vit-fruit-classifier

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

• Loss: 1.0194
• Accuracy: 0.9699

Training and evaluation data

This model was fine-tuned on the Fruits-30 dataset, a collection of images featuring 30 different types of fruits. Each image has been preprocessed and standardized to a size of 224x224 pixels for uniformity.

Dataset Composition

• Number of Classes: 30
• Image Resolution: 224x224 pixels
• Total Images: 826

Training and Evaluation Split

The dataset was split into training and evaluation sets using dataset.train_test_split function with a 80/20 train-test split, resulting in:

• Training Set: 660 images
• Evaluation Set: 166 images

Splitting Strategy

• The data was shuffled (shuffle=True) before splitting to ensure a random distribution of classes across the training and evaluation sets.
• Additionally, stratification was applied based on the "label" column (stratify_by_column='label') to maintain a balanced class distribution across both sets. This helps prevent the model from biasing towards classes with more samples in the training data.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
2.668	2.38	100	2.0731	0.9217
1.6565	4.76	200	1.4216	0.9518
1.1627	7.14	300	1.1256	0.9578
0.9571	9.52	400	1.0224	0.9639

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2