288	0.8148
0.1337	14.2857	100	0.7178	0.7963
0.1326	15.7143	110	0.7403	0.7778
0.119	17.1429	120	0.7099	0.7778
0.1193	18.5714	130	0.7626	0.7778
0.1227	20.0	140	0.7125	0.7963
0.1102	21.4286	150	0.7493	0.7963
0.1134	22.8571	160	0.7396	0.7963
0.1173	24.2857	170	0.7187	0.7963

Framework versions

• Transformers 4.48.3
• Pytorch 2.6.0
• Datasets 3.3.0
• Tokenizers 0.21.0