Hi @wehos,
Great and well organized code :)
I see that the spatial information are being learned in the attention mechanism in flowformer.
To my understanding, biology requires you make sure that you feed all the tissue cells of a region or a patient at once otherwise you will lose some spatial proximity information because cells are not fed in their spatial proximity order. For example if my batch size is 128 cells, flowformer will only learn the spatial proximity information about this local patch of the tissue (if they are even spatially close to each other) and will miss learning the spatial proximity information of between the cells of the current batch and the next ones.
I see that there is a fixed number for max_batch_size you can choose in config and that means you do not always feed all the cells at once which is biologically problematic. What do you think? and please correct me if I my understanding is wrong.
Thanks
Mohammed
Hi @wehos,
Great and well organized code :)
I see that the spatial information are being learned in the attention mechanism in flowformer.
To my understanding, biology requires you make sure that you feed all the tissue cells of a region or a patient at once otherwise you will lose some spatial proximity information because cells are not fed in their spatial proximity order. For example if my batch size is 128 cells, flowformer will only learn the spatial proximity information about this local patch of the tissue (if they are even spatially close to each other) and will miss learning the spatial proximity information of between the cells of the current batch and the next ones.
I see that there is a fixed number for max_batch_size you can choose in config and that means you do not always feed all the cells at once which is biologically problematic. What do you think? and please correct me if I my understanding is wrong.
Thanks
Mohammed