ViTForImageClassification

This model is a fine-tuned version of google/vit-base-patch16-224-in21k on the CIFAR10 dataset. It achieves the following results on the evaluation set:

• Loss: 0.1199
• Accuracy: 0.9678

Model description

A detailed description of model architecture can be found here

Training and evaluation data

CIFAR10

Training procedure

Straightforward tuning of all model's parameters.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 128
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
0.2995	0.27	100	0.3419	0.9108
0.2289	0.53	200	0.2482	0.9288
0.1811	0.8	300	0.2139	0.9357
0.0797	1.07	400	0.1813	0.946
0.1128	1.33	500	0.1741	0.9452
0.086	1.6	600	0.1659	0.9513
0.0815	1.87	700	0.1468	0.9547
0.048	2.13	800	0.1393	0.9592
0.021	2.4	900	0.1399	0.9603
0.0271	2.67	1000	0.1334	0.9642
0.0231	2.93	1100	0.1228	0.9658
0.0101	3.2	1200	0.1229	0.9673
0.0041	3.47	1300	0.1189	0.9675
0.0043	3.73	1400	0.1165	0.9683
0.0067	4.0	1500	0.1145	0.9697

Framework versions

• Transformers 4.34.1
• Pytorch 2.0.1+cu117
• Datasets 2.12.0
• Tokenizers 0.14.1