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

Loss: 0.0659
Accuracy: 0.9762

Model description

More information needed

Intended uses & limitations

To use, combine known genuine signature with questioned signature into a single image, then run inference.

python

from PIL import Image

def create_comparison_image(img1_path, img2_path):
    # Open images
    img1 = Image.open(img1_path).convert("RGB")
    img2 = Image.open(img2_path).convert("RGB")
    
    # Resize to same height
    height = max(img1.height, img2.height)
    width1 = int(img1.width * (height / img1.height))
    width2 = int(img2.width * (height / img2.height))
    
    img1 = img1.resize((width1, height), Image.LANCZOS)
    img2 = img2.resize((width2, height), Image.LANCZOS)
    
    # Create new image with space for both images
    total_width = width1 + width2
    comparison = Image.new('RGB', (total_width, height))
    
    # Paste images side by side
    comparison.paste(img1, (0, 0))
    comparison.paste(img2, (width1, 0))
    
    return comparison

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2.1687073595562957e-05
train_batch_size: 29
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 5
total_train_batch_size: 145
optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 1
num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
0.3114	0.9991	470	0.1464	0.9477
0.2831	1.9991	940	0.0803	0.9697
0.2806	2.9991	1410	0.0727	0.9756
0.2779	3.9991	1880	0.0744	0.9758
0.2588	4.9991	2350	0.0659	0.9762

Framework versions

Transformers 4.51.3
Pytorch 2.7.0+cu126
Datasets 3.6.0
Tokenizers 0.21.1

aevalone/vit-base-patch16-224-finetuned-forgery

vit-base-patch16-224-finetuned-forgery