finetuned-vit-base-patch16-224-upside-down-detector

This model is a fine-tuned version of vit-base-patch16-224-in21k on the custom image orientation dataset adapted from the beans dataset. It achieves the following results on the evaluation set:

• Accuracy: 0.8947

Training and evaluation data

The custom dataset for image orientation adapted from beans dataset contains a total of 2,590 image samples with 1,295 original and 1,295 upside down. The model was fine-tuned on the train subset and evaluated on validation and test subsets. The dataset splits are listed below:

Split	# examples
train	2068
validation	133
test	128

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-04
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 32
• num_epochs: 5

Training results

Epoch	Accuracy
0	0.8609
1	0.8835
2	0.8571
3	0.8941
4	0.8941

Framework versions

• Transformers 4.17.0
• Pytorch 1.9.0+cu111
• Pytorch/XLA 1.9
• Datasets 2.0.0
• Tokenizers 0.12.0