the simplicity of the transformer is quite refreshing. especially in vision where the Vision Transformer with linear patch encodings replaces complex intertwine

the simplicity of the transformer is quite refreshing. especially in vision where the Vision Transformer with linear patch encodings replaces complex intertwined decisions about filter size, striding, pooling, #filters, depth, etc., with the simpler decision of how to allocate your FLOPS between dimensionality, #heads, and #layers.