Vision transformer for mobilenet size and speed

A method of using a vision transformer, comprising:

• embedding an image using a convolution stem;
• capturing local information using a unified feed forward network (FNN);
• modeling spatial dependencies of the image using global multi head self attenuation (MHSA) blocks;
• and
• performing position encoding using a learnable attention bias.

A non-transitory computer readable medium storing program code, which when executed, is operative to cause a vision transformer to perform the steps of:

• embedding an image using a convolution stem;
• capturing local information using a unified feed forward network (FNN);
• modelling spatial dependencies of the image using global multi head self attenuation (MHSA) blocks;
• and
• performing position encoding using a learnable attention bias, wherein the vision transformer optimizes speed and size.