dinov2-base-1k_1L-boulderspot

This is a model fine-tuned to classify whether an aerial/satellite image contains a climbing area or not.

You can find some images to test inference with in this old repo from the original project

Model description

This model is a fine-tuned version of facebook/dinov2-base-imagenet1k-1-layer on the pszemraj/boulderspot dataset.

It achieves the following results on the evaluation set:

• Loss: 0.0519
• Accuracy: 0.9810
• F1: 0.9809
• Precision: 0.9808
• Recall: 0.9810
• Matthews Correlation: 0.8501

Intended uses & limitations

Classification of aerial/satellite imagery, ideally with spacial resolution 10-25 cm (i.e. for 10 cm, each pixel in the image corresonds to approx. 10 cm x 10 cm area on the ground). It may be suitable outside of that, but should be validated as other resolutions were not present in the training data.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 7395
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 5.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1	Precision	Recall	Matthews Correlation
0.1596	1.0	203	0.0733	0.9766	0.9759	0.9757	0.9766	0.8079
0.0635	2.0	406	0.1276	0.9474	0.9522	0.9619	0.9474	0.6845
0.1031	3.0	609	0.0602	0.9751	0.9755	0.9760	0.9751	0.8118
0.0587	4.0	813	0.0512	0.9737	0.9734	0.9732	0.9737	0.7905
0.038	4.99	1015	0.0519	0.9810	0.9809	0.9808	0.9810	0.8501

Framework versions

• Transformers 4.39.2
• Pytorch 2.4.0.dev20240328+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2