Thank you for sharing your work! It sounds like you're dealing with larger memory usage during the training process, whether using DSVT_P or DSVT_V. If I want to reduce the Sparse_shape, i meet some bug of model, such as set_attition and pointpillar_scatter. Apart from adjusting INPUT_LAYER.sparse_shape, output_shape, and MAP_TO_BEV.input_shape, anything else need to be adjust? And this params is decision by what?
Thank you for sharing your work! It sounds like you're dealing with larger memory usage during the training process, whether using DSVT_P or DSVT_V. If I want to reduce the Sparse_shape, i meet some bug of model, such as set_attition and pointpillar_scatter. Apart from adjusting INPUT_LAYER.sparse_shape, output_shape, and MAP_TO_BEV.input_shape, anything else need to be adjust? And this params is decision by what?