vit-lr-step

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

• Loss: 0.5312
• Accuracy: 0.8245
• Precision: 0.8216
• Recall: 0.8245
• F1: 0.8048

Training procedure

Early stopping is employed with a patience of 10 and validation loss as the stopping criteria.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: StepLR(optimizer, step_size = 1600, gamma = 0.5, last_epoch=-1)
• num_epochs: 100
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Precision	Recall	F1
0.6607	1.0	321	0.5487	0.8141	0.8096	0.8141	0.8033
0.4016	2.0	642	0.5312	0.8245	0.8216	0.8245	0.8048
0.2341	3.0	963	0.6710	0.8173	0.8126	0.8173	0.8001
0.1273	4.0	1284	0.6510	0.8419	0.8486	0.8419	0.8434
0.0855	5.0	1605	0.8303	0.8339	0.8345	0.8339	0.8251
0.0129	6.0	1926	0.7846	0.8516	0.8568	0.8516	0.8530
0.0008	7.0	2247	0.8298	0.8637	0.8623	0.8637	0.8604
0.0001	8.0	2568	0.8349	0.8644	0.8621	0.8644	0.8613
0.0001	9.0	2889	0.8528	0.8641	0.8617	0.8641	0.8610
0.0001	10.0	3210	0.8711	0.8634	0.8609	0.8634	0.8603
0.0001	11.0	3531	0.8797	0.8634	0.8609	0.8634	0.8603
0.0	12.0	3852	0.8891	0.8634	0.8609	0.8634	0.8603

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2