Screen AI

Implementation of the ScreenAI model from the paper: "A Vision-Language Model for UI and Infographics Understanding". The flow is: img + text -> patch sizes -> vit -> embed + concat -> attn + ffn -> cross attn + ffn + self attn -> to out. PAPER LINK:

Install

pip3 install screenai

Usage

import torch
from screenai.main import ScreenAI

# Create a tensor for the image
image = torch.rand(1, 3, 224, 224)

# Create a tensor for the text
text = torch.randn(1, 1, 512)

# Create an instance of the ScreenAI model with specified parameters
model = ScreenAI(
    patch_size=16,
    image_size=224,
    dim=512,
    depth=6,
    heads=8,
    vit_depth=4,
    multi_modal_encoder_depth=4,
    llm_decoder_depth=4,
    mm_encoder_ff_mult=4,
)

# Perform forward pass of the model with the given text and image tensors
out = model(text, image)

# Print the shape of the output tensor
print(out)

License

MIT

Citation

@misc{baechler2024screenai,
    title={ScreenAI: A Vision-Language Model for UI and Infographics Understanding}, 
    author={Gilles Baechler and Srinivas Sunkara and Maria Wang and Fedir Zubach and Hassan Mansoor and Vincent Etter and Victor Cărbune and Jason Lin and Jindong Chen and Abhanshu Sharma},
    year={2024},
    eprint={2402.04615},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Todo

• Implement the nn.ModuleList([]) in the encoder and decoder