vit-artworkclassifier

This model returns the artwork style of any image input.

This model is a fine-tuned version of google/vit-base-patch16-224-in21k on the imagefolder dataset. This is a subset of the artbench-10 dataset (https://www.kaggle.com/datasets/alexanderliao/artbench10), with a train set of 1000 artworks per class and a validation set of 100 artworks per class. It achieves the following results on the evaluation set:

• Loss: 1.1392
• Accuracy: 0.5948

Model description

You can find a description of the project that this model was trained for here: https://medium.com/@oliverpj.schamp/training-and-evaluating-stable-diffusion-for-artwork-generation-b099d1f5b7a6

Intended uses & limitations

This model only contains 9 out of the 10 artbench-10 classes - it does not contain ukiyo_e. This was due to availability and formatting issues.

Training and evaluation data

Train: 1000 randomly selected images from artbench-10 (per class). Val: 100 randomly selected images from artbench-10 (per class).

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 4
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
1.5906	0.36	100	1.4709	0.4847
1.3395	0.72	200	1.3208	0.5074
1.1461	1.08	300	1.3363	0.5165
0.9593	1.44	400	1.1790	0.5846
0.8761	1.8	500	1.1252	0.5902
0.5922	2.16	600	1.1392	0.5948
0.4803	2.52	700	1.1560	0.5936
0.4454	2.88	800	1.1545	0.6118
0.2271	3.24	900	1.2284	0.6039
0.207	3.6	1000	1.2625	0.5959
0.1958	3.96	1100	1.2621	0.6005

Framework versions

• Transformers 4.26.1
• Pytorch 1.13.1+cu117
• Datasets 2.9.0
• Tokenizers 0.13.2

Code to Run

def vit_classify(image):
    vit = ViTForImageClassification.from_pretrained("oschamp/vit-artworkclassifier")
    vit.eval()
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    vit.to(device)
    
    model_name_or_path = 'google/vit-base-patch16-224-in21k'
    feature_extractor = ViTFeatureExtractor.from_pretrained(model_name_or_path)

    #LOAD IMAGE

    encoding = feature_extractor(images=image, return_tensors="pt")
    encoding.keys()

    pixel_values = encoding['pixel_values'].to(device)

    outputs = vit(pixel_values)
    logits = outputs.logits

    prediction = logits.argmax(-1)
    return prediction.item() #vit.config.id2label[prediction.item()]